UbiCC Journal & ICIT 2009 conference - Applied Computing

Special Issue on ICIT 2009 Conference - Applied Computing



THE INFLUENCE OF BLENDED LEARNING MODEL ON DEVELOPING LEADERSHIP SKILLS OF SCHOOL ADMINISTRATORS

Tufan AYTAÇ



The Ministry of National Education, Ankara, TURKEY taytac1@yahoo.com

ABSTRACT The usage of b-learning approach on in-service education activities in Turkish education system are getting more and more important these days. Generally, traditional education and computer based education applications are used on in-service education activities. Blended learning (b-learning) combines online learning with face-to-face learning. The goal of blended learning is to provide the most efficient and effective learning experience by combining learning environments. The purpose of this research is to find out the effect of b-learning approach on developing administrators’ leadership skills. To identify what the school administrators’ educational needs and to know their existing leadership skills, needs assessment questionnaire was applied to 72 school administrators who were selected from 33 primary schools in 11 region of Ankara capital city. According to the descriptive statistical analysis results of questionnaire, in-service training programme was prepared for the development of school administrators’ leadership skills. The school administrators were separated into three groups as computer based learning (CBL) (25 participants), blended learning (BL) (23 participants) and traditional learning (TL) (24 participant) groups. These groups were trained separately with these three different learning environments by using the in-service training programme. According to the results of pre-test, post test and achievements score means, it was observed that BL groups’ score is the highest when compared to TL and CBL groups. As a result of this research, in terms of achievements and effectiveness, b-learning was found to be the most effective learning environment when compared to the others. Both learners and tutors findings strongly suggest that blended learning is available alternative delivery method for inservice education activities.1 Keywords: Blended Learning, e-Learning, Information Technology, In-service education



1 INTRODUCTION Blended Learning (b-Learning or Hybrid Learning) consists of the combination of e-Learning and traditional education approach. Blended learning combines online learning with face-to-face learning. The goal of blended learning is to provide the most efficient and effective learning experience by combining different learning environments. b-Learning stands in the forefront in respect of interactivity with target learner group, enriching learning process and integration of technology into education [1,2,3,16,21]. E-learning has had an interesting impact on the learning environment. Blended learning is the most logical and natural evolution of our learning agenda. It suggests an elegant solution to the challenges of tailoring learning and development. It represents an opportunity to integrate the



innovative and technological advances offered by online learning with the interaction and participation offered in the best of the traditional learning [20]. The ground of blended learning approach constitutes the powerfull side of traditional education and computer based educations instead of using one or the other on its own. Basic characteristics of Blended learning which reflects the values of 21st century education are [2];  Providing a new way of learning and teaching,  Teaching how to learn,  Creating digital learners,  Be more economical,  Focusing on technology and communication  Improving project-based learning,  And improving teaching process.



1 This research project article has been supported by The Scientific and Technological Research Council of Turkey (TÜBİTAK) (SOBAG 1001 Programme).



UbiCC Journal – Volume 4 No. 3



538



Special Issue on ICIT 2009 Conference - Applied Computing



Blended Learning practices provide project based learning opportunities for active learning and interaction among learners and especially provides as a way to meet the educational needs of the learners. Blended learning programs may include several forms of learning tools, such as real-time virtual/collaboration software, self-paced web-based courses, electronic performance support systems (EPSS) embedded within the learning-task environment, and knowledge management systems. Blended learning contains various event-based activities, including face-to-face learning, elearning, and self-paced learning activities. Blended learning often occurs as a mixture of traditional instructor-led training, synchronous online training, asynchronous self-paced study, and structured task based training from a teacher or mentor. The aim of blended learning is to combine the best of classroom face-to-face learning experiences with the best of online learning experiences. Overall, blended learning refers to the integration (or the so-called blending) of e-learning tools and techniques with traditional face-to-face teaching delivery methods. The two important factors here are the time spent on online activities and the amount of technology utilized, see Concept of Blended Learning figure 1 below: [3,4,6,7,8,9,10,11,12,15,16,19].



course materials. While such uses may be unique and engaging, they are not exactly novel [13].



Figure 2: A Blend of Learning Theories By applying learning theories of Keller, Gagné, Bloom, Merrill, Clark and Gery, (see Figure 2) five key ingredients emerge as important elements of a blended learning process (see Figure 2): 1. Live Events: Synchronous, instructor-led learning events in which all learners participate at the same time, such as in a live “virtual classroom.” 2. Self-Paced Learning: Learning experiences that the learner completes individually, at his own speed and on his own time, such as interactive, Internet-based or CD-ROM training. 3. Collaboration: Environments in which learners communicate with others, for example, e-mail, threaded discussions or online chat. 4. Assessment: A measure of learners’ knowledge. Preassessments can come before live or self-paced events, to determine prior knowledge, and post-assessments can occur following live or self-paced learning events, to measure learning transfer. 5. Performance Support Materials: On-the-job reference materials that enhance learning retention and transfer, including PDA downloads, and printable references, summaries, and job aids. 2

PURPOSE



Fig. 1: Concept of Blended Learning If two or more of these learning environments which are stated above are used to teach an educational objective, it can be said that blended learning is realized. However blended learning has more meaning than showing a web page during a lesson in the classroom and using information immediately in the web page to explain the lesson. Blended learning is a learning of environment which combines environments of face to face learning and web-based distance learning. Blended learning overcomes this limitation of an elearning only approach [12]. Today blended learning primarily functions as a replacement for extension of face-to face environments. For instance, it might be used to foster learning communities, extend training events, offer follow-up resources in a community of practice, access guest experts, provide timely mentoring or coaching, present online lab or simulation activities, and deliver prework or supplemental



The purpose of this research is to find out the effects of b-learning approach on developing school administrators’ leadership skills. 3 RESEARCH DESIGN To determine what the school administrators’ educational needs on leadership skills, needs assessment questionnaire was applied to 72 school administrators who were selected from 33 primary schools in 11 regions within Ankara capital city. According to the results of this questionnaire, in-service



UbiCC Journal – Volume 4 No. 3



539



Special Issue on ICIT 2009 Conference - Applied Computing



training programme on developing school administrator’s leadership skill was prepared. The most needed leadership skills of school administrators according to the results of needs assessment were determined as human relations in administration, basic management skills for school principles, job satisfaction at organizations and motivation. After that, content and learning activities of "School Administrators Leadership Skills Development In-service Programme" were prepared. Beside that course notes as training materials were prepared to be distributed to the participants in the form of CDROM and printed documents. The school administrators were separated into three groups as Computer Based Learning (CBL) (25 participants), Blended Learning (BL) (23 participant) and Traditional Learning (TL) (24 participant) groups. These groups were trained according to three different methods by preparing education programme. The groups were given two days course. Before the in-service training the school administrators who were in BL group reached the digital content and studied learning activities included in "School Administrators Leadership Skills Development In-service Programme" which is prepared by using Moodle Learning Managing System Softwware and published on http://beg.meb.gov.tr:8088/ website. The school administrators who are in the BL group were entered to the http://beg.meb.gov.tr:8088/ webpage by using their usernames and passwords given to them three weeks ago, before the in-service training. The interface of the website is shown in the Fig. 2. The school administrators in this group shared information, chatted, and studied activities with their colleagues and subject area specialist about the content and learning activities included in the site whenever they want. As online learner, school administrators build their confidence and learning processes as they get used to working independently online. Blended learning activities included online knowledge gathering and construction in teams or groups, publishing of electronic content, interactive elements like online brainstorming, discussion, several forms of feedback, evaluation and assessment, as well as other blended learning techniques. Lecturers posted messages to the BL group as a whole and to each administrators individually to meet their need for support. They posted explanation to guide learners in more complex tasks, encouraged them to communicate, to do their individual assignments, and to use the Moddle platform tools. They have at their disposal to facilitate their work. Tutors controlled and marked the online assignments, filled in learners’ performance reports, and write feedback on their performance in their online portfolios. Lecturers followed school administrators learning improvements and gave encouragement when motivation level began to falter. And after that this group was trained by lecturer as subject area specialist. Lecturer trained this group by using face to face education, computer based education and online training website prepared by moodle software.



Figure. 3: The Moodle interface On the other hand; all the in-service training content and activities were taught to CBL group by lecturer with aid of computer and projector. Finally, TL group was trained in a traditional way by using blackboard Multiple choice test which was made up of 20 questions were applied to the groups to investigate their achievements on leadership skills. This test was shown to content experts to identify its content validity. To find out the statistical significant difference among three groups score means, oneway Anova and Scheffe test were used. This test was applied to all groups as pre-test at the beginning and as post-test at the end of in-service training [5]. Blended Learning Model which was used on the research process showed Figure 3.



Figure. 3: The Process of Blended Learning Model



UbiCC Journal – Volume 4 No. 3



540



Special Issue on ICIT 2009 Conference - Applied Computing



4 RESULTS When three groups’ pre-test score means were compared, it was seen that there were significant differences among them (F (2-69)=53,350, p ti)  T}  = {(  ti)  T}  = {(  ti)  T} Time Interval  ={( = i)  T}  = {( , i, …, j, >,



where Allocation i, j is the processor node i, j for Task 1, N; Start 1, N, End 1, N – run time and stop time for Task 1, N execution. Time interval [Start, End] is treated as so called walltime (WT), defined at the resource reservation time [15] in the local batch-job management system. Figure 1 shows some examples of job graphs in strategies with different degrees of distribution, task details, and data replication policies [19]. The first type strategy S1 allows scheduling with fine-grain computations and multiple data replicas, the second type strategy S2 is one with fine-grain computations and a bounded number of data replicas, and the third type S3 implies coarse-grain computations and constrained data replication. The vertices P1, …, P6, P23, and P45 correspond to tasks, while D1, …, D8, D12, D36, and D78 correspond to data transmissions. The transition from graph G1 to graphs G2 and G3 is performed through lumping of tasks and reducing of the parallelism level. The job graph is parameterized by prior estimates of the duration Tij of execution of a task Pi for a



UbiCC Journal – Volume 4 No. 3



560



Special Issue on ICIT 2009 Conference - Applied Computing



processor node nj of the type j, of relative volumes Vij of computations on a processor (CPU) of the type j, etc. (Table 1). It is to mention, such estimations are also necessary in several methods of priority scheduling including backfilling in Maui cluster scheduler.

P2 D1 P1 P3 D2 D6 P2 P4 D3 D4 D5 P5 D8 G2 P4 D7 P6 G1



The processor node load level LLj is the ratio of the total time of usage of the node of the type j to the job run time. Schedules in Fig. 2, b and Fig. 2, c are related to strategies S2 and S3. 2.2 Critical Works Method Strategies are generated with a critical works method [20]. The gist of the method is a multiphase procedure. The first step of any phase is scheduling of a critical work – the longest (in terms of estimated execution time Tij for task Pi) chain of unassigned tasks along with the best combination of available resources. The second step is resolving collisions cased by conflicts between tasks of different critical works competing for the same resource. (a)

n1 n2 n3 n4 Nodes P1 P2 P3 P6 P2 P3 P4 P5 P2 P3 P1 5 10 15 P4 P5 P4 P5 LL1=0.35 LL2=0.10 LL3=0.15 LL4=0.50 CF=41 P6 LL1=0.35 LL2=0 LL3=0.65 LL4=0.50 CF=37 P6 LL1=0.35 LL2=0.10 LL3=0.15 LL4=0.50 CF=41 20 Time



P1 D12 P3 D12 P1 P23 D36 P45 D36 P5 D78 D78



P6



G3



P6



Figure 1: Examples of job graphs.



Figure 2 shows fragments of strategies of types S1, S2, and S3 for jobs in Fig. 1. The duration of all data transmissions is equal to one unit of time for G1, while the transmissions D12 and D78 require two units of time and the transmission D36 requires four units of time for G2 and G3. We assume that the lumping of tasks is characterized by summing of the values of corresponding parameters of constituent subtasks (see Table 1). Table 1: User's task estimations. Tij, Vij Ti1 Ti2 Ti3 Ti4 Vij P1 2 4 6 8 20 P2 3 6 9 12 30 Tasks P3 P4 1 2 2 4 3 6 4 8 10 20 P5 1 2 3 4 10 P6 2 4 6 8 20



Nodes n1 P1 n2 n3 n4 Nodes n1 n2 n3 n4 0

Nodes n1 P1 n2 n3 n4 0 Nodes n1 P1 n2 n3 n4 0



(b)

P4



P2 P3



P6



P5 5 10 15



LL1=0.60 LL2=0 LL3=0.15 LL4=0.20 CF=39 20 Time



(c)

P45



P23



P6



LL1=1 LL2=0 LL3=0 LL4=0 CF=25 20 Time



5



10



15



Supporting schedules in Fig. 2, a present a subset of a Pareto-optimal strategy of the type S1 for tasks Pi, i=1, …, 6 in G1. The Pareto relation is generated by a vector of criteria CF, LLj, j=1, …, 4. A job execution cost-function CF is equal to the sum of Vij/Ti, where Ti is the real load time of processor node j by task Pi rounded to nearest notsmaller integer. Obviously, actual solving time Ti for a task can be different from user estimation Tij (see Table 1).



Figure 2: Fragments of scheduling strategies S1 (a), S2 (b), S3 (c). For example, there are four critical works 12, 11, 10, and 9 time units long (including data transfer time) on fastest processor nodes of the type 1 for the job graph G1 in Fig. 1, a (see Table 1): (P1, P2, P4, P6), (P1, P2, P5, P6), (P1, P3, P4, P6), (P1, P3, P5, P6). The schedule with CF=37 has a collision (see Fig. 2, a), which occurred due to simultaneous attempts of



UbiCC Journal – Volume 4 No. 3



561



Special Issue on ICIT 2009 Conference - Applied Computing



tasks P4 and P5 to occupy processor node n4. This collision is further resolved by the allocation of P4 to the processor node n3 and P5 to the node n4. Such reallocations can be based on virtual organization economics – in order to take higher performance processor node, user should “pay” more. Cost-functions can be used in economical models [14] of resource distribution in virtual organizations. It is worth noting that full costing in CF is not calculated in real money, but in some conventional units (quotas), for example like in corporate non-commercial virtual organizations. The essential point is different – user should pay additional cost in order to use more powerful resource or to start the task faster. The choice of a specific schedule from the strategy depends on the state and load level of processor nodes, and data storage policies. 2.3 Examples of Scheduling Strategies Let us assume that we need to construct a conditionally optimal strategy of the distribution of processors according to the main scheme of the critical works method from [20] for a job represented by the information graph G1 (see Fig. 1). Prior estimates for the duration Tij of processing tasks P1, …, P6 and relative computing volumes Vij for four types of processors are shown in Table 1, where i = 1, …, 6; j = 1, …, 4. The number of processors of each type is equal to 1. The duration of all data exchanges D1, …, D8 is equal to one unit of time. The walltime is given to be WT = 20. The criterion of resource-use efficiency is a cost function CF. We take a prior estimate for the duration Tij that is the nearest to the limit time Ti for the execution of task Pi on a processor of type j, which determines the type j of the processor used. Table 2: The strategy of the type MS1.

Schedule 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Duration T1 2 2 10 2 2 2 10 2 10 2 10 2 2 10 T2 3 3 3 3 3 11 3 11 3 11 3 3 3 3 T3 3 3 3 3 3 11 3 11 3 11 3 3 3 3 T4 2 10 2 2 10 2 2 2 2 2 2 2 10 2 T5 2 10 2 2 10 2 2 2 2 2 2 2 10 2 T6 10 2 2 10 2 2 2 2 2 2 2 10 2 2 A1 1 1 4 1 1 1 1 1 2 1 3 1 1 4 A2 1 1 1 1 1 4 1 4 1 3 1 1 1 1



The conflicts between competing tasks are resolved through unused processors, which, being used as resources, are accompanied with a minimum value of the penalty cost function that is equal to the sum of Vij/Tij (see Table 1) for competing tasks. It is required to construct a strategy that is conditionally minimal in terms of the cost function CF for the upper and lower boundaries of the maximum range for the duration Tij of the execution of each task Pi (see Table 1). It is a modification of the strategy S1 with fine-grain computations, active data replication policy, and the best- and worst execution time estimations. The strategy with a conditional minimum with respect to CF is shown in Table 2 by schedules 1, 2, and 3 (Ai is allocation of task Pi, i = 1, …, 6) and the scheduling diagrams are demonstrated in Fig. 2, a. The strategies that are conditionally maximal with respect to criteria LL1, LL2, LL3, and LL4 are given in Table 2 by the cases 4-7; 8, 9; 10, 11; and 12-14, respectively. Since there are no conditional branches in the job graph (see Fig. 1), LLj is the ratio of the total time of usage of a processor of type j to the walltime WT of the job completion. The Pareto-optimal strategy involves all schedules in Table 2. The schedules 2, 5, and 13 have resolved collisions between tasks P4 and P5. Let us assume that the load of processors is such that the tasks P1, P2, and P3 can be assigned with no more than three units of time on the first and third processors (see Table 2). The metascheduler runs through the set of supporting schedules and chooses a concrete variant of resource distribution that depends on the actual load of processor nodes.



Allocation A3 3 3 3 3 3 1 3 2 2 4 3 3 3 3 A4 1 3 1 1 4 1 1 1 1 1 1 1 3 1 A5 2 4 2 2 1 2 2 2 2 2 2 2 4 2 A6 4 1 1 1 1 1 1 1 1 1 1 4 1 1 CF 41 37 41 41 38 39 41 39 41 41 41 41 39 41 LL1 0.35 0.35 0.35 0.85 0.85 0.85 0.85 0.30 0.35 0.30 0.35 0.35 0.35 0.35



Criteria LL2 0.10 0 0.10 0.10 0 0.10 0.10 0.65 0.75 0.10 0.10 0.10 0 0.10 LL3 0.15 0.65 0.15 0.15 0.15 0 0.15 0 0 0.55 0.60 0.15 0.65 0.15 LL4 0.50 0.50 0.50 0 0.50 0.55 0 0.55 0 0.55 0 0.50 0.50 0.50



UbiCC Journal – Volume 4 No. 3



562



Special Issue on ICIT 2009 Conference - Applied Computing



Then, the metascheduler should choose the schedules 1, 2, 4, 5, 12, and 13 as possible variants of resource distribution. However, the concrete schedule should be formulated as a resource request and implemented by the system of batch processing subject to the state of all four processors and possible runtimes of tasks P4, P5, and P6 (see Table 2). Suppose that we need to generate a Paretooptimal strategy for the job graph G2 (see Fig. 1) in the whole range of the duration Ti of each task Pi, while the step of change is taken to be no less than the lower boundary of the range for the most performance processor. The Pareto relation is generated by the vector of criteria CF, LL1, … , LL4. The remaining initial conditions are the same as in the previous example. The strategies that are conditionally optimal with respect to the criteria CF, LL1, LL2, LL3, and LL4 Table 3: The strategy of the type S2.

Schedule 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 Duration T1 2 4 4 5 2 2 2 2 4 4 4 5 2 4 4 4 5 2 2 2 2 4 4 4 5 2 2 2 2 4 4 4 5 T2 3 3 3 3 3 3 3 6 3 3 4 3 6 3 3 4 3 3 3 3 6 3 3 4 3 3 3 3 6 3 3 4 3 T3 3 3 3 3 3 3 3 6 3 3 4 3 6 3 3 4 3 3 3 3 6 3 3 4 3 3 3 3 6 3 3 4 3 T4 4 2 3 2 2 4 5 2 2 3 2 2 2 2 3 2 2 2 4 5 2 2 3 2 2 2 4 5 2 2 3 2 2 T5 4 2 3 2 2 4 5 2 2 3 2 2 2 2 3 2 2 2 4 5 2 2 3 2 2 2 4 5 2 2 3 2 2 T6 3 3 2 2 5 3 2 2 3 2 2 2 2 3 2 2 2 5 3 2 2 3 2 2 2 5 3 2 2 3 2 2 2 A1 1 2 2 2 1 1 1 1 1 1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 A2 1 1 1 1 1 1 1 2 1 1 1 1 2 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 2 1 1 1 1



are presented in Table 3 by the schedules 1-4, 5-12, 13-17, 18-25, and 26-33, respectively. The Paretooptimal strategy does not include the schedules 2, 5, 12, 14, 16, 17, 22, and 30. Let us consider the generation of a strategy for the job represented structurally by the graph G3 in Fig. 1 and by summing of the values of the parameters given in Table 1 for tasks P2, P3 and P4, P5. As a result of the resource distribution for the model G3, the tasks P1, P23, P45, and P6 turn out to be assigned to one and the same processor of the first type. Consequently, the costs of data exchanges D12, D36, and D78 can be excluded. Because there can be no conflicts in this case between processing tasks (see Fig. 1), the scheduling obtained before the exclusion of exchange procedures can be revised.



Allocation A3 3 3 3 3 3 3 3 1 3 3 4 3 4 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 4 3 3 4 3 A4 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 1 1 1 1 1 1 2 2 1 1 1 1 1 A5 4 2 3 2 2 4 4 2 2 3 2 2 2 2 2 2 2 2 3 3 2 2 3 2 2 2 4 4 2 2 3 2 2 A6 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 CF 39 41 40 43 43 39 41 42 41 40 41 43 43 41 40 41 43 43 39 40 42 41 40 41 43 43 39 40 42 41 40 41 43 LL1 0.40 0.40 0.40 0.35 0.60 0.60 0.60 0.60 0.60 0.60 0.60 0.60 0.30 0.40 0.40 0.40 0.35 0.35 0.40 0.35 0.30 0.40 0.40 0.40 0.35 0.35 0.40 0.35 0.30 0.40 0.40 0.40 0.35



Criteria LL2 0.20 0.30 0.20 0.35 0.10 0 0 0.30 0.10 0 0.10 0.10 0.40 0.45 0.50 0.50 0.50 0.35 0.20 0.25 0.40 0.30 0.20 0.30 0.35 0.35 0.20 0.25 0.40 0.30 0.20 0.30 0.35 LL3 0.15 0.15 0.30 0.15 0.15 0.15 0.15 0 0.15 0.30 0 0.15 0 0 0 0 0 0.15 0.35 0.40 0.30 0.15 0.30 0.20 0.15 0.15 0.15 0.15 0 0.15 0.30 0 0.15 LL4 0.20 0 0 0 0 0.20 0.25 0 0 0 0.20 0 0.30 0 0 0 0 0 0 0 0 0 0 0 0 0 0.20 0.25 0.30 0 0 0.20 0



UbiCC Journal – Volume 4 No. 3



563



Special Issue on ICIT 2009 Conference - Applied Computing



Table 4: The strategy of the type S3. ScheDuration dule T1 T23 T45 T6 1 2 8 6 4 2 4 8 3 5 3 6 4 6 4 4 8 4 3 5 5 10 4 3 3 6 11 4 3 2



Allocation A1 1 1 1 1 1 1 A23 1 1 1 1 1 1 A45 1 1 1 1 1 1 A6 1 1 1 1 1 1 CF 25 24 24 27 29 32 LL1 1 1 1 1 1 1



Criteria LL2 0 0 0 0 0 0 LL3 0 0 0 0 0 0 LL4 0 0 0 0 0 0



The results of distribution of processors are presented in Table 4 (A23, A45 are allocations, and T23, T45 are run times for tasks P23 and P45). Schedules 1-6 in Table 4 correspond to the strategy that is conditionally minimal with respect to CF with LL1 = 1. Consequently, there is no sense in generating conditionally maximal schedules with respect to criteria LL1, …, LL4. 2.4 Coordinated Scheduling with the Critical Works Method The critical works method was developed for application-level scheduling [19, 20]. However, it can be further refined to build multifactor and multicriteria strategies for job-flow distribution in virtual organizations. This method is based on dynamic programming and therefore uses some integral characteristics, for example total resource usage cost for the tasks that compose the job. However the method of critical works can be referred to the priority scheduling class. There is no conflict between these two facts, because the method is dedicated for task co-allocation of compound jobs. Let us consider a simple example. Fig. 3 represents two jobs with walltimes WT1 = 110 and WT2 = 140 that are submitted to the distributed environment with 8 CPUs. If the jobs are submitted one-by-one the metascheduler (Section 3) will also schedule them one-by-one and will guarantee that every job will be scheduled within the defined time interval and in most efficient way in terms of a selected cost function and maximize average load balance of CPUs on a single job scale (Fig. 4). Job-flow execution will be finished at WT3 = 250. This is an example of application-level scheduling and no integral jobflow characteristics are optimized in this case. To combine application-level scheduling and job-flow scheduling and to fully exploit the advantages of the approach proposed, one can submit both jobs simultaneously or store them in buffer and execute the scheduling for all jobs in the buffer after a certain amount of time (buffer time). If the metascheduler gets more than one job to



schedule it runs the developed mechanisms that optimize the whole job-flow (two jobs in this example). In that case the metascheduler will still try to find an optimal schedule for each single job as described above and, at the same time, it will try to find the most optimal job assignment so that the average load of CPUs will be maximized on a jobflow scale.



Figure 3: Sample jobs. Fig. 5 shows, that both jobs are executed within WT4 = WT2 = 140, every data dependency is taken into account (e.g. for the second job: task P2 is executed only after tasks P0, P4, and P1 are ready), the final schedule is chosen from the generated strategy with the lowest cost function. Priority scheduling based on queues is not an efficient way of multiprocessor jobs co-allocating, in our opinion. Besides, there are several wellknown side effects of this approach in the cluster systems such as LL, NQE, LSF, PBS and others.



UbiCC Journal – Volume 4 No. 3



564



Special Issue on ICIT 2009 Conference - Applied Computing



For example, traditional First-Come-First-Serve (FCFS) strategy leads to idle standing of the resources. Another strategy, which involves job ranking according to the specific properties, such as computational complexity, for example LeastWork-First (LWF), leads to a severe resource fragmentation and often makes it impossible to execute some jobs due to the absence of free resources. In distributed environments these effects can lead to unpredictable job execution time and thereby to unsatisfactory quality of service.



Figure 4: Consequential scheduling. In order to avoid it many projects have components that make schedules, which are supported by preliminary resource reservation mechanisms [15, 16].



One to mention is Maui cluster scheduler, where backfilling algorithm is implemented. Remote Grid resource reservation mechanism is also supported in GARA, Ursala and Silver projects [16]. Here, only one variant of the final schedule is built and it can become irrelevant because of changes in the local job-queue, transporting delays etc. The strategy is some kind of preparation of possible activities in distributed computing based on supporting schedules (see Fig. 2, Tables 2, 3 and 4) and reactions to the events connected with resource assignment and advance reservations [15, 16]. The more factors considered as formalized criteria are taken into account in strategy generation, the more complete is the strategy in the sense of coverage of possible events [18, 19]. The choice of the supporting schedule [20] depends on the utilization state of processor nodes, data storage and relocation policies specific to the environment, structure of the jobs themselves and user estimations of completion time and resource requirements. It is important to mention that users can submit jobs without information about the task execution order as required by existing schedulers like Maui cluster scheduler were only queues are supported. Implemented mechanisms of our approach support a complex structure for the job, which is represented as a directed graph, so user should only provide data dependencies between tasks (i.e. the structure of the job). The metascheduler will generate the schedules to satisfy their needs by providing optimal plans for jobs (application-level scheduling) and the needs for the resource owners by optimizing the defined characteristics of the jobflow for the distributed system (job-flow scheduling). 3 METASCHEDULING FRAMEWORK



Figure 5: Job-flow scheduling.



In order to implement the effective scheduling and allocation to heterogeneous resources, it is very important to group user jobs into flows according to the strategy type selected and to coordinate jobflow and application-level scheduling. A hierarchical structure (Fig. 6) composed of a jobflow metascheduler and subsidiary job managers, which are cooperating with local batch-job management systems, is a core part of a scheduling framework proposed in this paper. It is assumed that the specific supporting schedule is realized and the actual allocation of resources is performed by the system of batch processing of jobs. This schedule is implemented on the basis of a user resource request with a requirement to the types and characteristics of resources (memory and processors) and to the system software as well as generated, for example, by the script of the job entry instruction qsub. Therefore, the formation and support of scheduling



UbiCC Journal – Volume 4 No. 3



565



Special Issue on ICIT 2009 Conference - Applied Computing



strategies should be conducted by the metascheduler, an intermediary link between the job flow and the system of batch processing.

Job-flows i j Metascheduler k



Job manager for strategy Sk



Job manager for strategies Si, Sj



Computer nodes



Job manager for strategy Si



Computer nodes



Computer node domains



Figure 6: framework.



Components



of



metascheduling



The advantages of hierarchically organized resources managers are obvious, e.g., the hierarchical job-queue-control model is used in the GrADS metascheduler [13] and X-Com system [2]. Hierarchy of intermediate servers allows decreasing idle time for the processor nodes, which can be inflicted by transport delays or by unavailability of the managing server while it is dealing with the other processor nodes. Tree-view manager structure in the network environment of distributed computing allows avoiding deadlocks when accessing resources. Another important aspect of computing in heterogeneous environments is that processor nodes with the similar architecture, contents, administrating policy are grouped together under the job manager control. Users submit jobs to the metascheduler (see Fig. 6) which distributes job-flows between processor node domains according to the selected scheduling and resource co-allocation strategy Si, Sj or Sk. It does not mean, that these flows cannot “intersect” each other on nodes. The special reallocation mechanism is provided. It is executed on the higherlevel manager or on the metascheduler-level. Job managers are supporting and updating strategies based on cooperation with local managers and simulation approach for job execution on processor nodes. Innovation of our approach consists in mechanisms of dynamic job-flow environment reallocation based on scheduling strategies. The nature of distributed computational environments itself demands the development of multicriteria and multifactor strategies [17, 18] of coordinated scheduling and resource allocation. The dynamic configuration of the environment, large number of resource reallocation events, user’s



and resource owner’s needs as well as virtual organization policy of resource assignment should be taken into account. The scheduling strategy is formed on a basis of formalized efficiency criteria, which sufficiently allow reflecting economical principles [14] of resource allocation by using relevant cost functions and solving the load balance problem for heterogeneous processor nodes. The strategy is built by using methods of dynamic programming [20] in a way that allows optimizing scheduling and resource allocation for a set of tasks, comprising the compound job. In contrast to previous works, we consider the scheduling strategy as a set of admissible supporting schedules (see Fig. 2, Tables 2 and 3). The choice of the specific variant depends on the load level of the resource dynamics and is formed as a resource query, which is sent to a local batch-job processing system. One of the important features of our approach is resource state forecasting for timely updates of the strategies. It allows implementing mechanisms of adaptive job-flow reallocation between processor nodes and domains, and also means that there is no more fixed task assignment on a particular processor node. While one part of the job can be sent for execution, the other tasks, comprising the job, can migrate to the other processor nodes according to the updated co-allocation strategy. The similar schedule correction procedure is also supported in the GrADS project [13], where multistage job control procedure is implemented: making initial schedule, its correction during the job execution, metascheduling for a set of applications. Downside of this approach is the fact, that it is based on the creation of a single schedule, so the metascheduler stops working when no additional resources are available and job-queue is then set to waiting mode. The possibility of strategy updates allows user, being integrated into economical conditions of virtual organization, to affect job start time by changing resource usage costs. In fact it means that the job-flow dispatching strategy is modified according to new priorities and this provides competitive functioning and dynamic jobflow balance in virtual organization with inseparable resources. 4 SIMULATIONS RESULTS STUDIES AND



4.1 Simulation System We have implemented an original simulation environment (Fig. 7) of the metascheduling framework (see Fig. 6) to evaluate efficiency indices of different scheduling and co-allocation strategies. In contrast to well-known Grid simulation systems such as ChicSim [12] or OptorSim [23], our simulator MetaSim generates



UbiCC Journal – Volume 4 No. 3



566



Special Issue on ICIT 2009 Conference - Applied Computing



multicriteria strategies as a number of supporting schedules for metascheduler reactions to the events connected with resource assignment and advance reservations. Strategies for more than 12000 jobs with a fixed completion time were studied. Every task of a job had randomized completion time estimations, computation volumes, data transfer times and volumes. These parameters for various tasks had difference which was equal to 2, ..., 3. Processor nodes were selected in accordance to their relative performance. For the first group of “fast” nodes the relative performance was equal to 0.66, …, 1, for the second and the third groups 0.33, …, 0.66 and 0.33 (“slow” nodes) respectively. A number of nodes was conformed to a job structure, i.e. a task parallelism degree, and was varied from 20 to 30. 4.2 Types of Strategies We have studied the strategies of the following types: S1 – with fine-grain computations and active data replication policy; S2 – with fine-grain computations and a remote data access; S3 – with coarse-grain computations and static data storage; MS1 – with fine-grain computations, active data replication policy, and the best- and worst execution time estimations (a modification of the strategy S1). The strategy MS1 is less complete than the strategy S1 or S2 in the sense of coverage of events in distributed environment (see Tables 2 and 3). However the important point is the generation of a strategy by efficient and economic computational procedures of the metascheduler. The type S1 has



more computational expenses than MS1 especially for simulation studies of integrated job-flow and application-level scheduling. Therefore, in some experiments with integrated scheduling we compared strategies MS1, S2, and S3. 4.3 Application-Level Scheduling Study We have conducted the statistical research of the critical works method for application-level scheduling with above-mentioned types of strategies S1, S2, S3. The main goal of the research was to estimate a forecast possibility for making application-level schedules with the critical works method without taking into account independent job flows. For 12000 randomly generated jobs there were 38% admissible solutions for S1 strategy, 37% for S2, and 33% for S3 (Fig. 8). This result is obvious: application-level schedules implemented by the critical works method were constructed for available resources non-assigned to other independent jobs. Along with it there is a conflict distribution for the processor nodes that have different performance (“fast” are 2-3 times faster, than “slow” ones): 32% for “fast” ones, 68% for “slow” ones in S1, 56% and 44% in S2, 74% and 26% for S3 (Fig. 9). This may be explained as follows. The higher is the task state of distribution in the environment with active data transfer policy, the lower is the probability of collision between tasks on a specific resource. In order to implement the effective scheduling and resource allocation policy in the virtual organization we should coordinate application and job-flow levels of the scheduling.



Figure 7: Simulation environment of hierarchical scheduling framework based on strategies.



UbiCC Journal – Volume 4 No. 3



567



Special Issue on ICIT 2009 Conference - Applied Computing



S1



S1



S2



S2



S3 Figure 8: Percentage of admissible application-level schedules. 4.4 Job-Flow and Application-Level Scheduling Study For each simulation experiment such factors as job completion “cost”, task execution time, scheduling forecast errors (start time estimation), strategy live-to-time (time interval of acceptable schedules in a dynamic environment), and average load level for strategies S1, MS1, S2, and S3 were studied. Figure 10 shows load level statistics of variable performance processor nodes which allows discovering the pattern of the specific resource usage when using strategies S1, S2, and S3 with coordinated job-flow and application-levels scheduling. The strategy S2 performs the best in the term of load balancing for different groups of processor nodes, while the strategy S1 tries to occupy “slow” nodes, and the strategy S3 - the processors with the highest performance (see Fig. 10).



S3 Figure 9: Percentage of collisions for “fast” processor nodes in application-level scheduling.

Average node load level, % 80 60 40 20 0



S1 0.66-1



S2 0.33-0.66



S3 0.33



Relative processor nodes performance



Figure 10: Processor node load level in strategies S1, S2, and S3. Factor quality analysis of S2, S3 strategies for the whole range of execution time estimations for the



UbiCC Journal – Volume 4 No. 3



568



Special Issue on ICIT 2009 Conference - Applied Computing



selected processor nodes as well as modification MS1, when best- and worst-case execution time estimations were taken, is shown in Figures 11 and 12.

Relative job completion cost 1 Relative task execution time 1



0.5



0.5



0 MS1 Сost S2 S3 Execution time



0



Figure 11: Job completion cost and task execution time in strategies MS1, S2, and S3. Lowest-cost strategies are the “slowest” ones like S3 (see Fig. 11); they are most persistent in the term of time-to-live as well (see Fig. 12).

Relative time-to-live 1 Start time deviation/ job run time 1



0.5



0.5



0 MS1 Time-to-live S2 S3 Deviation



0



Figure 12: Time-to-live and start deviation time in strategies MS1, S2, and S3. The strategies of the type S3 try to monopolize processor resources with the highest performance and to minimize data exchanges. Withal, less persistent are the “fastest”, most expensive and most accurate strategies like S2. Less accurate strategies like MS1 (see Fig. 12) provide longer task completion time, than more accurate ones like S2 (Fig. 11), which include more possible events, associated with processor node load level dynamics. 5 CONCLUSIONS AND FUTURE WORK



scenarios, e.g., in our experiments we use FCFS management policy in local batch-job management systems. Afore-cited research results of strategy characteristics were obtained by simulation of global job-flow in a virtual organization. Inseparability condition for the resources requires additional advanced research and simulation approach of local job passing and local processor nodes load level forecasting methods development. Different jobqueue management models and scheduling algorithms (FCFS modifications, LWF, backfilling, gang scheduling, etc.) can be used here. Along with it local administering rules can be implemented. One of the most important aspects here is that advance reservations have impact on the quality of service. Some of the researches (particularly the one in Argonne National Laboratory) show, that preliminary reservation nearly always increases queue waiting time. Backfilling decreases this time. With the use of FCFS strategy waiting time is shorter than with the use of LWF. On the other hand, estimation error for starting time forecast is bigger with FCFS than with LWF. Backfilling that is implemented in Maui cluster scheduler includes advanced resource reservation mechanism and guarantees resource allocation. It leads to the difference increase between the desired reservation time and actual job starting time when the local request flow is growing. Some of the quality aspects and job-flow load balance problem are associated with dynamic priority changes, when virtual organization user changes execution cost for a specific resource. All of these problems require further research. ACKNOWLEDGEMENT. This work was supported by the Russian Foundation for Basic Research (grant no. 09-01-00095) and by the State Analytical Program “The higher school scientific potential development” (project no. 2.1.2/6718). 6 [1] REFERENCES I. Foster, C. Kesselman, and S. Tuecke: The Anatomy of the Grid: Enabling Scalable Virtual Organizations, Int. J. of High Performance Computing Applications, Vol. 15, No. 3, pp. 200 – 222 (2001) V.V. Voevodin: The Solution of Large Problems in Distributed Computational Media, Automation and Remote Control, Vol. 68, No. 5, pp. 32 – 45 (2007) D. Thain, T. Tannenbaum, and M. Livny: Distributed Computing in Practice: the Condor Experience, Concurrency and Computation: Practice and Experience, Vol. 17, No. 2-4, pp. 323 - 356 (2004)



The related works in scheduling problems are devoted to either job scheduling problems or application-level scheduling. The gist of the approach described is that the resultant dispatching strategies are based on the integration of job-flows and application-level techniques. It allows increasing the quality of service for the jobs and distributed environment resource usage efficiency. Our results are promising, but we have bear in mind that they are based on simplified computation



[2]



[3]



UbiCC Journal – Volume 4 No. 3



569



Special Issue on ICIT 2009 Conference - Applied Computing



[4]



[5]



[6]



[7]



[8]



[9]



[10]



[11]



[12]



[13]



A. Roy and M. Livny: Condor and Preemptive Resume Scheduling, In: J. Nabrzyski, J.M. Schopf, and J.Weglarz (eds.): Grid resource management. State of the art and future trends, Kluwer Academic Publishers, pp. 135 – 144 (2003) V.V. Krzhizhanovskaya and V. Korkhov: Dynamic Load Balancing of Black-Box Applications with a Resource Selection Mechanism on Heterogeneous Resources of Grid, In: 9th International Conference on Parallel Computing Technologies, Springer, Heidelberg, LNCS, Vol. 4671, pp. 245 – 260 (2007) F. Berman: High-performance Schedulers, In: I. Foster and C. Kesselman (eds.): The Grid: Blueprint for a New Computing Infrastructure, Morgan Kaufmann, San Francisco, pp. 279 – 309 (1999) Y. Yang, K. Raadt, and H. Casanova: Multiround Algorithms for Scheduling Divisible Loads, IEEE Transactions on Parallel and Distributed Systems, Vol. 16, No. 8, pp. 1092 – 1102 (2005) A. Natrajan, M.A. Humphrey, and A.S. Grimshaw: Grid Resource Management in Legion,” In: J. Nabrzyski, J.M. Schopf, and J.Weglarz (eds.): Grid resource management. State of the art and future trends, Kluwer Academic Publishers, pp.145 – 160 (2003) J. Beiriger, W. Johnson, H. Bivens et al.: Constructing the ASCI Grid, In: 9th IEEE Symposium on High Performance Distributed Computing, IEEE Press, New York, pp. 193 – 200 (2000) J. Frey, I. Foster, M. Livny et al.: Condor-G: a Computation Management Agent for Multiinstitutional Grids, In: 10th International Symposium on High-Performance Distributed Computing, IEEE Press, New York, pp. 55 – 66 (2001) D. Abramson, J. Giddy, and L. Kotler: High Performance Parametric Modeling with Nimrod/G: Killer Application for the Global Grid?, In: International Parallel and Distributed Processing Symposium, IEEE Press, New York, pp. 520 – 528 (2000) K. Ranganathan and I. Foster: Decoupling Computation and Data Scheduling in Distributed Data-intensive Applications, In: 11th IEEE International Symposium on High Performance Distributed Computing, IEEE Press, New York, pp. 376 – 381 (2002) H. Dail, O. Sievert, F. Berman et al.: Scheduling in the Grid Application Development Software project, In: J. Nabrzyski, J.M. Schopf, and J.Weglarz (eds.): Grid resource management. State of the art and



[14]



[15]



[16]



[17]



[18]



[19]



[20]



[21]



[22]



[23]



future trends, Kluwer Academic Publishers, pp. 73 – 98 (2003) R. Buyya, D. Abramson, J. Giddy et al.: Economic Models for Resource Management and Scheduling in Grid Computing, J. of Concurrency and Computation: Practice and Experience, Vol. 14, No. 5, pp. 1507 – 1542 (2002) K. Aida and H. Casanova: Scheduling Mixedparallel Applications with Advance Reservations, In: 17th IEEE International Symposium on High-Performance Distributed Computing, IEEE Press, New York, pp. 65 – 74 (2008) D.B. Jackson: GRID Scheduling with Maui/Silver, In: J. Nabrzyski, J.M. Schopf, and J.Weglarz (eds.): Grid resource management. State of the art and future trends, Kluwer Academic Publishers, pp. 161 – 170 (2003) K. Kurowski, J. Nabrzyski, A. Oleksiak, and J. Weglarz: Multicriteria Aspects of Grid Resource Management, In: J. Nabrzyski, J.M. Schopf, and J.Weglarz (eds.): Grid resource management. State of the art and future trends, Kluwer Academic Publishers, pp. 271 – 293 (2003) V. Toporkov: Multicriteria Scheduling Strategies in Scalable Computing Systems, In: 9th International Conference on Parallel Computing Technologies, Springer, Heidelberg, LNCS, Vol. 4671, pp. 313 – 317 (2007) V.V. Toporkov and A.S. Tselishchev: Safety Strategies of Scheduling and Resource Coallocation in Distributed Computing, In: 3rd International Conference on Dependability of Computer Systems, IEEE CS Press, pp. 152 – 159 (2008) V.V. Toporkov: Supporting Schedules of Resource Co-Allocation for Distributed Computing in Scalable Systems, Programming and Computer Software, Vol. 34, No. 3, pp. 160 – 172 (2008) M. Tang, B.S. Lee, X. Tang, et al.: The Impact of Data Replication on Job Scheduling Performance in the Data Grid, Future Generation Computing Systems, Vol. 22, No. 3, pp. 254 – 268 (2006) N.N. Dang, S.B. Lim, and C.K. Yeo: Combination of Replication and Scheduling in Data Grids, Int. J. of Computer Science and Network Security, Vol. 7, No. 3, pp. 304 – 308 (2007) W.H. Bell, D. G. Cameron, L. Capozza et al.: OptorSim – A Grid Simulator for Studying Dynamic Data Replication Strategies, Int. J. of High Performance Computing Applications, Vol. 17, No. 4, pp. 403 – 416 (2003)



UbiCC Journal – Volume 4 No. 3



570



Special Issue on ICIT 2009 Conference - Applied Computing



Least and greatest fixed points of a while semantics function

Fairouz Tchier Mathematics department, King Saud University P.O.Box 22452 Riyadh 11495, Saudi Arabia ftchier@hotmail.com May 1, 2009



Abstract

The meaning of a program is given by specifying the function (from input to output) that corresponds to the program. The denotational semantic definition, thus maps syntactical things into functions. A relational semantics is a mapping of programs to relations. We consider that the input-output semantics of a program is given by a relation on its set of states. In a nondeterministic context, this relation is calculated by considering the worst behavior of the program (demonic relational semantics). In this paper, we concentrate on while loops. We will present some interesting results about the fixed points of the while semantics function; f (X) = Q ∨ P 2 X where P , are the monotypes defined by the equations (a) R = id ∧ 1 ◦ R. These operators can also be characterized by Galois connections(see [2, 2]). For each relation R and each monotype a, R ≤ a ⇔ R ≤ 1 ◦ a. The domain and codomain operators are linked by the equation R> = R , • (b) R\a := (R\(a 2 1)) ≤ a, • b ≤ R\a ⇔ (R ◦ b)< ≤ a. We have to use exhaustively the complement of the domain of a relation R, i.e the monotype a such that a = R< ∼ . To avoid the notation R< ∼ , we adopt the Notation 2



In the calculus of relations, there are two ways for viewing sets as relations; each of them has its own advantages. The first is via vectors: a relation x is



UbiCC Journal – Volume 4 No. 3



572



Special Issue on ICIT 2009 Conference - Applied Computing



R := R< ∼ . Because we assume our relation algebra to be complete, least and greatest fixed points of monotonic functions exist. We cite [12] as a general reference on fixed points. Let f be a monotonic function. The following properties of fixed points are used below: (a) µf = {X|f (X) = X} = {X|f (X) ≤ X}, (b) νf = {X|f (X) = X} = {X|X ≤ f (X)}, (c) µf ≤ νf, (d) f (Y ) ≤ Y ⇒ µf ≤ Y, (e) Y ≤ f (Y ) ⇒ Y ≤ νf. In what follows, we describe notions that are useful for the description of the set of initial states of a program for which termination is guaranteed. These notions are progressive finiteness and the initial part of a relation. A relation R is progressively finite in terms of points iff there are no infinite chains s0 , ..., si such that si Rsi+1 ∀i, i ≥ 0. I.e there is no points set y which are the starting points of some path of infinite length. For every point set y, y ≤ R ◦ y ⇒ y = 0. The least set of points which are the starting points of paths of finite length i.e from which we can proceed only finitely many steps is called initial part of R denoted by I(R). This topic is of interest in many areas of computer science, mathematics and is related to recursion and induction principle. (5) Definition. (a) The initial part of a relation R, denoted I(R), is given by • I(R) = {a | a ≤ id : a/R = a} = {a | • • a ≤ id : a/R ≤ a} = µ(a : a ≤ id : a/R), where a is a monotype. (b) A relation R is said to be progressively finite [28] iff I(R) = id.

• The description of I(R) by the formulation a/R = a • shows that I(R) exists, since (a | a ≤ id : a/R) is monotonic in the first argument and because the set of monotypes is a complete lattice, it follows from the fixed point theorem of Knaster and Tarski that this function has a least fixed point. Progressive finiteness of a relation R is the same as well-foundedness



of R . Then, I(R) is a monotype. In a concrete setting, I(R) is the set of monotypes which are not the origins of infinite paths (by R): A relation R is progressively finite iff for a monotype a, a ≤ (R ◦ a)< ⇒ a = 0 equivalently ν(a : a ≤ id : (R ◦ a)< ) = 0 equivalently µ(a : a ≤ • id : a/R) = id. The next theorem involves the function wa (X) := Q ∨ P ◦ X, which is closely related to the description of iterations. The theorem highlights the importance of progressive finiteness in the simplification of fixed point-related properties. (6) Theorem. Let f (X) := Q ∨ P ◦ X be a function. If P is progressively finite, the function f has a unique fixed point which means that ν(f ) = µ(f ) = P ∗ ◦ Q [1]: As the demonic calculus will serve as an algebraic apparatus for defining the denotational semantics of the nondeterministic programs, we will define in what follows these operators.



3



Demonic refinement ordering



We now define the refinement ordering (demonic inclusion) we will be using in the sequel. This ordering induces a complete join semilattice, called a demonic semilattice. The associated operations are demonic join ( ), demonic meet ( ) and demonic composition ( 2 ). We give the definitions and needed properties of these operations, and illustrate them with simple examples. For more details on relational demonic semantics and demonic operators, see [5, 8, 6, 7, 14]. (7) Definition. We say that a relation Q refines a relation R [23], denoted by Q R, iff R< ◦ Q ≤ < < R and R ≤ Q . (8) Proposition. Let Q and R be relations, then (a) The greatest lower (wrt ) of Q and R is, Q R = Q< ◦ R< ◦ (Q ∨ R), If Q< = R< then we have i.e Q R = Q ∨ R. 3 and ∨ coincide



UbiCC Journal – Volume 4 No. 3



573



Special Issue on ICIT 2009 Conference - Applied Computing



(b) If Q and R satisfy the condition Q< ∧ R< = (Q ∧ R)< , their least upper bound is Q R = Q ∧ R ∨ Q ◦ R ∨ R ◦ Q, otherwise, the least upper bound does not exist. If Q< ∧ R< = 0 then we have and ∧ coincide i.e Q R = Q ∧ R. For the proofs see [9, 14]. (9) Definition. The demonic composition of rela• tions Q and R [5] is Q 2 R = (R< /Q) ◦ Q ◦ R. In what follows we present some properties of (10) Theorem. (a) (P

2 2



• (a) S(R) = I(P ) ◦ [(P ∨ Q)< /P ∗ ] ◦ P ∗ ◦ Q., with the restriction



(b) P < ∧ Q< = 0 Our goal is to show that the operational semantics a is equal to the denotational one which is given as the greatest fixed point of the semantic function Q ∨ P 2 X in the demonic semilattice. In other words, we have to prove the next equation: (a) S(R) = {X|X Q∨P

2



.



X};



Q) 2 R = P



2



(Q 2 R),



(b) R total ⇒ Q 2 R = Q ◦ R, (c) Q function ⇒ Q 2 R = Q ◦ R. See [5, 6, 7, 14, 35]. Monotypes have very simple and convenient properties. Some of them are presented in the following proposition. (11) Proposition. Let a and b be monotypes. We have (a) a = a = a2 ,



by taking P := t 2 B and Q := t∼ , one gets the demonic semantics we have assigned to while loops in previous papers [14, 35]. Other similar definitions of while loops can be found in [19, 25, 29]. Let us introduce the following abbreviations: (12) Abbreviation. Let P , Q and X be relations subject to the restriction P < ∧ Q< = 0 (b) and x a monotype. The Abbreviations wd , wa , w< , a and l are defined as follows: wd (X) := Q ∨ P 2 X, • a := (P ∨ Q)< /P ∗ , wa (X) := Q ∨ P ◦ X, l := I(P ). w< (x) := Q< ∨ (P 2 x)< = Q ∨ (P 2 x)< (Mnemonics: the subscripts a and d stand for angelic and demonic, respectively; the subscript < refers to the fact that w< is obtained from wd by composition with <; the monotype a stands for abnormal, since it represents states from which abnormal termination is not possible; finally, l stands for loop, since it represents states from which no infinite loop is possible.) In what follows we will be concerned about the fixed point of wa , w< and wd . (13) Theorem. Every fixed point Y of wa (Abbreviation 12) verifies P ∗ ◦ Q ≤ Y ≤ P ∗ ◦ Q ∨ l∼ 2 1, and the bounds are tight (i.e. the extremal values are fixed points). The next lemma investigates the relationship between fixed points of w< and those of wd (cf. Abbreviation 12). 4



(b) a 2 b = a ∧ b = b 2 a, (c) a ∨ a∼ = id and a ∧ a∼ = 0, (d) a ≤ b ⇔ b∼ ≤ a∼ , (e) a∼ 2 b∼ = (a ∨ b)∼ , (f ) (a ∧ b)∼ = (a 2 b)∼ = a∼ ∨ b∼ , (g) a 2 b∼ ∨ b = a ∨ b, (h) a ≤ b ⇔ a 2 1 ≤ b 2 1. In previous papers [14, 13, 31, 35], we found the semantics of the while loop given by the following P    graph: - e
- s 

Q



UbiCC Journal – Volume 4 No. 3



574



Special Issue on ICIT 2009 Conference - Applied Computing



(14) Lemma. Let h(X) := (P ∨ Q) ∨ (P ◦ X)< and h1 (x) := (P ∨ Q) 2 1 ∨ P ◦ x. (a) Y = wd (Y ) ⇒ w< (Y < ) = Y < , (b) w< (Y < ) = Y < ⇒ h(Y ) = Y , (c) h(Y ) = Y ⇒ h1 (Y

2



4



Application



1) = Y



2



1,



(15) Lemma. Let Y be a fixed point of wd and b be a fixed point of w< (Abbreviation 12). The relation b 2 Y is a fixed point of wd . (16) Lemma. If Y and Y are two fixed points of wd (Abbreviation 12) such that Y < = Y < and Y < ◦P is progressively finite, then Y = Y . The next theorem characterizes the domain of the greatest fixed point, wrt , of function wd . This domain is the set of points for which normal termination is guaranteed (no possibility of abnormal termination or infinite loop). (17) Theorem. Let W be the greatest fixed point, wrt to , of wd (Abbreviation 12). We have W < = a 2 l. The following theorem is a generalization to a nondeterministic context of the while statement verification rule of Mills [24]. It shows that the greatest fixed point W of wd is uniquely characterized by conditions (a) and (b), that is, by the fact that W is a fixed point of wd and by the fact that no infinite loop is possible when the execution is started in a state that belongs to the domain of W . Note that we also have W < ≤ a (see Theorem 17), but this condition is implicitly enforced by condition (a). Half of this theorem (the ⇐ direction) is also proved by Sekerinski (the main iteration theorem [29]) in a predicative programming set-up. (18) Theorem. A relation W is the greatest fixed point, wrt , of function wd (Abbreviation 12), iff the following two conditions hold: (a) W = wd (W ), (b) W < ≤ l. In what follows we give some applications of our results. 5



In [6, 7], Berghammer and Schmidt propose abstract relation algebra as a practical means for the specification of data types and programs. Often, in these specifications, a relation is characterized as a fixed point of some function. Can demonic operators be used in the definition of such a function? Let us now show with a simple example that the concepts presented in this paper give useful insights for answering this question. In [6, 7], it is shown that the natural numbers can be characterized by the relations z and S (zero and successeur ) the laws (a) Ø = z = zL ∧ zz ⊆ I (z is a point), SS = I ∧ S S ⊆ I (S is a one to one application.), Sz = Ø (z has a predecessor), = L = {x|z ∪ S x x} (generation principle). For the rest of this section, assume that we are given a relation algebra satisfying these laws. In this algebra, because of the last axiom, the inequation (a) z ∪ S X ⊆ X obviously has a unique solution for X, namely, X = L. Because the functiong(X) := z ∪ S X is ∪continuous, this solution can be expressed as (a) L =

n≥0



g n (Ø) =



n≥0



S



n



z,



where g 0 (Ø) = Ø, g n+1 (Ø) = g(g n (Ø)), S 0 = I and S n+1 = S S n . However, it is shown in [6, 7] that z S 2 X ⊆ X, obtained by replacing the join and composition operators in a by their demonic counterparts, has infinitely many solutions. Indeed, from Sz = Ø and the Schr¨der rule, it follows that o (a) z ∩ S L = Ø, so that, by definition of demonic join (8(a)) and demonic composition (9), z S 2 X = (z ∪ S 2 X) ∩ z ∩ (S 2 X)L ⊆ z ∩ S L = Ø. Hence, any relation R is a solution to z S 2 X ⊆ X. Looking at previous papers [14, 32, 33, 34, 31], one



UbiCC Journal – Volume 4 No. 3



575



Special Issue on ICIT 2009 Conference - Applied Computing



immediately sees why it is impossible to reach L by joining anything to z (which is a point and hence is an immediate predecessor of Ø), since this can only lead to z or to Ø. Let us now go ‘fully demonic’ and ask what is a solution to z S 2 X X. By the discussion above, this is equivalent to Ø X, which has a unique solution, X = Ø. This raises the question whether it is possible to find some fully demonic inequation similar to (a), whose solution is X = L. Because L is in the middle of the demonic semilattice, there are in fact two possibilities: either approach L from above or from below. For the approach from above, consider the inequation X z S

2



how the universal relationL arises as the greatest lower bound n≥0 S n 2 z of this set of points. Note that, whereas there is a unique solution to a, there are infinitely many solutions to 4 (equivalently, to a), for example n≥k S n (= n≥k S n ), for any k. For the upward approach, consider z X 2S X.



X.



Using Theorem 10(c), we have z S 2X = z S X, since S is deterministic (axiom a(b)). From a, z ⊆ S L; this implies z ⊆ S XL and S X ⊆ z, so that, by definition of , z S X = z ∩ S X ∪ z ∩ S XL ∪ z ∩ S X = z ∪ S X. This means that 4 reduces to (a) X z ∪ S X.



By definition of refinement (7), this implies that z ∪ S XL ⊆ XL; this is a variant of (a), thus having XL = L as only solution. This means that any solution to 4 must be a total relation. But L is total and in fact is the largest (by ) total relation. It is also a solution to 4 (since by axiom a(d), z ∪ S L = L) so that L = {X|X z S 2 X}; that is, L is the greatest fixed point in (BL , ) of n2 f (X) := z S 2 X. Now consider z, n≥0 S n where S is a n-fold demonic composition defined by S 0 = I and S n+1 = S 2 S n . By axiom a(b), S is deterministic, so that, by 10(c) and associativity of demonic composition, conS n 2 z = S n z. Hence, It is easy to show that for any n ≥ 0, S n z is a point (it is the n-th successor of zero) and that m = n ⇒ S m z = S n z. Hence, in (BL , ), {S n z|n ≥ 0} (i.e. {S n 2 z|n ≥ 0}) is the set of immediate predecessors of Ø; looking at [31] shows 6



Here also there are infinitely many solutions to this inequation; in particular, any vector v, including Ø and L, is a solution to 4. Because (BL , ) is only a join semilattice, it is not at all obvious that the least fixed point of h(X) := z X 2 S exists. It does, however, since the following derivan tion shows that n≥0 z 2 S n (= n≥0 h (z ), 0 where h (z ) = z ) is a fixed point of h and hence is obviously the least solution of 4: Because z and S are mappings, property 10(c) implies that z 2 S n = z S n , for any n ≥ 0. But z S n is also a mapping (it is the inverse of the point S n z) and hence is total, from which, by Proposition 8(a) n and equation a, n≥0 z 2 S n = = n≥0 z S n n z S = ( n≥0 S z)˘ = L = L. This n≥0 means that L is the least upper bound of the set of mappings {z 2 S n |n ≥ 0}. Again, a look at [31] gives some intuition to understand this result, after recalling that mappings are minimal elements in (BL , ) (though not all mappings have the form z 2 S n ). Thus, building L from below using the set of mappings {z 2 S n |n ≥ 0} is symmetric to building it from above using the set of points {S n 2 z|n ≥ 0}.



5



Conclusion



We presented a theorem that can be also used to find the fixed points of functions of the form f (X) := Q ∨ P 2 X (no restriction on the domains of P and Q). This theorem can be applied also to the program verification and construction (as in the precedent example). Half of this theorem (the ⇐ direction) is also proved by Sekerinski (the main iteration theorem [29]) in a predicative programming set-up. Our theorem is more general because there is no restriction on the domains of the relations P and Q.



UbiCC Journal – Volume 4 No. 3



576



Special Issue on ICIT 2009 Conference - Applied Computing



The approach to demonic input-output relation presented here is not the only possible one. In [19, 20, 21], the infinite looping has been treated by adding to the state space a fictitious state ⊥ to denote nontermination. In [8, 18, 22, 26], the demonic input-output relation is given as a pair (relation,set). The relation describes the input-output behavior of the program, whereas the set component represents the domain of guaranteed termination. We note that the preponderant formalism employed until now for the description of demonic input-output relation is the wp-calculus. For more details see [3, 4, 17].



[6] Berghammer, R.: Relational Specification of Data Types and Programs. Technical report 9109, Fakult¨t f¨r Informatik, Universit¨t der a u a Bundeswehr M¨nchen, Germany, Sept. 1991. u [7] Berghammer, R. and Schmidt, G.: Relational Specifications. In C. Rauszer, editor, Algebraic Logic, 28 of Banach Center Publications. Polish Academy of Sciences, 1993. [8] Berghammer, R. and Zierer, H.: Relational Algebraic Semantics of Deterministic and Nondeterministic Programs. Theoretical Comput. Sci., 43, 123–147 (1986). [9] Boudriga, N., Elloumi, F. and Mili, A.: On the Lattice of Specifications: Applications to a Specification Methodology. Formal Aspects of Computing, 4, 544–571 (1992). [10] Chin, L. H. and Tarski, A.: Distributive and Modular Laws in the Arithmetic of Relation Algebras. University of California Publications, 1, 341–384 (1951). [11] Conway, J. H.: Regular Algebra and Finite Machines. Chapman and Hall, London, 1971. [12] Davey, B. A. and Priestley, H. A.: Introduction to Lattices and Order. Cambridge Mathematical Textbooks. Cambridge University Press, Cambridge, 1990. [13] J. Desharnais, B. M¨ller, and F. Tchier. Kleene o under a demonic star. 8th International Conference on Algebraic Methodology And Software Technology (AMAST 2000), May 2000, Iowa City, Iowa, USA, Lecture Notes in Computer Science, Vol. 1816, pages 355–370, SpringerVerlag, 2000. [14] Desharnais, J., Belkhiter, N., Ben Mohamed Sghaier, S., Tchier, F., Jaoua, A., Mili, A. and Zaguia, N.: Embedding a Demonic Semilattice in a Relation Algebra. Theoretical Computer Science, 149(2):333–360, 1995. 7



References

[1] Backhouse, R. C., and Doombos, H.: Mathematical Induction Made Calculational. Computing science note 94/16, Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands, 1994. [2] Backhouse, R. C., Hoogendijk, P., Voermans, E. and van der Woude, J.:. A Relational Theory of Datatypes. Research report, Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands, 1992. [3] R. J. R. Back. : On the correctness of refinement in program development. Thesis, Department of Computer Science, University of Helsinki, 1978. [4] R. J. R. Back and J. von Wright.: Combining angels, demons and miracles in program specifications. Theoretical Computer Science,100, 1992, 365–383. [5] Backhouse, R. C. and van der Woude, J.: Demonic Operators and Monotype Factors. Mathematical Structures in Comput. Sci., 3(4), 417– 433, Dec. (1993). Also: Computing Science Note 92/11, Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands, 1992.



UbiCC Journal – Volume 4 No. 3



577



Special Issue on ICIT 2009 Conference - Applied Computing



[15] Desharnais, J., Jaoua, A., Mili, F., Boudriga, N. and Mili, A.: A Relational Division Operator: The Conjugate Kernel. Theoretical Comput. Sci., 114, 247–272 (1993). [16] Dilworth, R. P.: Non-commutative Residuated Lattices. Trans. Amer. Math. Sci., 46, 426–444 (1939). [17] E. W. Dijkstra. : A Discipline of Programming. Prentice-Hall, Englewood Cliffs, N.J., 1976. [18] H. Doornbos. : A relational model of programs without the restriction to Egli-Milner monotone constructs. IFIP Transactions, A-56:363–382. North-Holland, 1994. [19] C. A. R. Hoare and J. He. : The weakest prespecification. Fundamenta Informaticae IX, 1986, Part I: 51–84, 1986. [20] C. A. R. Hoare and J. He. : The weakest prespecification. Fundamenta Informaticae IX, 1986, Part II: 217–252, 1986. [21] C. A. R. Hoare and al. : Laws of programming. Communications of the ACM, 30:672–686, 1986. [22] R. D. Maddux. : Relation-algebraic semantics. Theoretical Computer Science, 160:1–85, 1996. [23] Mili, A., Desharnais, J. and Mili, F.: Relational Heuristics for the Design of Deterministic Programs. Acta Inf., 24(3), 239–276 (1987). [24] Mills, H. D., Basili, V. R., Gannon, J. D. and Hamlet,R. G.: Principles of Computer Programming. A Mathematical Approach. Allyn and Bacon, Inc., 1987. [25] Nguyen, T. T.: A Relational Model of Demonic Nondeterministic Programs. Int. J. Foundations Comput. Sci., 2(2), 101–131 (1991). [26] D. L. Parnas. A Generalized Control Structure and its Formal Definition. Communications of the ACM, 26:572–581, 1983 [27] Schmidt, G.: Programs as Partial Graphs I: Flow Equivalence and Correctness. Theoretical Comput. Sci., 15, 1–25 (1981). 8



[28] Schmidt, G. and Str¨hlein, T.: Relations and o Graphs. EATCS Monographs in Computer Science. Springer-Verlag, Berlin, 1993. [29] Sekerinski, E.: A Calculus for Predicative Programming. In R. S. Bird, C. C. Morgan, and J. C. P. Woodcock, editors, Second International Conference on the Mathematics of Program Construction, volume 669 of Lecture Notes in Comput. Sci. Springer-Verlag, 1993. [30] Tarski, A.: On the calculus of relations. J. Symb. Log. 6, 3, 1941, 73–89. [31] F. Tchier.: S´mantiques relationnelles e d´moniaques et v´rification de boucles non e e d´terministes. Theses of doctorat, D´partement e e de Math´matiques et de statistique, Universit´ e e Laval, Canada, 1996. [32] F. Tchier.: Demonic semantics by monotypes. International Arab conference on Information Technology (Acit2002),University of Qatar, Qatar, 16-19 December 2002. [33] F. Tchier.: Demonic relational semantics of compound diagrams. In: Jules Desharnais, Marc Frappier and Wendy MacCaull, editors. Relational Methods in computer Science: The Qu´bec seminar, pages 117-140, Methods Pube lishers 2002. [34] F. Tchier.: While loop d demonic relational semantics monotype/residual style. 2003 International Conference on Software Engineering Research and Practice (SERP03), Las Vegas, Nevada, USA, 23-26, June 2003. [35] F. Tchier.: Demonic Semantics: using monotypes and residuals. IJMMS 2004:3 (2004) 135160. (International Journal of Mathematics and Mathematical Sciences) [36] M. Walicki and S. Medal.: Algebraic approches to nondeterminism: An overview. ACM computong Surveys,29(1), 1997, 30-81. [37] L.Xu, M. Takeichi and H. Iwasaki.: Relational semantics for locally nondeterministic



UbiCC Journal – Volume 4 No. 3



578



Special Issue on ICIT 2009 Conference - Applied Computing



programs. New Generation Computing 15, 1997, 339-362.



9



UbiCC Journal – Volume 4 No. 3



579



Special Issue on ICIT 2009 Conference - Applied Computing



CASE STUDIES IN THIN CLIENT ACCEPTANCE

Paul Doyle, Mark Deegan, David Markey, Rose Tinabo, Bossi Masamila, David Tracey School of Computing, Dublin Institute of Technology, Ireland WiSAR Lab, Letterkenny Institute of Technology {paul.doyle, mark.deegan, david.markey}@dit.ie,{rose.tinabo, bossi.masamila}@student.dit.ie david.tracey@lyit.ie



ABSTRACT Thin Client technology boasts an impressive range of financial, technical and administrative benefits. Combined with virtualisation technology, higher bandwidth availability and cheaper high performance processors, many believe that Thin Clients have come of age. But despite a growing body of literature documenting successful Thin Client deployments there remains an undercurrent of concern regarding user acceptance of this technology and a belief that greater efforts are required to understand how to integrate Thin Clients into existing, predominantly PC-based, deployments. It would be more accurate to state that the challenge facing the acceptance of Thin Clients is a combination of architectural design and integration strategy rather than a purely technical issue. Careful selection of services to be offered over Thin Clients is essential to their acceptance. Through an evolution of three case studies the user acceptance issues were reviewed and resolved resulting in a 92% acceptance rate of the final Thin Client deployment. No significant bias was evident in our comparison of user attitudes towards desktop services delivered over PCs and Thin Clients. Keywords: Thin Clients, Acceptance, Virtualisation, RDP, Terminal Services.



1



INTRODUCTION



It is generally accepted that in 1993 Tim Negris coined the phrase “Thin Client” in response to Larry Ellison’s request to differentiate the server centric model of Oracle from the desktop centric model prevalent at the time. Since then the technology has evolved from a concept to a reality with the introduction of a variety of hardware devices, network protocols and server centric virtualised environments. The Thin Client model offers users the ability to access centralised resources using full graphical desktops from remotely located, low cost, stateless devices. While there is sufficient literature in support of Thin Clients and their deployment, the strategies employed are not often well documented. To demonstrate the critical importance of how Thin Clients perform in relation to user acceptance we present a series of case studies highlighting key points to be addressed in order to ensure a successful deployment. 1.1 Research Aim The aim of this research has been to identify a successful strategy for Thin Client acceptance within an educational institute. There is sufficient literature which discusses the benefits of Thin Client adoption, and while this was referenced it was not central to the aims of this research as the barrier to obtaining these benefits was seen to be acceptance of the



technology. Over a four year period, three Thin Client case studies were run within the Dublin Institute of Technology with the explicit aim of determining the success factors in obtaining user satisfaction. The following data criteria were used to evaluate each case study in addition to referencing the Universal Theory of User Acceptance Testing (UTUAT) [1]. 1) Login events on the Thin Clients. 2) Reservation of the Thin Client facility. 3) The cost of maintaining the service. 1.2 Paper Structure In section 2 we review the historical background and trends of Thin Client technology to provide an understanding of what the technology entails. Section 3 discusses the case for Thin Clients within existing literature including a review of deployments within industry and other educational institutes. Section 4 provides details of the three case studies discussing their design, evaluating the results, and providing critical analysis. Section 5 takes a critical look at all of the data and sections 6 and 7 provide conclusions and identify future work. This paper is aimed at professionals within educational institutes seeking ways to realize the benefits of Thin Client computing while maintaining the support and acceptance of users. It provides a balance between



UbiCC Journal – Volume 4 No. 3



585



Special Issue on ICIT 2009 Conference - Applied Computing

the hype of Thin Clients and the reality of their deployment. 2 THIN CLIENT EVOLUTION The challenge faced by Thin Client technology is to deliver on these lower costs and mobility, while continuing to provide a similarly rich GUI user experience to that provided by the desktop machine (a challenge helped by improved bandwidth, but latency is still often a limiting factor [4]) and the flexibility with regard to applications they have on their desktop. Typically, current Thin Client systems have an application on a server (generally Windows or Linux) which encodes the data to be rendered into a remote display protocol. This encoded data is sent over a network to a Thin Client application running on a PC or a dedicated Thin Client device to be decoded and displayed. The Thin Client will send user input such as keystrokes to the application on the server. The key point is that the Thin Client does not run the code for the user's application, but only the code required to support the remote display protocol. While the term Thin Client was not used for dumb terminals attached to mainframes in the 1970's, the mainframe model shared many of the attributes of Thin Client computing. It was centralised, the mainframe ran the software application and held the data (or was attached to the data storage) and the terminal could be shared by users as it did not retain personal data or applications, but displayed content on the screen as sent to it by the mainframe. From a desktop point of view, the 1980's were dominated by the introduction and adoption of the Personal Computer. Other users requiring higher performance and graphics used Unix Workstations from companies like Apollo and Sun Microsystems. The X Window System [5] was used on many Workstations and X terminals were developed as a display and input terminal and provided a lower cost alternative to a Unix Workstation, with the X terminal connecting to a central machine running an X display manager. As such, they shared some of the characteristics of a Thin Client system, although the X terminal ran an X Server making it more complicated than Thin Client devices. The 1990's saw the introduction of several remote display protocols, such as Citrix's ICA [6] Microsoft's RDP [7] and AT&T's VNC [8] for Unix that took advantage of the increasing bandwidth available on a LAN to provide a remote desktop to users. Terminal Services was introduced as part of Windows NT4.0 in 1996 and it offered support for the Remote Desktop Protocol (RDP) allowing access to Windows applications running on the Server, giving users access to a desktop on the Server using an RDP client on their PC. RDP is now offered on a range of Windows platforms [9]. Wyse and vendors such as Ncomputing launched terminals, which didn't run the Windows operating system, but accessed Windows applications on a Windows Server using RDP, which is probably still the



The history of Thin Clients is marked by a number of overly optimistic predictions that it was about to become the dominant model of desktop computing. In spite of this there have been a number of marked developments in this history along with those of desktop computing in general which are worth reviewing to set the context for examining the user acceptance of this technology. Thin Clients have established a role in desktop computing although not quite the dominant one initially predicted. These developments have usually been driven by increases in processing power (and reductions in the processor costs) in line with Moore's law, but the improvements in bandwidth and storage capacity are having an increasing effect on desktop computing and on Thin Client computing [2] driving the move towards more powerful lower cost desktops but also the possibilities of server virtualisation and Thin Client computing with the ability to run Thin Clients over WANs. The first wave of computing was one where centralised mainframe computers provided the computing power as a shared resource which users accessed using dumb terminals which provided basic text based input and output and then limited graphics as they became graphics terminals. These mainframes were expensive to purchase and were administered by specialists in managed environments and mostly used for specific tasks such as performing scientific calculations and running highly specialised bespoke payroll systems. The next wave was that of personal computing, whereby users administered their own systems which provided a platform for their personal applications, such as games, word-processor, mail and personal data. Since then the personal computer has undergone a number of significant changes, but the one of most interest was the nature of the interface provided to the user which has grown into a rich Graphical User Interface where the Personal Computer became a gateway to the Internet with the Web browser evolving into a platform for delivery of rich media content, such as audio and video. This move from a mainframe centralised computing model to a PC distributed one resulted in a number of cost issues related to administration. This issue was of particular concern for corporate organizations, in relation to licensing, data security, maintenance and system upgrades. For these cost reasons and the potential for greater mobility for users, the use of Thin Clients is often put forward as a way to reduce costs using the centralised model of the Thin Client architecture. This also offers lower purchase costs and reduces the consumption of energy [3].



UbiCC Journal – Volume 4 No. 3



586



Special Issue on ICIT 2009 Conference - Applied Computing

dominant role of dedicated hardware Thin Clients. Similarly VNC is available on many Linux and Unix distributions and is commonly used to provide remote access to a user's desktop. These remote display protocols face increasing demands for more desktop functionality and richer media content, with ongoing work required in how, where and when display updates are encoded, compressed or cached [10]. Newer remote display protocols such as THINC have been designed with the aim of improving these capabilities [11]. In 1999, Sun Microsystems took the Thin Client model further with the SunRay, which was a simple network appliance, using its own remote display protocol called ALP. Unlike some of the other Thin Clients which ran their own operating system, SunRay emphasized its completely stateless nature [12]. This stateless nature meant that no session information or data was held or even cached (not even fonts) on the appliance itself and enabled its session mobility feature, whereby a smart card was used to identify a user with a session so that with the smartcard the user could login from any SunRay connected to the session's server and receive the desktop as it was previously. Many of these existing players have since focused on improving their remote desktop protocols and support for multimedia or creating new hardware platforms. There have also been some newer arrivals like Pano Logic and Teradici who have developed specific client hardware to create “zero” clients, with supporting server virtualisation to render the remote display protocols. Also, there are a number of managed virtual desktops hosted in a data centre now being offered. One of the drivers behind Thin Client Technology, particularly when combined with a dedicated hardware device, is to reduce the cost of the client by reducing the processing requirement to that of simply rendering content, but a second driver (and arguably more important one) is to gain a level of universality by simplifying the variations in the client side environment. This has been met in a number of new ways using Virtual Machine players and USB memory in Microsoft's research project “Desktop on a Keychain” (DOK) [13] and also the Moka5 product [14], allowing the mobility (and security) benefits attributed to Thin Clients. This can be enhanced with the use of network storage to cache session information [15]. It can be seen that Thin Clients have evolved along with other desktop computing approaches, often driven by the same factors of increasing processing power, storage capacity and bandwidth. However, newer trends that are emerging with regard to virtualisation, internet and browser technologies, together with local storage, present new challenges and opportunities for Thin Client technology to win user acceptance. As Weiser said in 1999 in this new era, “hundreds or thousands of computers do our bidding. The relationship is the inverse of the mainframe era: the people get the air conditioning now, and the nice floors, and the computers live out in cyberspace and sit there waiting eagerly to do something for us”. [16] 3 THE CASE FOR THIN CLIENTS



There are many stated benefits for Thin Clients all of which are well documented [17][18]. While there is no single definitive list, potential system designers may have different aims when considering Thin Clients, these benefits should be clearly understood prior to embarking on any deployment and are discussed below. 3.1 Reduced cost of software maintenance The administrative cost benefit of the Thin Client model, according to Jern [19] is based on the simple observation that there are fewer desktop images to manage. With the combination of virtualisation environments and Windows Terminal Service (WTS) systems it would not be uncommon for twenty five or more desktop environments to be supported from a single installation and configuration. This reduces the number of upgrades and customizations required for desktop images in computer laboratories where the aim is to provide a consistent service from all systems. Kissler and Hoyt [20] remind us that the “creative use of Thin Client technology can decrease both management complexity and IT staff time.” In particular they chose Thin Client technology to reduce the complexity of managing a large number of kiosks and quick-access stations in their new thirty three million dollar library. They have also deployed Thin Client devices in a range of other roles throughout Valparaiso University in Indiana. Golick [21] on the other hand suggests that the potential benefits of a Thin Client approach include the lower mean time to repair (MTTR) and lower distribution costs. It is interesting to note that he does suggest that the potential cost savings for hardware are a myth, but that administration savings still make a compelling case for using Thin Client technology. Enhanced Security Speer and Angelucci [22] suggest that security concerns should be a major factor in the decision to adopt Thin Client systems and this becomes more apparent when referencing the Gartner Thin Client classification model. The Thin Client approach ensures that data is stored and controlled at the datacentre hosting the Thin Client devices. It is easy to argue that the user can retain the mobility of laptops but with enhanced security and the data is not mobile, just the access point. The argument is even easier to make when we consider recent high-profile cases of the theft of unencrypted laptops containing sensitive medical or financial records. The freedom



3.2



UbiCC Journal – Volume 4 No. 3



587



Special Issue on ICIT 2009 Conference - Applied Computing

conferred on users of corporate desktop and laptop PCs undermines the corporation’s obligations in relation to data privacy and security. Steps taken to protect sensitive data on user devices are often too little and too late. Strassmann [23] states that the most frequent use of a personal computer is for accessing web applications and states that the Thin Client model demonstrates significantly lower security risks for the corporation. Five security justifications for adopting the Thin Client model were proposed. 1) 2) 3) 4) 5) Zombie Prevention Theft Dodging File Management Software Control Personal Use Limitations 1) 2) 3) 4) 5.4 million kWh reduction, 2,800 tonnes of CO2 saved annually Servers reduced by a factor of 20 IT budget cut by a fifth



Strassmann concedes that Thin Clients are not necessarily best for every enterprise and every class of user, but for enterprises with a large number of stationary “non-power” users, “Thin Clients may present the best option in terms of security, cost effectiveness and ease of management.” 3.3 User Mobility User mobility can refer to the ability of a user to use any device, typically within the corporation’s intranet, as a desktop where the user will see a consistent view of the system, for example, SunRay hot-desking. While user profiles in Microsoft Windows support this, it is often only partially implemented. Session mobility can be viewed as the facility for users to temporarily suspend or disconnect their desktop session and to have it reappear, at their request, on a different device at a later time. This facility removes the need for users to log-out or to boot-up a desktop system each time they wish to log-in. Both of these potential features of Thin Client technologies help to break the sense of personal ownership that users often feel for their desktop or laptop computers. It is this sense of personal ownership which makes the maintenance and replacement of corporate PCs a difficult task, and this feeling of ownership and control is often a reason why users resist the adoption of a centrally controlled Thin Client to replace their desktop, whereas this is exactly why IT management may want to adopt it. Environmental Costs In the article “An Inefficient Truth” Plan [24] reveals a series of “truths” supported by a number of case studies directed at the growing costs of Information and Communication Technologies. One such case study is of Reed Managed Services where 4,500 PCs were replaced with Thin Clients, and a centralised blade server providing server based virtualised desktops. Savings are reported as follows:



Indeed there are many deployments focused on obtaining energy savings through the use of Thin Clients. In a case study where SunRay systems were introduced into Sparkasse a public German Bank, Bruno-Britz [25] reports that the savings in electricity costs alone were enormous. The University of Oxford has deployed SunRay Thin Client devices in their libraries citing the cooler and quieter operation as factors in their decision. These devices, having no local hard disk and no fan operate at a lower temperature and more quietly than traditional PCs. This characteristic has environmental implications from noise, cooling and power consumption perspectives. 3.5 Summary of Benefits In summary, we can extract the benefits observed within literature and case studies as follows: 1) Increased security as data maintained centrally 2) Reduced cost of hardware deployment and management and faster MTTR 3) Reduced administration support costs 4) Environmental costs savings 5) Reduced cost of software maintenance 6) Reduced cost of software distribution 7) Zero cost of local software support 8) The ability to leverage existing desktop hardware and software 9) Interface portability and session mobility 10) Enhanced Capacity planning 11) Centralised Usage Tracking and Capacity Planning 3.6 Thin Clients vs. Fat Clients Thin Client technology has evolved in sophistication and capability since the middle of the 1990s, however the “thickness” (the amount of software and administration required on the access device) of the client is a source of distinction for many vendors [26][27]. Regardless of “thickness”, Thin Clients require less configuration and support when compared to Fat Clients (your typical PC). In the early 1990s Gartner provided a client-server reference design shown in Figure 1. This design provides clarity for the terms “thin” and “fat” clients by viewing applications in terms of the degree of data access, application and presentation logic present on the server and client sides of the network. The demand for network based services such as email, social networking and the World Wide Web has driven bandwidth and connectivity requirements to higher and higher levels of reliability and performance [28]. As we progress to an “always on”



3.4



UbiCC Journal – Volume 4 No. 3



588



Special Issue on ICIT 2009 Conference - Applied Computing

network infrastructure the arguments focused against Thin Clients based on requiring an offline mode of usage are less relevant. The move from Fat Client to Thin Client is however often resisted as individuals find themselves uncomfortable with the lack of choice provided when the transition is made, as observed by Wong et al.[29]. incomplete and flawed technology. In the case of Thin Clients, it should be accepted that there are tradeoffs to be made. One of the appealing aspects of the Fat client is its ability to be highly flexible which facilitates extensive customization. However not every user will require that flexibility and customization. Thin Clients are not going to be a silver bullet addressing all users needs all of the time. All three case studies were evaluated under the following headings in order to allow a direct comparison between each. These criteria were selected to ensure that there was a balance between the user acceptance of the technology and the technical success of each deployment. 1) Login events on the Thin Clients 2) Reservation of the Thin Client facility 3) The cost of maintaining the service



Figure 1: Gartner Group Client/Server Reference Design



4



CASE STUDIES



No matter how well documented the benefits of Thin Clients may be, there is still an issue of acceptance to be addressed. While it may be tempting to assume that the implementation of technology is a technical issue and that simply by building solutions a problem is effectively solved, evidence would point to the contrary. As there can often be a disparity between what is built and what is required or needed. Too often requirements gathering, specification definition and user consultation are forgotten in the rush to provide new services which are believed to be essential. In essence the notion of “if we build it they will come” is adopted, inevitably causing confusion and frustration for both service provider and the user. For example, during Sun Microsystems’ internal deployment of its own SunRay Thin Client solution many groups and functions sought exemptions from the deployment as they believed that their requirements were sufficiently different to the “generic user” to warrant exclusion from the project. The same arguments still exist today and it is often those with a more technical understanding of the technology who are the agents of that technology’s demise. By providing interesting and often creative edge cases which identify the limitations of a technology, they can, by implication, tarnish it as an



Figure 2: Case Study 1



4.1



DIT Case Study 1 In 2005 the DIT introduced the SunRay Thin Client technology into the School of Computing. In a similar approach to many other technology deployments the strengths of the technology were reviewed and seen as the major selling points of the deployment. In the case of SunRay there was a cheap appliance available which would provide the service of graphical based Unix desktops. Centralised administration ensured that the support costs would be low and the replacement requirements for systems for the next five years would be negligible. In essence the technological and administrative advantages were the focus of this deployment. Few of the services offered within the existing PC infrastructure were included in the deployment. This deployment sought to offer new services to students and introduced Thin Clients for the first time to both students and staff.



UbiCC Journal – Volume 4 No. 3



589



Special Issue on ICIT 2009 Conference - Applied Computing

4.1.1 Design A single laboratory was identified for deploying the SunRay systems and all PC in that lab were replaced with SunRay 150 devices. A private network interconnect was built which ensured that all data sent from the clients traversed a private network to the SunRay server. The initial design of this case study is shown in Figure 2 and it allowed students within this new Thin Client lab access to the latest version of Solaris using a full screen graphical environment as opposed to an SSH command-line Unix shell which was the traditional method still used from existing computing laboratories. A new authentication system was introduced based on LDAP which required students to have a new username and password combination which was different to the credentials already in use within the Active Directory domain used for the existing PC network. The reason for this alternative authentication process was due to the difficulty of authenticating on a Unix system using Active Directory. Once the server was running, the Thin Client laboratory was ready to provide graphical based Unix login sessions at a considerable reduced price when compared to an investment of Unix workstations for each desk. In total 25 Thin Client devices were installed which were all connected to a single Solaris server. In summary the key components within the design were as follows: 1) 2) 3) 4) 5) 6) The service was on a private network New authentication process was introduced New storage mechanism was introduced Devices were all in the same location Service provided was a CDE desktop on Solaris Graphical desktops running on Linux servers also accessible Given that the nature of the service did not significantly change over the course of the three years that the system was in place with the exception of semester activity in line with student presence in the institute, it is clear that there was low utilization of the service. The graph shows raw data plotted, where login events were less than 10 per day.



14 12 10



Login Events per day



8 6 4 2 0 Feb 05 Feb 06 Feb 07 Feb 08



Figure 3: User Login Events



Reservation of the Thin Client Facility: Each laboratory may be reserved by staff for the delivery of tutorial sessions and exercises. The hourly reservations for this laboratory were reduced as a result of the introduction of Thin Clients with only 1 to 2 hours being reserved per day. One of the primary reasons for the reduction in the use of this facility was the fact that it had now become special purpose and the bookings for the room were limited to the courses which could be taught within it. The Cost of Maintaining the Service: A detailed analysis of cost savings associated with the introduction of Thin Clients within our institute and specifically the costs associated with this case study was performed by Reynolds and Gleeson, [30]. In their study they presented evidence of savings in relation to the cost of support, the cost of deployment and a basic analysis of the power consumption costs. They review both the system and the software distribution steps associated with Thin Clients and PC systems and present a point of quantifiable comparison between the two. Key findings of this analysis were as follows: 1) Time spent performing system upgrades and hardware maintenance was reduced to virtually zero as no hardware or software upgrades were required. 2) A single software image was maintained at the central server location and changes were made available instantly to all users. 3) No upgrade costs were incurred on the Thin Clients or server hardware. All systems have



4.1.2 Results The login events are a measure of the general activity of the devices themselves and were considered to be a reasonable benchmark for comparison with existing laboratories within the institute. One interesting point is that the comparison of facilities is not necessarily relevant when the facilities provide different services. Due to the fact that Unix instead of Windows was provided meant that, with the exception of those taking courses involving Unix, the majority of students were unfamiliar with the technology and did not seek to use the systems. Login events on the Thin Clients: The login events were extracted from the Solaris server by parsing the output of the last command which displays the login and logout information for users which it extracts from the /var/adm/wtrmpx file. The number of login events per day was calculated and plotted in the graph shown in Fig. 3. Immediately obvious was the low use of the system.



UbiCC Journal – Volume 4 No. 3



590



Special Issue on ICIT 2009 Conference - Applied Computing

remained in place throughout both case studies. The devices in this lab are now 8 years old and are fulfilling the same role today as they did when first installed. 4) The Thin Client lab is a low power consumption environment due to the inherent energy efficiency of the Thin Client hardware over existing PCs. This can provide up to 95% energy savings when compared to traditional PCs [24]. 4.1.3 Analysis There has been extensive research in the area of user acceptance of technology, but perhaps the most relevant work in this area is the Unified Theory of Acceptance and Use of Technology (UTAUT) [1] which identifies four primary constructs or factors; a) b) c) d) Performance Expectancy Effort Expectancy Social Influence Facilitating Conditions This is defined as the degree to which there is a perception of how others will view or judge them based on their use of the system. Clearly by isolating the devices and having it associated with specialized courses, there was no social imperative to use the labs. Unix as a desktop was relatively uncommon in the School at the time of the case study and there would have been a moderate to strong elitist view of those who were technical enough to use the systems. d) Facilitating Conditions This is defined as the degree to which an individual believes in the support for a system. At first glance this does not appear to be a significant factor considering that the services were created by the support team and there was considerable vested interest in seeing it succeed. However additional questions asked by the UTAUT include the issue of compatibility with systems primarily used by the individual. Each of the UTAUT factors can be considered significant for Case Study 1. Many of the issues raised hang on the fundamental issue that the new services offered on the Thin Client were different to existing services and for all practical purposes seen as incompatible with the majority of systems available to students elsewhere. The fact that the technology itself may have worked flawlessly, and may have delivered reduced costs was irrelevant as the service remained under utilized. Given that the reason for this lack of acceptance was potentially inherent in the implementation of services and not due to failings in the technology itself it was clear that a second case study was required which would address the issue of service.



While there are additional factors such as Gender, Age and Experience, within the student populations these are for the most part reasonably consistent and will be ignored. It should be stressed that although the UTAUT was developed for an industry based environment it is easily adapted for our purposes. It was felt that this model serves as a relevant reference point when discussing the performance of the case studies. Clearly Case Study 1 failed to gain acceptance despite belief that it would in fact be highly successful at its inception. We review the case study under the four UTAUT headings to identify the source of the user rejection of the Thin Clients. a) Performance Expectancy This factor is concerned with the degree to which the technology will assist in enhancing a users own performance. Clearly however the services provided an advantage to those students who wished to use Unix systems. Since the majority of courses are based on the Windows operating system it would be reasonable to assume that there was no perceived advantage in using a system which was not 100% compatible with the productivity applications used as part of the majority of courses. b) Effort Expectancy This factor is concerned with the degree of ease associated with the use of the system. One of the clear outcomes of Case Study 1 was that students rejected the Unix systems as it was seen to be a highly complex system, requiring additional authentication beyond what was currently used in traditional laboratories. c) Social Influence



Figure 4: Case Study 2



UbiCC Journal – Volume 4 No. 3



591



Special Issue on ICIT 2009 Conference - Applied Computing

4.2 Case Study 2 The second case study is a modification of the basic implementation of the first case study with changes focused on increasing student acceptance of the Thin Client facility. Removing the Unix centric nature of the existing service was central to the system redesign. It was decided that additional services could be easily and cheaply offered to the Thin Client environment providing users with the ability to access more compatible services from within the Thin Client environment. Figure 4 identifies the key components within the design. b) Course specific Windows Terminal Servers for courses where there were specific software requirements not common to all students. c) Individual Virtualised desktops for students in specific modules where administration rights were required. d) All services were made available from both the Thin Client and PC labs as they were available over the Remote Desktop Protocol RDP. e) Provisioning of an easy access point to all services from within the Thin Client environment which was not available from PC systems. 4.2.2 Results The data gathered for Case Study 2 was evaluated under same three headings as per case study 1. 1) Login events on the Thin Clients 2) Reservation of the Thin Client facility. 3) The cost of maintaining the service. 25 20 Login Events per Day 15 10 5 0 08 Feb 22 Feb 08 Mar 22 Mar 05 Apr

Figure 5: User Login Event Comparison Case Study 2



4.2.1 Design The most important addition to the second case study was the provision of additional services which were similar to those available in PC labs. This was to ensure that students could use this facility and have an experience on a par with the PC labs. A new domain was created where Unix and Windows shared a common authentication process. Due to difficulties integrating Unix and the existing Windows authentication process, the new Domain was built on the LDAP system with SAMBA providing the link between the new Windows Terminal Servers and the LDAP system. While students could now use the same username and password combination for Windows and Unix systems this was not integrated into the existing Windows authentication process. Students were still required to have two sets of credentials, the first for the existing PC labs, and the second for access to a new domain containing a number of Windows Terminal Servers and the original graphical Unix desktop. While the Thin Clients now provided Windows and Unix graphical desktops, the new Windows Domain was also accessible from existing PC labs via RDP connections to the Terminal Servers. This allowed classes to be scheduled either inside or outside of the Thin Client laboratory. In addition to providing Windows Terminal Services (WTS), student owned virtual machines were now also available. Due to the fact that most services were now available from all locations, the ease of access to the services from within the Thin Client lab was improved by providing users with a menu of destinations upon login. This new login script effectively provided a configurable redirection service to the WTS and Virtualisation destinations using the rdesktop utility [31] which performed a full screen RDP connection to specified destinations. An interesting outcome of this destination chooser was that any RDP based destination could be included regardless of the authentication process used. This would however require a second authentication process with the connecting service. The new services provided were as follows: a) A general purpose Windows Terminal Server with mounted storage for all students and staff.



Case Study 1



Login events on the Thin Clients: Figure 5 shows a comparison of activity during the same time period for the two case studies. To identify trends in the data a displacement forward moving average was performed on the data as shown in Eq. (1). (1)



It is clear that for the same time period there was a significant increase in the use of the system as the number of login events increased by a factor of 4. Once again the login events were extracted from the Solaris server by parsing the output of the last command. Reservation of the Thin Client Facility: The changes to the Thin Client facility were announced at the start of the second academic semester as a PC upgrade and the number of room bookings increased as shown in Figure 6 from 6 hours a week to 20 hours a week. This was due to the use of the room as a Windows based laboratory



UbiCC Journal – Volume 4 No. 3



592



Special Issue on ICIT 2009 Conference - Applied Computing



using the new WTS and virtualisation services. 8



Hours per day



Case Study 1 Case Study 2



6 4 2 0 Mon Tue Wed



Thurs



Fri



Figure 6: Thin Client Room Reservations



The Cost of Maintaining the Service: All of the benefits observed from the first case study were retained within this case study. The addition of terminal services reduced the reliance of students on Fat Client installations. Students are now using virtual machines and terminal servers on a regular basis from all labs.



4.2.3 Analysis This second case study certainly saw an improvement over its earlier counterpart and students and staff could now access more familiar services from the Thin Client lab. Given the dramatic increase relative to the earlier results it could be stated that the introduction of the more familiar services increased the acceptance of the facility. Both case studies demonstrated equally well that it is possible to obtain the total cost of ownership benefits using a Thin Client model, but the services offered has a dramatic affect on user acceptance. It is useful to review the outcome in relation to the UTUAT. a) Performance Expectancy Given that new services such as personalised virtual machines were now available, staff and students could identify a clear advantage to the system where administration rights could be provided in a safe manner, allowing more complex and previously unsupported activities to take place. For example, the Advanced Internet module for the MSc. students could now build and administer full web servers which could remain private to the student ensuring that no other student could access or modify a project which was a work in progress. b) Effort Expectancy Considerable improvements were made in this case study to allow users to access well known environments from both the Thin Clients and PC systems. Students who were taught modules using the new WTS or virtual environments were trained on how to access the systems, and once they used them they continued to do so throughout the year. Those who did not have



modules being taught using these new services were still required go through a new login/access process which was not well documented. For example within the Thin Client labs the new username/password combination was required to access the choice of destinations from the devices. This acted as a barrier to use even though emails were sent to students and information on how to access these accounts were posted in the labs. Usernames were based on existing student ID numbers. c) Social Influence Little changed in this case study for those who did not have a teaching requirement based on the new services. d) Facilitating Conditions With the provision of WTS services and virtual machines which provided Windows environments the issue of compatibility was reduced. However two issues remained which were not addressed. Firstly while users could now share a common data store between systems on this new domain there was no pre-packaged access to the data store on the existing PC domain. While it was technically possible to combine both under a single view, this required user intervention and additional training which was not provided. Secondly the sequence of steps required to access choices from the Thin Clients was a non-standard login process which now required a second login, the first of which was at a Unix graphical login screen. For many this initial login step remained as a barrier to using the system.



The most striking result from this case study is that while the second case study demonstrated significant increase in acceptance and use, the PC environments remained the system of choice for students, as shown in Figure 7. In this graph we show the typical use PC laboratory within the same faculty. Thin Client use remained less than one third of the use of the busiest computer laboratory. Thin Clients are shown to be capable of providing services equally well to both Windows and Unix users by introducing the ability of students to access their own private desktop from many locations, however this feature alone was not enough to entice users from the existing PC infrastructure. Clearly the introduction of virtualisation to the infrastructure allowed new services to be developed and used from Thin and Fat clients which could be seen as a potential for migrating users to a Thin Client/Virtualisation model, which indeed is a future planned initiative. The results show a definite increase in the use of the Thin Client facilities with data being gathered from the same period over both case studies to eliminate any bias which might occur due to module schedule



UbiCC Journal – Volume 4 No. 3



593



Special Issue on ICIT 2009 Conference - Applied Computing



differences at different time periods during the year. The timing and method used to announce the changes was critical to the increase in acceptance. The announcement of the systems as a PC upgrade removed some of the barriers which existed for users who did not feel comfortable with a Unix environment but failed to attract a majority of the students. 80 70 60 50 40 30 20 10 0

08 Feb



PC Lab 1



Case Study 2 Case Study 1

22 Feb 08 Mar 22 Mar 05 Apr



Figure 7: Comparison with PC Computer Labs



4.3



Case Study 3



The third case study was designed using the experiences of the first two case studies and was extended beyond the School of Computing. It was aimed at demonstrating the capability of the Thin Client technology in two different demographic environments, the first was one of the Institute Libraries where PCs were used by students from many different faculties and the second was within the Business faculty where computer system use was provided in support of modules taught within that faculty. This case study expressed the following aims at the outset 1) To demonstrate the use of Thin Client technology within the student population and determine the level of student acceptance of that technology. 2) To implement a number of alternative technologies in order to provide a point of comparison with respect to their overall performance and acceptance. 3) To determine the capability of the existing network infrastructure to support Thin Clients. 4.3.1 Design Unlike the previous case studies the aim was to insert Thin Clients into the existing environment as invisibly as possible. This meant that existing authentication processes were to be maintained. There were two different authentication processes in place which needed to be support, Novell Client for the Business faculty and Active Directory for the Library. In both cases a WTS system was built which joined to the respective domains. Applications were installed on the Thin Client in order to mirror those that were present on existing PCs in the chosen



locations. It was essential that the Thin Clients were not to be identifiable by students if at all possible, and to co-locate them with existing PC systems. To ensure that all devices behaved in a consistent manner to PCs they must boot and present the same login screen as would be expected on a PC in the same location. To achieve this all Thin Client devices with the exception of the SunRay systems used a Preboot Execution Environment (PXE) [32] boot process to connect to a Linux Terminal Server Project server (LTSP). The server redirected the user session to the correct WTS using rdesktop where the user was presented with a Windows login screen identical to those on adjacent PC systems. The SunRay systems were run in Kiosk mode which allowed the boot sequence to redirect the session to a WTS also via the rdesktop utility. The WTS were installed on a VMWare ESX Server to allow rollback and recovery of the servers. This however was not central to the design of the case study and only served as a convenience in sharing hardware resources between multiple servers. The only concern was the potential performance of the WTS under a virtualised model. Given that the applications were primarily productivity applications such as word processing and browsing, and that the maximum number of users allowable on any WTS was 25 (based on the number of devices which were directly connected to the WTS) this was considered to be within the acceptable performance range of the architecture. This assumption was tested prior to the case study being made accessible to students with no specific issues raised as to warrant further restructuring of the architecture Seventy five Thin Clients were deployed in six locations. The following Thin Client devices were used as shown in Figure 8 and Table 1.



Login Events per Day



Figure 8: Case Study 3



UbiCC Journal – Volume 4 No. 3



594



Special Issue on ICIT 2009 Conference - Applied Computing



Table 1: Thin Clients deployed Device Boot Mode Quantity Dell GX260 PXE Boot PC 15 Dell FX 160 PXE Boot TC 25 HP T5730 PXE Boot TC 8 Fujitsu FUTRO S PXE Boot TC 2 SunRay 270 SunRay 25 4.3.2 Linux Terminal Server Project LTSP works by configuring PCs or suitable Thin Clients to use PXE-Boot to obtain the necessary kernel and RDP client used as part of this project. These are obtained from a TFTP server whose IP address is provided as a DHCP parameter when the client PXE-Boots. As part of the DHCP dialogue, devices configured to PXE-Boot are given settings by the DHCP server. These include; TFTP Boot Server Host Name and Bootfile Name. The necessary settings were configured on each of the DHCP servers serving the relevant locations within the DIT so as to point any PXE-Boot devices to the relevant LTSP boot server and to specify the kernel to be loaded by the PXE-Boot client. Using these settings the PXE-Boot clients load a Linux kernel and then an RDP client which connects to one of the three WTS used as part of this case study.



remotely from the primary labs within the Business faculty and traditionally did not have high use. Lab 2 was a more central location and again as expected this exhibited greater user activity. The systems remained in operation continually for the period of the case study which was over one month during which data was collected from the three WTS systems. 4.3.4 User Survey Once the case study was running a desktop satisfaction survey which employed the Likert scale [33] was conducted to obtain feedback from students using the Thin Client systems. The design of the questionnaire was such that students were asked to identify their desktop using a colour coded system which was known only to the authors. Each of the Thin Clients and a selection of PC systems (which were not PXE booted) where targeted for the survey to allow a comparative analysis between all Thin Clients and existing PC systems to be performed. The survey did not reference Thin Clients in any of the questions but rather sought feedback on application use and overall satisfaction with the performance of the system through a series of questions. There were 234 responses recorded for the survey. The key questions in the survey were as follows. 1) Please rate the overall performance of the machine you are currently using 2) Please identify the primary reason you used this computer 3) How would you rate your overall satisfaction with this desktop? 4) Would you use this desktop computer again? 80% User Satisfaction Ratings 75% 70% 65% 60%

PC-Fat SunRay PXE HP TC Dell TC Client Boot PC



140 Login Events per Day 120 100 80 60 40 20 0 17 Apr 24 Apr 1 May 8 May 15 May

Figure 9: User Login Event Comparison



Library Lab 2 Lab 1



4.3.3 Results Use of the Thin Clients was recorded using login scripts on the Windows Terminal Servers which recorded login and logout events. As expected the use of the Library systems exceed the use of the laboratories but both were in line with typical use patterns expected for each location. What was immediately obvious was that each location had a higher utilization than the previous two case studies but comparable with the PC labs shown in Figure 9. One of the difficulties with the comparison however is that the final case study was performed at a different point in the teaching semester and use of the systems declined as students prepared for examinations. Lab 1 was a “quiet lab” located



All Applications



Browsers



Figure 10: User satisfaction rating of desktop performance



The issue of overall performance was broken down by the device used which was identified using the colour coded scheme described earlier. Figure 10 below represents the average rating of satisfaction reported by users broken down by device and primary application in use. Since over 50% of responses identified “Browsing” as the primary reason for using the machine there are two



UbiCC Journal – Volume 4 No. 3



595



Special Issue on ICIT 2009 Conference - Applied Computing



User Satisfaction Ratings



satisfaction ratings provided as a point of comparison. Figure 11 shows the combined rating of users responses to overall satisfaction with desktop, desktop performance and application performance.



Non-USB Storage 100% 90% 80% 70% 60% 50% 40%

PC



USB Only



80% User Satisfaction Ratings 78% 76% 74% 72% 70% 68% 66%

PC SunRay PXE Boot HP DELL



SunRay PXE Boot



HP



Dell



Figure 13: Storage Satisfaction Rating



Figure 11: Combined rating of desktop performance



4.3.5 Analysis This final case study while shorter in length than the other case studies demonstrated significant progress in user acceptance. As part of the survey users were asked if they would consider reusing the system and as can be seen in Figure 12 there was significant support for the systems. The small number of responses representing those who did not wish to reuse the system cited USB performance as the primary cause of their dissatisfaction. This was identified early in the testing of the Thin Clients that all systems performed noticeably slower than the PC systems in this respect. Questions regarding the primary storage method used by students were added to the survey as was a satisfaction rating. From the results in Figure 13 it is clear that while the PC systems did perform better when users primarily used USB storage, the satisfaction in storage performance for all other options were comparable. The HP satisfaction rate had a low survey response rate and hence was not considered significant in our analysis given the small number of data points. NO, 8%



By making the Thin Clients as invisible as possible and comparing satisfaction and user access to the existing PC systems it was clear that for the majority of users there was no apparent change to the services provided. Integrating into the existing authentication process was an essential feature of this case study as was the presenting of a single authentication process at the WTS login screen. Efforts were also made to ensure that the applications installed on the WTS were configured to look and feel the same as those on the standard PC. As with the previous case studies it is useful to review the case study in relation to the UTUAT. a) Performance Expectancy With the exception of increasing the number of desktops in the Library, the primary deployment mainly replaced existing systems, so users were not provided with any reminders that they were using a different system. In effect there was no new decision or evaluation by the user to address the questions which were relevant in the previous case studies. b) Effort Expectancy The reuse of the existing login/access procedure which was well known and part of the normal process for students using existing PC systems again allowed for this factor to become mainly irrelevant. Usernames, passwords, applications and system behaviour were identical to those on the PCs. c) Social Influence Without perceiving a difference in service, social influence as a factor was also eliminated. Only the SunRay systems had different keyboards and screens, and as these screens were of higher resolution than existing PCs they were if anything seen as a more popular system. d) Facilitating Conditions Unlike the previous case studies support for the facility was more complex. Different levels of expertise and engagement were required. Thin



YES, 92%



Figure 12: User Response "Would you use this system again"



UbiCC Journal – Volume 4 No. 3



596



Special Issue on ICIT 2009 Conference - Applied Computing



Clients were now part of a larger support structure where many individuals were not core members of the technical team who built the systems. However given that only three support calls were raised during the case study there was little pressure on this factor either. The calls raised were not in fact directly related to the Thin Client devices, but rather the network and the virtual environments used to host the centralised servers. 5 CRITICAL ANALYSIS



case study. These three case studies provide data centric analysis of user acceptance and identify the evolving designs of our deployments. To gain acceptance of Thin Clients within an educational institute our case studies identified these key factors. 1) Locate the Thin Clients among the existing PC systems, do not separate them or isolate them. 2) Ensure that the login process and credentials users are identical to the existing PC systems. 3) Ensure that the storage options are identical to the existing PC systems 4) Focus on providing exactly the same services that already exist as opposed to focusing on out new services. By ensuring we ran a blind test on the user population where Thin Clients co-existed with PC systems, and where the services offered were indistinguishable by the user, we were able to show a user satisfaction rating of 92%. No significant bias was evident in our comparison of user attitudes of desktop services delivered over PCs and Thin Clients. 7 FUTURE WORK



The UTUAT provides a useful reference point in understanding some of the factors affecting acceptance of the Thin Clients. In the first case study the primary barrier to acceptance was the incompatibility of the new system with the existing system. Students were not motivated to use the new system as there were few advantages to doing so and considerable effort in learning how to use it. The second case study while more successful still failed to gain acceptance despite the expansion of services offered being comparable with existing Windows services. The session mobility and access from anywhere feature, while useful did not overcome the resistance of users to migrate to the Thin Clients. Thin Clients still required separate credentials and the login process was still different to the PC systems. The third and final case study was designed to provide the same existing services as the PC only using a centralised server and Thin Client model. No new services for the user were provided. The primary aim was to have the systems indistinguishable from the existing installation of PCs, effectively running a blind test for user acceptance. Once the users accepted the new systems, further machines could be deployed quickly and cheaply. The total cost of ownership and centralised support savings demonstrated in the first two case studies were just as relevant in the third case study. 6 CONCLUSION



Additional case studies are planned which will focus on acceptance of Thin Clients within the academic staff population and will evaluate the relevance of some of the proposed core technological advantages within that environment such as session mobility, Desktop as a Service, and dynamic lab reconfiguration and remote access using WAN and not just LAN environments. 8 REFERENCES [1] V. Venkatesh, M.G. Morris, G.B. Davis, and F.D. Davis, “User acceptance of information technology: Toward a unified view,” Mis Quarterly, 2003, pp. 425-478. [2] J.D. Northcutt, “CYB Newslog - Toward Virtual Computing Environments.” [3] D. Tynan, “Think thin,” InfoWorld, Jul. 2005. [4] S.J. Yang, J. Nieh, M. Selsky, and N. Tiwari, “The Performance of Remote Display Mechanisms for Thin-Client Computing,” IN PROCEEDINGS OF THE 2002 USENIX ANNUAL TECHNICAL CONFERENCE, 2002. [5] T. Richardson, F. Bennett, G. Mapp, and A. Hopper, “Teleporting in an X window system environment,” IEEE Personal Communications Magazine, vol. 1, 1994, pp. 6-13. [6] Citrix Systems, “Citrix MetaFrame 1.8 Backgrounder,” Jun. 1998.



There is considerable literature in support of Thin Client technology, and while there may be debate regarding the finer points of its advantages the issue has been and continues to be one of acceptance. Acceptance for Thin Clients as a technology is often confused with the non technical issues arising from the deployment. The UTUAT helps distinguish between technical and non-technical issues and as shown within our case studies, the way in which the technology was presented to the user had a higher impact on acceptance than had the technology itself. This point is highlighted by the fact that the Thin Client devices which were not widely used in first case study were integrated seamlessly into the third



UbiCC Journal – Volume 4 No. 3



597



Special Issue on ICIT 2009 Conference - Applied Computing



[7] Microsoft Corporation, “Remote Desktop Protocol: Basic Connectivity and Graphics Remoting Specification,” Technical White Paper, Redmond, 2000. [8] T. Richardson, Q. Stafford-Fraser, K. Wood, and A. Hopper, “Virtual network computing,” Internet Computing, IEEE, vol. 2, 1998, pp. 33-38. [9] Microsoft Corporation, “Microsoft Windows NT Server 4.0, Terminal Server Edition: An Architectural Overview,” Jun. 1998. [10] J. Nieh, S.J. Yang, and N. Novik, “A comparison of thin-client computing architectures,” Network Computing Laboratory, Columbia University, Technical Report CUCS-022-00, 2000. [11] R.A. Baratto, L.N. Kim, and J. Nieh, “Thinc: A virtual display architecture for thin-client computing,” Proceedings of the twentieth ACM symposium on Operating systems principles, ACM New York, NY, USA, 2005, pp. 277-290. [12] B.K. Schmidt, M.S. Lam, and J.D. Northcutt, “The interactive performance of SLIM: a stateless, thin-client architecture,” Proceedings of the seventeenth ACM symposium on Operating systems principles, Charleston, South Carolina, United States: ACM, 1999, pp. 32-47. [13] M. Annamalai, A. Birrell, D. Fetterly, and T. Wobber, Implementing Portable Desktops: A New Option and Comparisons, Microsoft Corporation, 2006. [14] “MokaFive, Virtual Desktops,” http://www.mokafive.com/. [15] R. Chandra, N. Zeldovich, C. Sapuntzakis, and M.S. Lam, “The Collective: A cache-based system management architecture,” Proceedings of the 2nd USENIX Symposium on Networked Systems Design and Implementation (NSDI’05). [16] M. Weiser, “How computers will be used differently in the next twenty years,” Security and Privacy, 1999. Proceedings of the 1999 IEEE Symposium on, 1999, pp. 234-235. [17] M. Jern, “"Thin" vs. "fat" visualization clients,” Proceedings of the working conference on Advanced visual interfaces, L'Aquila, Italy: ACM, 1998, pp. 270-273. [18] S. Kissler and O. Hoyt, “Using thin client technology to reduce complexity and cost,” Proceedings of the 33rd annual ACM SIGUCCS conference on User services, ACM New York, NY, USA, 2005, pp. 138-140. [19] M. Jern, “"Thin" vs. "fat" visualization clients,” L'Aquila, Italy: ACM, 1998, pp. 270-273. [20] S. Kissler and O. Hoyt, “Using thin client technology to reduce complexity and cost,” New York, NY, USA: ACM, 2005, pp. 138– 140.



[21] J. Golick, “Network computing in the new thinclient age,” netWorker, vol. 3, 1999, pp. 3040. [22] S.C. Speer and D. Angelucci, “Extending the Reach of the Thin Client.,” Computers in Libraries, vol. 21, 2001, pp. 46 - . [23] P.A. Strassmann, “5 SECURE REASONS FOR THIN CLIENTS.,” Baseline, 2008, p. 27. [24] G.A. Plan, “An inefficient truth,” PC World, 2007. [25] M. Bruno-Britz, “Bank Sheds Pounds.,” Bank Systems & Technology, vol. 42, 2005, p. 39. [26] “Sun Ray White Papers,” http://www.sun.com/sunray/whitepapers.xml. [27] B.K. Schmidt, M.S. Lam, and J.D. Northcutt, “The interactive performance of SLIM: a stateless, thin-client architecture,” Charleston, South Carolina, United States: ACM, 1999, pp. 32-47. [28] S. Potter and J. Nieh, “Reducing downtime due to system maintenance and upgrades,” San Diego, CA: USENIX Association, 2005, pp. 66. [29] I. Wong-Bushby, R. Egan, and C. Isaacson, “A Case Study in SOA and Re-architecture at Company ABC,” 2006, p. 179b. [30] G. Reynolds and M. Gleeson, “Towards the Deployment of Flexible and Efficient Learning Tools: The Thin Client,” The Proceedings of the 4th China-Europe International Symposium on Software. China (Guanzhou). Sun Yat-Sen University. (2008). [31] “rdesktop: A Remote Desktop Protocol client.” [32] B. Childers, “PXE: not just for server networks anymore!,” Linux J., vol. 2009, 2009, p. 1. [33] R. Likert, “A Technique for the Measurement of Attitudes,” Archives of Psychology, vol. 140, 1932, pp. 1–55.



UbiCC Journal – Volume 4 No. 3



598



Special Issue on ICIT 2009 Conference - Applied Computing



AN INTERACTIVE COMPOSITION OF WORKFLOW APPLICATIONS BASED ON UML ACTIVITY DIAGRAM

Yousra Bendaly Hlaoui, Leila Jemni Ben Ayed Research Unit of Technologies of Information and Communication Tunis, Tunisia Yousra.bendalyhlaoui@esstt.rnu.tn Leila.jemni@fsgt.rnu.tn



ABSTRACT In today's distributed applications, semi automatic and semantic composition of workflows from Grid services is becoming an important challenge. We focus in this paper on how to model and compose interactively workflow applications from Grid services without considering lower level description of the Grid environment. To reach this objective, we propose a Model-Driven Approach for developing such applications based on semantic and syntactic descriptions of services available on the Grid and abstract description provided by UML activity diagram language as well. As there are particular needs for modeling composed workflows interactively from Grid services, we propose to extend the UML activity diagram notation. These extensions deal with additional information allowing an interactive and semi automatic composition of workflows. In addition this specific domain language contains appropriate data to describe matched Grid services that are useful for the execution of the obtained workflows. Keywords: Grid services, Interactive, semantic, composition, Workflow application, UML activity diagrams.



1



INTRODUCTION



Today’s distributed applications [23] are developed by integrating web or Grid services [13, 14] in a workflow. Due to the very large number of available services and the existence of different possibilities for constructing workflow from matching services, the problem of building such applications is usually a non trivial task for a developer. This problem requires finding and orchestrating appropriate Grid services in a workflow. Therefore, we propose an approach that allows semi automatic and interactive composition of workflow applications from Grid services. To describe and model workflow applications we use UML [25] activity diagrams. Recently, several solutions were proposed to compose applications from Grid services such as works presented in [8, 17, 18]. However, the proposed solutions need interaction with user and guidelines or rules in the design of the composed applications. Consequently, the resulting source code is neither re-usable nor it promotes dynamic adaptation facilities as it should. However, for applications composed of Grid services, we need an abstract view not only of the offered services but also of the resulting application [31]. This abstraction allows the reuse of the elaborated



application and on the other reduces the complexity of the composed applications. There are several architectural approaches for distributed computing applications [22] which make easy the development process. However, these approaches need rigorous development methods to promote the reuse of components in future Grid development application [16]. It has been proven from past experience that using structured engineering methods makes easy the development process of any computing system and reduces the complexity when building large Grid application [22]. To reduce this complexity and allow the reuse of Grid service applications, we adopt a model-driven approach [24]. Thus we introduce in this paper a new approach to build, interactively, workflow applications by following OMG(s) principals of the MDA in the development process [2, 3, 4]. In this approach [2, 3, 4], our focus is to compose and model workflows from existing Grid services that represent the main aspect in the development of Grid services applications. The workflow modeling identifies the control and data flows from one depicted Grid service's operation to the next to build and compose the whole application. To model and express the composed workflow of Grid services, we use as abstract language the



UbiCC Journal – Volume 4 No. 3



599



Special Issue on ICIT 2009 Conference - Applied Computing



activity diagrams of UML [25]. The provided model forms the Platform Independent Model (PIM) of the proposed MDA approach. This model is more understandable for the user than an XML [35] based workflow description languages like BPEL4WS [15] which represent the Platform Specific Model (PSM). This paper is organized as follows. Section 2 presents the related work. Section 3 introduces the different components of the composition system; section 4 specifies our proposed UML profile, composition patterns and different steps of the interactive composition process. Finally, section 5 concludes the paper and proposes areas for further research. 2 RELATED WORK



Many works were carried out in the field of Grid and Web services composition, such as works presented in [8, 17, 18, 19, 20, 28, 29, 30]. In [28] authors were interested in the semi automatic composition of web services and proposed a validation approach based on the semantic descriptions of services and on a logic based language to describe and validate the resulting composite Web services. However, the resulting composed web service is not clear for user who is not familiar with logic based languages. In our contribution, we propose a solution not only to compose workflows from available Grid services, but also to provide graphical and comprehensive models of the resulting workflows. In the same framework, authors in [29] proposed a composition approach of Web services based on Symbolic Transition Systems (STS). They developed a sound and complete logical approach for identifying the existence of available composition. They have emphasized upon the abstract representation of the composition request (the goal of the composition) and the representation of the resulting composite Web service. For the representation, authors have used UML state machine diagrams [25] which are suitable only to describe a sequence of component services without addressing the other forms of matching services in a workflow such as parallel branches or and-branches. On the other hand, UML activity diagrams that we use in our modelling approach support all kind of workflow composition patterns [10] such as parallelism, split and fork. The authors in [19, 20, 30] have proposed a Model Driven Approach for composing manually Web services. They were based on UML activity diagrams to describe the composite Web service and on UML class diagrams to describe each available Web Service. The user depicts the suitable Web service and matches it in the workflow representing the composite Web service using UML activity diagrams. This approach would have been better if the composition were automatically elaborated, since



the number of available services is in increase with the existence of several forms and manners to compose such services. Based on domain ontology description, we lead the user through to the composition process. Also, we provide for this user a graphical interface based on a domain specific UML language for automatic grid service composition. This UML profile [5] is based on stereotypes, tagged values and workflow patterns [5] that we propose to ensure the automatic composition. In the field of Grid services composition the most related work is the work presented by Gubala et al in [8, 17, 18]. In this work, the authors have developed a tool for semi automatic and assisted composition of scientific Grid application workflows. The tool uses domain specific knowledge and employs several levels of workflow abstractness in order to provide a comprehensive representation of the workflow for the user and to lead him in the process of possible solution construction, dynamic refinement and execution. The originality of our contribution is that firstly we save the effort of the user from the dynamic refinement and execution as we propose a Model Driven Approach which separates the specific model from the independent model. Secondly, we use UML activity diagrams to deliver the functionality in a more natural way for the human user. The use of UML activity diagrams in the description of workflow application is argued in several works such as works presented in [1, 10, 12, 27]. Thus, the advantage of UML activity diagrams is that they provide an effective visual notation and facilitate the analysis of workflows composition. In our approach, we propose an UML profile for composing systematically a workflow application from Grid services [5]. 3 THE INTERACTIVE COMPOSITION SYSTEM WORKFLOW



The system allows an interactive and semantic composition of workflows from Grid services. As shown in figure 1, the system is composed of three components: a Grid Services workflows composer, an ontological Grid Services registry and a workflows execution system also we call it activity machine. The Grid services workflow composer This system is composed of three components: the composition tool, the transformation tool and the verification tool. 3.1.1 The composition tool It provides a graphical interface in the form of UML activity diagrams editor allowing to the user an interactive, systematic and semantic workflow 3.1



UbiCC Journal – Volume 4 No. 3



600



Special Issue on ICIT 2009 Conference - Applied Computing



Figure 1: Different components of the workflow composition system Composition [6]. This composition is based on the composition process which will be detailed in section 4.3. In the Grid registry, services are described in an ontological form with statements regarding the service operation's inputs, outputs, pre-conditions and effects (the IOPE set) [26]. Through these notions, the composition system is able to match different grid service’s operations into a workflow following a reverse traversal approach. Thus, and by associating the required data with the produced output, the composer constructs a data flow between Grid service’s operations using our workflow composition patterns and UML profile[5]. The composer may also use a specific notion of effect that may bind two operations together with non-data dependency. If the Grid registry fails to find the right operation, the composition process stops. Otherwise, the composition process will stop when all workflow dependencies are resolved. The request is sent to the Ontological Grid Registry in the form of SPARQL query [34]. This language provides a higher-level access to the ontology transcribed knowledge for the automatic discovery and semantic matching of services. Therefore, once the workflow model is built, it should be validated and verified to ensure its reliability before being executed and reused as subworkflow. 3.1.2 The transformation tool To support the verification and the execution of workflow models described in UML activity diagrams (UML-AD), the transformation tool translates the activity diagram into a Hyper-Graph (HG). This HG will be translated as well by the transformation tool into a NuSMV format file according to a relative semantic. The details of these semantics may not be relevant to the topic for which the paper is submitted. However these details could be made available. 3.1.3 The verification tool Checking errors in design models like UML activity diagrams is essential since correcting an error while the system is alive and running is usually very costly [21]. Consequently, workflow activity diagram models should be spotted and corrected as early as possible [6]. Several techniques are used in the field of behavioural design verification such as theorem proving and model checking [11]. The latter is the most useful because it is fully automatic and gives feedback in case of detected errors. It verifies whether some given finite state machine satisfies some given property specified in temporal logic [9]. For activity diagrams, symbolic model checking has proven to be an efficient verification technique [11]. Thus, our verification tool is based on NuSMV symbolic model checker [9] that supports strong fairness property which is necessary to be verified in a workflow model to obtain realistic results. With the model checker, arbitrary propositional requirements can be checked against the input model. If a requirement fails to hold, an error trace is returned by the model checker. The transformation tool translates systematically the error trace into an activity diagram trace by high-



UbiCC Journal – Volume 4 No. 3



601



Special Issue on ICIT 2009 Conference - Applied Computing



lighting a corresponding path in the given activity diagram. The Grid registry During the workflow composition process, the Grid registry provides the composer system with the description of services available at the moment and provides reasoning capabilities to enable proper matchmaking of services inputs and outputs. The Grid registry [6] is an ontological distributed repository of Grid services and subworkflows. This registry is responsible for storing and managing documents which contains descriptions of syntax and semantics of services and their operations expressed in an RDF file [33]. The semantic Web is making available technologies which support automate knowledge sharing. In particular there are several existing initiatives such as OWL-S [26] which proves that ontologies have a key role in the automating service discovery and composition. That knowledge is based on semantic descriptions of service classes published by the service developers and provided in the Grid environment [16]. Our Grid registry is based on an ontological description of services and workflows. The service ontology [7] provides concepts and properties that allow description and matchmaking of available services. A part of this ontology is common to all services and it is based on a standard semantic web service description ontology OWL-S [26] which makes interoperability with existing services. A part from the common ontology, there is a domain specific part of the ontology. The domain service ontology [7] allows users to extend the common ontology schema in order to provide a better specification of services as well as their inputs and outputs. For these we define a data ontology [7] which provides concepts and properties for describing services input and outputs. Ontology alignment [7] is a process for finding semantic relationships among the entities of ontologies. Its main activity is to find similar concept in ontologies being aligned, in order to map them. The measures for similarity computation can be divided into two general groups; namely lexical measures and structural measures. Lexical measures are based on surface similarity such as title or label of entities. In contrary, structural measures try to recognize similarities by considering structures of ontology graphs. The most advanced similarity algorithms use combination of multiple similarity measures to obtain more information about concepts similarity. In our Grid registry, we adopt an approach using a combination of lexical and structural similarity [7]. We use similarity measures for mapping domain ontology as initial selection and then the selection will be refined with using structural similarity method [7]. 3.2



The workflow execution system The reliable workflow model is sent to the workflow execution system [6] which produces implementation code for handling control flow and data flow. The activity diagram describing the workflow model is translated into a specific XML file which will be the input of the execution system. A workflow execution system executes different workflow activities specified in the workflow XML document in the correct order and with their required inputs and outputs data. The execution of an activity corresponds to the invocation of a Grid service’s operation. The workflow execution system monitors these activities using the tagged values information expressed in the activities but does not perform them. An activity of the activity diagram modelling the workflow represents a state of the workflow execution system in which the system waits for an invoked grid service operation to complete its work. Hence, the defined semantics of activity diagrams for the verification describe the behaviour of the execution system. When the system enters a state relative to an invocation grid service node or activity ai, it invokes a piece of behaviour that is executed by the service or system environment. While the latter is in ai (activity ai is active), it waits for the termination event of the invoked piece of behaviour. When termination event occurs, the system reacts by executing the outgoing edge E: it leaves the E's sources and enters the E's targets and the execution process continues for the other activity nodes until the final node is reached. 4 UML BASED INTERACTIVE COMPOSITION OF WORKFLOWS FROM GRID SERVICES



3.3



In order to match and compose different Grid service’s operations, we need to analyze constructs of workflow models at higher abstraction level. Since UML [25] is the core of the MDA [24], we use its activity diagram language to model composed workflows. The composition system provides to the user a graphical interface to compose its request using a UML profile specific for the domain of composing systematically 4.1 UML Profile for composing workflows In this section, we present our UML profile which is based on Domain Specific Language (DSL) for customizing UML activity diagrams for the systematic composition of workflows from Grid services [5]. In our DSL (See Figure 2), an activity of an UML activity diagram represents a Grid service's operation, while object flows represent the types of results which flow from one activity to another.



UbiCC Journal – Volume 4 No. 3



602



Special Issue on ICIT 2009 Conference - Applied Computing



Effects binding two operations are presented with control flows [5]. The name of an activity in the diagram represents the name of the Grid service's operation. This name must be specified as a Grid service could have more than one operation often called interface which are specified in its relative WSDL file [32]. There are two different types of activities: yetunresolved activities and established activities of the composed workflow. The former represent the need for a Grid Service's operation to be inserted in order to complete the workflow. However, the latter represent abstract operations that are already included into the workflow. As there are two different activity types in a Grid service workflow model, an activity needs to be typed and specified. To fulfil this, we propose to use the DSL modelling element invoke to stereotype an established activity which is used to invoke an external Grid service's operation and yetunresolved to stereotype activities which are not yet resolved. Object nodes of an established activity are data stereotyped. Unknown input and output for a yet-unresolved activity are unknown stereotyped. In our UML profile, an object node could be relatedto a final node as composed workflow of Grid application should always deliver a result. 4.2 UML-AD composition patterns We identify, in this section, how UML activity diagrams support some of basic Grid service composition patterns [5]. These patterns are



essential in the systematic building of workflow applications from Grid services. The use of these patterns depends on the number of the depicted Grid Service's operations and their inputs and outputs [5]. These operations are results of the semantic research elaborated by the ontological Grid services registry. This research is invoked by a request given by the composition system in order to complete an unresolved activity in the workflow. The Grid service registry provides zero, one or more operations producing the intended output. Operations are depicted to be inserted in the workflow interactively with the user. 4.2.1 Sequence Pattern When the Grid registry provides one Grid service's operation that is able to produce the required result or the user selects one operation from the provided operation set; the composition system uses the sequence pattern to insert the operation in the workflow. In this case and as is illustrated by the figure 3, a single abstract operation or activity (e.g. GridService1Operation1) will be inserted in the workflow model described by the UML-AD language. This operation may also require some data for itself (e.g. GridService1Operation1Input) and thus it may introduce a new unresolved dependency (e.g. the yet-unresolved stereotyped activity). So, we use a follow-up method to build a simple pipeline-type sequential workflow: a sequence pattern.



Figure 2: Meta-model of Grid service workflow composition specific UML activity diagram language



UbiCC Journal – Volume 4 No. 3



603



Special Issue on ICIT 2009 Conference - Applied Computing



A Sequence pattern is composed with sequential activities which are related with control flow (non data operations dependency) or object flow (data operation dependency).



to object node representing the required input which both of them flow to a merge construct. Semantically, several services instances are invoked in parallel threads and the merge will only wait for the first flow to finish. We distinguish, in Figure 5, two different Grid service's operations, GridService1Operation1 and GridService2Operation1 providing the same output data DataOutput.



Figure 3: The sequence pattern 4.2.2 And-branches pattern The and-branches pattern is introduced when the introduced operation represented by an abstract UML activity has more than one input. This pattern is based on the Synchronization pattern presented in [9]. This pattern starts with object nodes, representing alternative operation inputs, which flow to a join node. The latter is linked to the abstract grid service's operation. This operation introduces some unresolved dependencies in the workflow. Semantically, several services instances are invoked in parallel threads and the join will wait for all flows to finish. As illustrated in Figure 4, the operation of the Grid service GridService1Operation1 needs two inputs data GridService1Operation1Input1 and GridService1Operation1Input2. The relative pattern produces two parallel threads in the workflow.



Figure 5: Alternative branches pattern 4.2.4 Alternative services pattern When composing workflows from Grid services, a specific matching based on semantic comparison could provide two or more different Grid services performing each of them the required operation. In such case and when the user do not choose one of the depicted Grid service’s operations, the composition system uses the alternative services pattern to involve the operations in the workflow model. In this pattern, the Grid service’s operation to insert is modelled by a composed super-activity with a specified input data object and specified output data object (Figure 6). The super-activity is stereotyped as AlternativeServiceInstance to indicate that its task may be accomplished by a set of alternative service's instances. These alternative service instances are described with sub-activities. The sub-activities shall be grid service instances and thus stereotyped as invoke. It was up to decision mechanism of the workflow execution engine to choose which service instance in such given workflow node is to be invoked and executed. In Figure 6, the data DataOutput is provided from GridServiceOperation service operation which could GridService1Operation1 provider or be GridService2Operation2 provider.



Figure 4: And-branches pattern 4.2.3 Alternative branches pattern When the Grid registry provides more than one operation able to produce the required result, and the user do not select one of them, the composition requires a specific pattern: the alternative branches pattern. This pattern combines the Exclusive Choice and Simple Merge patterns presented in [9]. In this pattern, each alternative service's operation is linked Figure 6: Alternative services pattern



UbiCC Journal – Volume 4 No. 3



604



Special Issue on ICIT 2009 Conference - Applied Computing



4.3 The composition process Figure 7 illustrates the scenario of the composition process of workflows from available Grid services. This composition is based on the domain specific UML activity diagram language presented in section 4.1. In the following, we comment the different process steps of the scenario presented in the figure 7. Step 1: The user builds its composition request by specifying what kind of outcome or result that it expects from the workflow application execution. Step 2: The composition system analyses the desired output and sends a SPARQL query to request the ontologies of the Grid registry describing the available Grid services. The composer requests the Grid registry for a Grid service’s operation having the specified result as output. Step 3: If the required operation is not found and all unknown results are resolved then the composition process stops.



Step 4: If the required operation is found then the system displays its characteristics to the user to confirm the choice. The register may provide more than one operation. In such case the user could choose the operation to insert in the workflow model from the given list. If it does not specify its operation, then the system inserts all the given operations using one of the composition patterns presented in section 4.2. Relatively to the number of depicted operations and their inputs and outputs, the composer chooses the right composition pattern. Step 5: For each input of inserted operation, the system defines one unresolved dependency as a workflow activity which is not yet established. This activity depends on some Grid service’s operation. For each unresolved dependency the composer asks the user if it wants to continue the composition process or not. If the response is positive the composer re-executes the process from the step 2 to resolve the current unresolved dependency.



Figure 7: Scenario of the interactive composition of Grid service workflows based on UML activity diagrams



UbiCC Journal – Volume 4 No. 3



605



Special Issue on ICIT 2009 Conference - Applied Computing



5



Illustration of composition



the



interactive



workflow



In the following, we illustrate the composition process through the example of the domain of the city traffic pollution analysis. This application, as presented in [23], targets the computation of traffic air pollutant emission in an urban area. Step 1: Figure 8 shows an example of initial workflow that represents a composition request for the results of the pollutant emission due to the city traffic. The desired result, PollutionEmission, is described by the rectangle representing the object node in the relative activity diagram.



Figure 8: Initial workflow as a composition request Step 2: Figure 9 represents the workflow of the computation of traffic air pollution analysis after one step of composition. The service’s operation, delivering the PollutionEmission result, is AirPollutionEmissionCalculator. This operation is the result of the composer query asked to the ontological Grid registry. The operation requires two inputs TrafficFlowFile and PathsLenght-File, thus it infers two unresolved dependencies in the activity diagram modelling the composed workflow.



the composer system is able to match different operations into a workflow following a reverse traversal approach. Thus, and by associating the required data with the produced output, the composer constructs a data flow between operation using workflow patterns and our UML profile [5]. The composer may also use a specific notion of effect that may bind two operations together with non-data dependency. In [10], five basic control patterns were defined to be supported by all workflow languages and workflow products. These patterns are Sequence pattern, Parallel split pattern, Synchronization pattern, Exclusive Choice and Simple Merge patterns. Figure 10 represents the example of city traffic analysis Workflow after the full composition activity. It involves several Grid service operations, sequence branches, parallel split branches, simple merge branches and a loop [5]. The loop is involved in the workflow diagram as the application iterates in order to analyze the possible traffic. The Figure shows also how UML activity diagrams support the five basic patterns in the composition specific domain of Grid services workflows [5]. In the example, some of object node or input data, such as VehiculeType and StartZonzId, are given by the user of the application; they do not have an operation provider. This illustrates the interaction between our composition system and the user.



Figure 9: An example of workflow after one step of composition Step 3: For every dependency that needs to be resolved .i.e. a yet-unresolved activity, the composer contacts the ontological registry in order to find suitable service’s operations that may produce the required result. The services are described in an ontological form with statements regarding the service operation’s inputs, outputs, preconditions and effects (the IOPE set) [26]. Through these notions,



Figure 10: The workflow application after the full composition



UbiCC Journal – Volume 4 No. 3



606



Special Issue on ICIT 2009 Conference - Applied Computing



6



CONCLUSION



In this paper, we have presented an approach for composing interactively workflows from Grid services [2, 3, 4, 6]. This composition is based on an UML profile for customizing UML activity diagrams to compose and model workflows [5] and on composition patterns [5] as well. The interactive composition process was illustrated through the example of city traffic pollution analysis domain [23] We have developed and implemented most of the presented components of the composition system. Actually, we are working on the implementation of the workflow execution system that invokes and executes the depicted Grid service instances and manages the control and data flows in a run time environment relatively to our proposed activity diagram semantic.



Interactive Composition of UML-AD for the Modelling of Workflow Applications, In. Proc. International Conference on Of the 4th Information Technology, ICIT'2009, Amman, Jordan (2009). [7] Y. Bendaly Hlaoui, L. Jemni Ben Ayed: Ontological Description of Grid Services Supporting Automatic Workflow Composition, In. Proc. Of the International Conference on Web and Information Technologies, ICWIT'2009, Kerkennah, Tunisia, ACM SIGAPP.fr, IHE éditions, pp. 233-243 (2009). [8] M. Bubak, R. Guballa, M. Kapalka, M. Malawski, K. Rycerz: Workflow Composer and service registry for grid applications, Journal of Future Generation Computer Systems, Vol. 21, pp. 79-86 (2005). [9] A. Cimatti, E. Clarck, A. Tacchella: Nusmv version 2: An opensource tool for symbolic model checking, In Proc. Of the International Conference on Computer-Aided Verification, CAV'02, Lecture Notes in Computer Science, Springer Verlag (2002). [10] M. Dumas, and A. H. M. ter Hofsetde: UML Activity Diagrams as a Workflow Speci_cation Language, In UML'2001 Conference, Toronto, Ontario, Canada, Lecture Notes in Computer Science (LNCS), Springer-Verlag, Heidelberg, Germany (2001). [11] R. Eshuis: Semantics and verification of UML Activity Diagrams for Workflows Modelling, PhD thesis, University of Twente (2002). [12] R. Eshuis and R. Wieringa: Comparing Petri net and Activity diagram variants for workflow modelling: A Quest for Reactive Petri Nets, Petri Net technology for communication based Systems, LNCS, Springer Verlag (2003). [13] I. Foster, D.Berry, A.Djaoui, A.Grimshaw, B.Horn, H.Kishimoto, F.Maciel, A.Savy, F.Siebenlist, R.Subramaniam, J.Treadwell, J.Von Reich: The Open Grid Services Architecture, Version 1.0. (2004). [14] I. Foster, C. Kesselman: Grid Services for Distributed System Integration, Journal of IEEE Computer, Vol. 35, No. 6, pp. 37-46 (2004). [15] T. Gardner: UML modelling of automated Business Processes with a Mapping to BPEL4WS, In. Proc. Of the European Conference on Object Oriented Programming (2003).



7



REFERENCES



[1] R. Bastos, D. Dubugras, A. Ruiz: Extending UML Activity Diagram for Workflow modelling in Productions Systems, In. Proc. Of the 35th Annual Hawaii International Conference on System Sciences, HICSS'02, IEEE Cs Press (2002). [2] Y. Bendaly Hlaoui, L. Jemni Ben Ayed: Toward an UML-based composition of grid services workflows, In Proc. Of the 2nd international workshop on Agent-oriented software engineering challenges for Ubiquitous and Pervasive Computing, AUPC’08, ACM Digital Library, pp. 21-28 (2008). [3] Y. Bendaly Hlaoui, L. Jemni Ben Ayed: An extented UML activity Diagram for Composing Grid Services Workflows, In Proc. Of the IEEE international Conference on Risks and Security of Internet and Systems, CriSIS’08, Tozeur, Tunisia, p. 207-212 (2008).

[4] Y. Bendaly Hlaoui, L. Jemni Ben Ayed: A



MDA approach for semi automatic grid services workflows composition, In Proc. Of the IEEE international conference on Industrial Engineering and engineering Managment, IEEM’08, p.1433-1437 (2008). [5] Y. Bendaly Hlaoui, L. Jemni Ben Ayed: Patterns for Modeling and Composing Workflows from Grid Services, In. Proc. Of the 11th International Conference on Enterprise Information Systems, ICEIS'2009, Milan, Italy, LNBIP, SpringerVerlag, Vol. 24, pp. 615-626 (2009). [6] Y. Bendaly Hlaoui, L. Jemni Ben Ayed: An



UbiCC Journal – Volume 4 No. 3



607



Special Issue on ICIT 2009 Conference - Applied Computing



[16] C. Goble, D. de Roure: The grid: an application of the semantic web, ACM SIGMOD RecordSpecial section on semantic web and data management, Vol. 31, No.4, pp. 65-70 (2002). [17] T. Gubala, D. Herezlak, M. Bubak, M. Malawski: Semantic Composition of Scientific Workflows Based on the Petri Nets Formalism, In Proc. Of the Second IEEE International Conference on e-Science and Grid Computing, e-Science'06 (2006). [18] R. Guballa, A. Hoheisel, F. First: Highly Dynamic workflow Orchestration for scientific Applications, CoreGRID Technical Report, Number TR-0101 (2007). [19] R. Gronomo, I. Solheim: Towards Modelling Web Service Composition in UML, In The 2nd International Workshop on Web Services: Modelling, Architecture and Infrastructure, Porto, Portugal (2004). [20] R. Gronomo, MC. Jaeger: Model Driven Semantic Web Service Composition, In Proc. Of the 12th Asia-Pacific Software Engineering Conference, APSEC'05 (2005). [21] M. Laclavik, E.Gatial, Z. Balogh, O. Habala, G. Nguyen, L.Hluchy: Experience Management Based on Text Notes, In. Proc. Of e-Challenges 2005 Conf. (2005). [22] W. Li, C. Huang, Q. Chen, H. Bian: A ModelDriven Aspect Framework for Grid Service Development, In Proc. Of the IEEE International Conference on Internet and Web Applications and Services, ICIW’06, pp. 139-146 (2006). [23] M. Masetti, S. Bianchi, G. Viano: Application of K-Wf Grid technology to Coordinated Traffic Management. http://grid02.softeco.it/site/projectinfo.html [24] Model Driven Architecture (MDA). Document nomber omrsc/2001-07-01 (2001) [25] Object Management Group. UML Superstructure Specification. July (2005). 2.0



[27] Pllana, T. Fahringer, J. Testori, S. Benkner, I. Brandic: Towards an UML Based Graphical Representation of Grid Workflow Application, In. Proc. Of the 2nd Eu-ropean Across Grids Conference, Nicosia, Cyprus, Springer-Verlag (2004). [28] J. Rao, P. Kungas, M. Matskin: Logic-based web service composition: from service description to process model, In Proc. Of the IEEE International Conference on Web Services, ICWS 2004, San Diego, California, USA (2004). [29] E. Sirin, J. Hendler, B. Parsia: Semi automatic composition of web services using semantic descriptions, In Proc. Of the ICEIS-2003 Workshop on Web Services, Modeling, Architecture and Infrastructure, Angers, France (2003). [30] D. Skogan, R. Gronomo, I. Solheim: Web Service Composition in UML, In Proc. Of the 8th Intl Enterprise Distributed Object Computing Conference, EDOC'04 (2004). [31] M. Smith, T. Friese, B. Freisleben: Model Driven Development of Service-Oriented Grid Applications, In. Proc. Of the IEEE Asia-Pacific Conference on Services Computing, APSCC'06 (2006). [32] Web Services Descriptio Language (WSDL) 1.1. W3C Note 15 March (2001). [33]W3C: Resource Description Framework (RDF) Model and Syntax Specification, report num. TR/1999/REC-rdf-syntax-19990222 (1999). [34]W3C: SPARQL Query Language for RDF, report , 2008. [35] M. J. Young: XML Step by Step, Microsoft Press, ISBN: 2-84082-812-X (2001).



[26] OWL-S: Semantic Markup for Web Services. The OWL Services Coalition. OWL-S version 2.0.S.



UbiCC Journal – Volume 4 No. 3



608



Special Issue on ICIT 2009 Conference - Applied Computing



HOW TO MAP PERSPECTIVES

Gilbert Ahamer, Adrijana Car, Robert Marschallinger, Gudrun Wallentin, Fritz Zobl Institute for Geographic Information Science at the Austrian Academy of Sciences ÖAW/GIScience, Schillerstraße 30, A-5020 Salzburg, Austria gilbert.ahamer@oeaw.ac.at, adrijana.car@oeaw.ac.at, robert.marschallinger@oeaw.ac.at, gudrun.wallentin@oeaw.ac.at, fritz.zobl@oeaw.ac.at



ABSTRACT “Perspectives” are seen as the basic element of realities. We propose different methods to “map” time, sspace, economic levels and other perceptions of reality. IT allows views on new worlds. These worlds arise by applying new perspectives to known reality. IT helps to organise the complexity of the resulting views. Key Words: Geographic Information Science, mapping, time, space, perception. 0. LET’S START TO THINK 0.1 Our world is the entirety of perceptions. (Our world is not the entirety of facts.) 1. WRITING HELPS TO BECOME AWARE We ask: Is it possible to map = write 1. the distribution of material facts and elements in geometric space? (physics) 2. the distribution of factual events in global time? (history) 3. the distribution of real-world objects across the Earth? (geography) 4. the distribution of elements along material properties? (chemistry) 5. the distribution of growth within surrounding living conditions2? (biology) 6. the distribution of persons acting in relationships? (sociology) 7. the distribution of individuals between advantage and disadvantage? (economics) 8. the distribution of perspectives within feasible mindsets? (psychology) 9. the distribution of living constructs along selectable senses? (theology) We see: awareness results from reflection (Fig. 2).

elements objects events matter living conditions personalities advantages perspectives sense x y z space themes time = t



Figure 0: The human being perceives the world. Hence, every individual lives in a different world (Fig. 0). 0.2 The “indivisible unit”, the atom (ατομος1) of reality, is equal to one (human) perspective. Our world is made up of a multitude of perceptions, not of a multitude of realities and not of a multitude of atoms (Fig. 1).



Figure 1: The “primordial soup” of living, before the advent of (social) organisms: uncoordinated perspectives, uncoordinated world views. 0.3 In order to share one’s own conception with others, “writing” was invented. Similarly, complex structures, such as landscapes, are “mapped”. To map means to write structures.



Figure 2: Fundamental dimensions, along which to coordinate individual world views when reflecting.



                                                            

1



                                                            

2



what cannot be split any further (Greek)



životné prostredie (Slovak): living environment



UbiCC Journal – Volume 4 No. 3



609



Special Issue on ICIT 2009 Conference - Applied Computing



2. TIME CAN BE 1. an attribute of space (a very simple historic GISystem) 2. an independent entity (Einstein’s physics) 3. the source of space (cosmology). In terms of GIS item 2.1 is expressed as “t is one of the components of geo-data” i (Fig. 3).



Figure 3: The where-what-when components of geo-data, also known as triad (Peuquet 2002: 203). Time can be understood as • establishing an ordinal scale for events • driving changes (= Δ) of realities • something that unfortunately does not appear on paper. A proposed solution is to map changing realities (Δ) instead of mapping time. Time is replaced by what it produces. This is indicated in Fig. 4.

Δ elements Δ objects Δ events Δ matter (e.g. its path) Δ living conditions Δ personalities Δ advantages Δ perspectiv. Δ sense x y z space themes project! time = t



Figure 5: Notions of path in a geo-space: (a) Minard’s map of human losses during Napoleon’s 1812 campaign into Russia; and (b) its geovisualisation in a time cube (Kraak, 2009). Further examples such as landslides in geology, growth of plants, energy economics, economics will be shown in chapter 7. For implementing the idea to project the t axis onto the Δ axis we need to have clear insight how time quantitatively changes reality. In other words: we need a model, which (explaining how processes occur) determines the representation of time (Fig. 6). Examples are sliding geology, ΔGDP/cap, plant growth. One cannot perceive time (never!), only its effects: what was perceived in this time span (duration)4? This is why the t axis is projected onto another axis denoting the effect of elapsed time; what this means to the individual sciences is shown in Fig. 4. Very similarly, in physics nobody can feel force, only its effect (deformation, acceleration), and still forces have been undisputedly a key concept for centuries. What is time? Just a substrate for procedures. What is space? Just hooks into perceived reality. We retain from this chapter 2 that we need a clear model of how elapsing time changes reality. Then we can map time as suggested: by its effects.



Figure 4: The projection of time (t) onto the effects of time (the changes Δ) can apply to any science. This idea flips = projects the t axis onto one of the vertical axes. Time means then: how maps are changed by the envisaged procedures. Such procedures modify the variables along the axes, be they of physical (gravity force) or of social nature (war). A classical example is Minard’s map of Napoleon’s 1812 campaign into Russia3 (Fig. 5a, b).



                                                            

3



                                                            

4



Patriotic War (in Russian): Отечественная война



T. de Chardin’s (1950) concept of durée (French).



UbiCC Journal – Volume 4 No. 3



610



Special Issue on ICIT 2009 Conference - Applied Computing



3. HOW TO WRITE TIME? The big picture shows us various examples: 1. as a wheel (see the Indian flag): revolving zodiacs, rounds in stadiums, economic cycles, Kondratieff’s waves 2. as an arrow (see Cartesian coordinates): directed processes, causal determinism, d/dt, d²/dt² 3. as the engine for further improvement (evolutionary economics): decrease vs. increase in global income gaps, autopoietic systems, self-organisation 4. as the generator of new structures (institution building, political integration, progressive didactics): new global collaborative institutions, peer-review, culture of understanding, self-responsible learning, interculturality 5. as evolving construct (music). From this chapter 3 we only keep in mind that the concepts to understand and represent time are fundamentally and culturally different.



5. HOW TO MAP SPACE AND TIME? The detailed picture: it is obvious that a choice must be made for one mode of representation and for one view of one scientific discipline: 1. (x, y; t): cartography, GIS (Fig. 7) 2. (x, y, z; t): geology 3. (x, y, z; vx, vy, vz; t): landslides 4. (x, y, z; biospheric attributes; t): ecology, tree-line modelling 5. (countries; economic attributes; GDP/cap) or (social attributes; structural shifts; elapsing evolutionary time): economic and social facts in the “Global Change Data Base”6 (Fig. 8) 6. perceiving rhythms and structures: (only) these are “worth recognising”: music, architecture, fine arts.

objects seen by geographers



x y z space themes



harmonised world views!



time = t



Figure 6: All data representations require models. 4. HOW TO WRITE SPACE? The big picture shows us various examples: 1. as a container of any fact and any process (geography and GIS) 2. as result of human action (landscape planning) 3. as evolving construct (architecture). Examples span space as • received and prefabricated versus • final product of one’s actions, namely: 1. spaces as the key notion for one’s own science: everything that can be georeferenced means GIS 2. space as the product of human activity 3. expanding space into state space: the entirety of possible situations is represented by the space of all “state vectors” which is suitable only if procedures are smooth. The main thesis here is: the “effects of time” are structurally similar in many scientific disciplines, and they often imply “changes in structures” too. Information Technology (IT) is already providing scientific tools to visualise such structures.



5



Figure 7: Harmonising world views: GIS reunites world views by relating everything to its location. Different sciences may have considerably different outlooks on reality (Fig. 8). A humble attitude of recognising facts5 instead of believing in the theories one’s own discipline offers can empower people to survive even in the midst of other scientific specialties: Galileo’s (1632) spirit: give priority to observation, not to theories! This is the essential advantage of geography as a science: geographers describe realities, just as they appear. Such a model-free concept of science has promoted the usefulness of GIS tools to people independent of personal convictions, scientific models or theories.

objects seen by economists



x y z space themes



harmonised world views!



time = t



Figure 8: Different but again internally harmonised world views: explain facts from another angle.



                                                            

5



                                                            

6



datum (Latin): what is given (unquestionable)



This GCDB is described in Ahamer (2001)



UbiCC Journal – Volume 4 No. 3



611



Special Issue on ICIT 2009 Conference - Applied Computing



6. WHAT IT DOES, DID, AND COULD DO 6.1 IT helps to organise the multitude of views (= perceptions) onto data that are generated by humans: • IT constructs world views, such as: GIS, history, economics, geology, ecology etc. • IT has already largely contributed to demolishing traditional limitations of space and time: o Space: tele(-phone, -fax, -vision), virtual globes (Longley et al., 2001) o Time: e-learning, asynchronous web-based communication, online film storage (Andrienko & Andrienko 2006). 6.2 This paper investigates non-classical modes of geo-representation. We would like to point out that there are two already well-established fields that offer solutions to mapping (space and time, Fig. 9) views: Scientific and information visualisation are branches of computer graphics and user interface design which focus on presenting data to users, by means of interactive or animated digital images. The goal of this field7 is usually to improve the understanding of the data presented. If the data presented refers to human and physical environments, at geographic scales of measurement, then we talk about Geovisualisation, e.g. (MacEachren, Gahegan et al. 2004; Dykes, MacEachren et al. 2005, Dodge et al., 2008).



7. EXAMPLES The authors are members of the “Time and Space” project at their institution named “Geographic Information Science”8, a part of which explores the cognitive, social, and operational aspects of space & time in GIScience. This includes models of both social and physical space and consequences thereof for e.g. spatial analysis and spatial data infrastructures. We investigate how space and time are considered in these application areas, and how well the existing models of space and time meet their specified needs (see e.g. Fig. 9). This investigation is expected to identify gaps. Analysis of these gaps will result in improved or new spatio-temporal concepts particularly in support of the above mentioned application areas. 7.1 Sliding realities: geology The notion of the path in geography (x, y, t) is extended by the z axis (see item 5.2) which produces a map of “time”: Fig. 9 (Zobl, 2009).



Figure 9: Geology takes the (x, y, z; t) world view. The “effect of time” is sliding (luckily in the same spatial dimensions x, y, z): we take the red axis in Fig. 10. Space itself is sufficiently characteristic for denoting the effects of time.



Figure 9: Time series and 3 spatio-temporal data types (http://www.crwr.utexas.edu/gis/gishydro05/). 6.3 IT could develop tools that are then interchangeable across scientific disciplines, e.g. landslides that may structurally resemble institutional and economic shifts (see 7.1). IT could prompt scientists to also look at data structures from other disciplines. Whatever the disciplines may be, the issues are structures and structural change! Figure 10: These effects of time occur in space, most helpfully. Source: Brunner et al. (2003).



                                                            

8



                                                            

7



http://en.wikipedia.org/wiki/Scientific_Visualization



The overarching aim of the GIScience Research Unit is to integrate the “G” into Information Sciences (GIScience, 2009)



UbiCC Journal – Volume 4 No. 3



612



Special Issue on ICIT 2009 Conference - Applied Computing



7.2 Slices of realities: geology Despite the lucky coincidence that the effect of time (Δx, Δy, Δz) occurs in the same space (x, y, z) we try to produce slides carrying more information (item 5.3) and hence recur to the so-called attributes mentioned in Fig. 9 such as grey shades or colours. The speed of sliding (d/dt x, d/dt y, d/dt z) is denoted both by horizontal offsets and whitish colours in the spaghettis (Marschallinger, 2009) of Fig. 11.



7.5 Global deforestation One key driver for global change is deforestation; easy to map as change of land use category of a given area (Fig. 13).



  Figure 11: The (x, y, z; vx, vy, vz; t) view of a landslide process (shades of grey mean speed v).



Figure 13: The (x, y, z; Δ biospheric attributes; t): view of the global deforestation process in megatons carbon. Above: map of carbon flow, below: time series of GCDB data per nation symbolically geo-referenced by the location of their capitals. This representation is analogous to Fig. 11. In both, the focus shifts from maps(t) maps(t, Δt). Interest includes temporal dynamics: t = colour (above); Δt = height+colour (below), enriching the purely spatial interest. Even if to the aim is to enlarge the scope of the information delivered from the static map (Fig. 13 above) to the “dynamic map” (Fig. 13 below), readers will remain unsatisfied because no insight into the dynamic properties of deforestation is provided (Fig. 18). Increasingly, the viewer’s focus turns further from “facts” to “changes of facts”, to “relationships with driving parameters9” and to (complex social and political) “patterns10”. 7.6 Realities beyond slides But what if the information belongs to the social or economic realm (Fig. 14)? How to depict economic levels, education or policies? Figure 14: Example for graphic notation: one (hypothesised) parameter per nation (seen across the ).  Jordan =



7.3 Slide shows How to map spatial realities that are not any longer isotropic displacement vectors of space itself? For the example of changing tree lines in the Alps (Wallentin, 2009) a slide show is used to present the change of growth patterns made up of the multitude of individual agents (= trees = dots in Fig. 12). Moving spatial structures are depicted as a film of structures (item 5.4).



Figure 12: The (x, y, z; biospheric attributes; t) view of the Alpine tree line (above) and its shift induced by climate change as a slide show (below). In such processes which involve independent behaviour of autonomous agents (here: trees) it becomes seemingly difficult to apply a transformation of space itself, e.g. d/dt(x, y, z).



                                                            

see the suggested scenarios for water demand, water supply and water quality (Ahamer, 2008) 10 Patterns: name of the journal of the American Society for Cybernetics ASC

9



UbiCC Journal – Volume 4 No. 3



613



Special Issue on ICIT 2009 Conference - Applied Computing



7.7 Mapping social processes Social processes in social organisms can be described by the intensity of four different communicational dimensions (Fig. 15) along time: S = info, A = team, T = debate, B = integration. This type of writing (Fig. 16) resembles a score in musical notation11 and was invented for the webbased negotiation game “Surfing Global Change” (SGC), its rules are published in (Ahamer, 2004). The elementary particle of humanity’s progress – consensus building – is trained by SGC In this case, IT contributed to making communication independent from space and time: a web-platform enables asynchronous worldwide interaction of participants.



8. TRANSFORMATION OF COORDINATES 8.1 All the above examples have shown that • various “spaces” can be thought of • it would be suitable to enlarge the notion of “time”. 8.2 Suitably, a transformation of coordinates from time to “functional time” may be thought of. 8.3 In chapter 2, we suggested already to regard time as the substrate for procedures. Consequently, different “times” can be applied to different procedures. As an example, in theoretical physics, the notion of “Eigentime12” is common and means the system’s own time. 8.4 Similar to the fall line in the example of landslides in chapter 7.1 (red in Fig. 10) the direction of the functional time is the highest gradient of the envisaged process. This (any!) time axis is just a mental, cultural construction. 8.5 According to chapter 2 (Fig. 6) a clear understanding (mental model) is necessary to identify the main “effect of time”. We see that such an understanding can be culturally most diverse. Just consider the example of economic change: • optimists think that the global income gap decreases with development • pessimists believe that it increases, hampering global equity. 8.6 Therefore, any transformation of coordinates bears in itself the imponderability of complex social assumptions about future global development and includes a hypothesis on the global future. 8.7 Still, a very suitable transformation is t GDP/capita



Figure 15: Four basic components of any social procedure: learning information (Soprano S), forming a team (Alto A), debating (Tenor T), and integrating opposing views (Bass B).



(Fig. 17) both because of good data availability and increased visibility of paths of development. GDP/cap resembles evolutionary time.



time t = real time:



GDP/cap



≈ evolutionary time of development: complex graphic structure simpler graphical structure



Figure 16: A map of social processes in 4 dimensions during a negotiation procedure in a university course: participants show varying activity levels.



Figure 17: A suitable transformation of time uses the economic level, measured as GDP per capita.



                                                            

11



                                                            

12



partitura (Italian): score (in music)



literally (German): the own time (of the system)



UbiCC Journal – Volume 4 No. 3



614



Special Issue on ICIT 2009 Conference - Applied Computing



8.8 The strategic interest of such a transformation is “pattern recognition”, namely to perceive more easily structures in data of development processes. Examples for such “paths of development” are shown in Fig. 18 for the example of fuel shares in energy economics.



9. A FUTURISTIC VISION 9.1 Building on the vision of “Digital Earth” (Gore, 1998), the deliberations in this paper might eventually lead to the vision of “Digital Awareness”: the common perspective on realities valid for the global population, aided by (geo)graphic means. 9.2 The primordial element of (human and societal) evolution is consensus building. Without ongoing creation of consensus global “evolutionary time” is likely to fall back. The futuristic vision is to map global awareness.



  Figure 18: Structural shift of percentages of various fuels in all nations’ energy demand 196191. Data source: GCDB (Ahamer, 2001). 8.9 It is suggested here that implicitly during many mapping endeavours such transformation occurs. This is legitimate, but care must be taken to take into account the (silently) underlying model of human development. 8.10 Suitable transformation of coordinates can facilitate to see and communicate evolutionary structures, as it enables common views of humans and is therefore helpful for global consensus building. 8.11 Also the “effects of time” are projected into a common system of understanding which might give hope to facilitate common thinking independently of pre-conceived ideologies. This plan creates the “common reference system of objects”. 8.12 This paper suggests enlarging the concept of • “globally universal geo-referencing” (one of the legacies of IT) to • “globally universal view-referencing” • or “globally universal referencing of perspectives” 13. Fig. 19 illustrates this step symbolically. Figure 19: The global society perceives the world. 9.3 Much like the georeferenced satellites which circulate around the world produce a “Google, Virtual [or similar] Earth”, the individual spectators in Fig. 19 circle around the facts – and they create a “common virtual perception”: an IIS = Interperspective Information System.

the entirety seen by all global citizens



x y z space of themes



entirety of world views! time = t



                                                            

The facts themselves may well be delivered by endeavours such as Wikipedia but here it refers to the perspective on facts! A huge voluntarily generated database on people’s perceptions, views and opinions would be needed.

13



Figure 20: Divergent perceptions circulate around earthen realities. The entirety of world views creates the IIS (Interperspective Information System).



UbiCC Journal – Volume 4 No. 3



615



Special Issue on ICIT 2009 Conference - Applied Computing



9.4 Do we just mean interdisciplinarity? No. Nor do we simply refer to people looking into any direction. Fig. 21 shows the difference to IIS.



10. CONCLUSION Sciences are similar to “languages” spoken by people, they differ globally. Understanding for others’ languages is essential for global sustainable peace. Human perceptions are also strongly influenced by underlying models, assumptions and preconceived understandings. Studying geo-referenced data sets (GIS) can help to facilitate bridging interperceptional gaps. For the transformation of world views – to make them understandable – it is necessary to know about • the “effect of time”, namely the “path along the continuum of time” which a variable is expected to take • the speakers’ underlying model of a complex techno-socio-economic nature • the resulting perception of other humans. A future task and purpose of IT could be to combine the multitude of (e.g. geo-referenced) data and to rearrange it in an easily understandable manner for the viewpoints and perspectives of another scientific discipline or just another human being. Such a system is called Interperspective Information System IIS. Merging a multitude of perspectives to form a common view of the entire global population is the target of an IIS. Symbolically, a “Google Earth”-like tool would eventually develop into a “Google World Perspective”-like tool, or a “Virtual Earth”-like tool would become a “Virtual Perspective” tool encompassing all (scientific, social, personal, political, etc.) views in an easily and graphically understandable manner. In the above futuristic vision, IT can/should(!) become a tool to facilitate consensus finding. It can rearrange the same data for a new view. Symbolically speaking: similar to Google Earth which allows one to view the same landscape from different angles, a future tool would help to navigate the world concepts, the world views and the world perspectives of the global population. IT can reorganise extremely large data volumes (if technological growth rates continue) and could eventually share these according to the viewpoint of the viewer. Such a step of generalisation would lead from “Geographic Information Science” to “Interperspective Information Science”, implying the change of angles of perception according to one’s own discipline.



Figure 21: This is not IIS. 9.5 The science of the third millennium will allow dealing with a multitude of world views and world perspectives (see Tab. 1) with an emphasis on consensus building. When learning, the emphasis lies on social learning and may also make use of game-based learning (such as the web-based negotiation game “Surfing Global Change”) which allows to experimentally experiment with world views without any risk involved. Table 1: The science of the third millennium encompasses multiple perspectives element interaction perspective single ones manifold Mechanics Thermodynamics Logics Systems analysis Teaching



19th cent.



20th cent.



Social learning gaming, IIS



9.6 A suitable peaceful “common effort14” for a peaceful future of humankind would involve developing tools and visual aids in order to understand the opinions of other citizens of the globe. The future is dialogue. Or else there will be no future.



                                                            

(jihad in Arabic) also means: common effort of a society

14



UbiCC Journal – Volume 4 No. 3



21st cent.



616



Special Issue on ICIT 2009 Conference - Applied Computing



REFERENCES Ahamer, G. (2001), A Structured Basket of Models for Global Change. In: Environmental Information Systems in Industry and Public Administration (EnvIS). ed. by C. Rautenstrauch and S. Patig, Idea Group Publishing, Hershey, 101-136, http://www. oeawgiscience.org/ProjectFactSheets/Project FactSheet_GlobalChange.pdf. Ahamer, G., Wahliss, W. (2008), Baseline Scenarios for the Water Framework Directive. Ljubljana, WFD Twinning Project in Slovenia, http://www.oeaw-giscience.org/ProjectFact Sheets/ProjectFactSheet_EU_SDI.pdf. Andrienko, N., Andrienko G. (2006), Exploratory Spatial Analysis, Springer Brunner, F.K., Zobl, F., Gassner, G. (2003), On the Capability of GPS for Landslide Monitoring. Felsbau 2/2003, 51-54. de Chardin, T. (1950), La condition humaine [Der Mensch im Kosmos]. Beck, Stuttgart. Dodge, M., McDerby, M., Turner, M. (eds.) (2008) Geographic Visualisation, Wiley Dykes, J., A. MacEachren, et al. (2005). Exploring Geovisualization. Oxford, Elsevier. Galileo, G. (1632), Dialogo sopra i due massimi sistemi del mondo, tolemaico, e copernicano. Fiorenza. GIScience, (2008), Connecting Real and Virtual Worlds. Poster at AGIT’08, http://www.oeawgiscience.org/index.php?option=com_content&ta sk=blogcategory&id=43&Itemid=29. Gore, A. (1998). Vision of Digital Earth, http://www.isde5.org/al_gore_speech.htm. Kraak (2009), Minard’s map. www.itc.nl/personal/kraak/1812/3dnap.swf Longley, P.A. et al. (2001) Geographic Information. Science and Systems, Wiley MacEachren, A. M., M. Gahegan, et al. (2004). Geovisualization for Knowledge Construction and Decision Support. IEEE Computer Graphics & Applications 2004 (1/2): 13-17. Marschallinger, R. (2009), Analysis and Integration of Geo-Data. http://www.oeaw-giscience.org/. Peuquet, D. J. (2002). Representations of Space and Time. New York, The Guilford Press. Wallentin, G. (2009), Ecology & GIS. Spatiotemporal modelling of reforestation processes. See http://www.oeawgiscience.org/images/stories/Downloads/pecha% 20kucha%20technoz%20day.pdf Zobl, F. (2009), Mapping, Modelling and Visualisation of georelevant processes. http://www.oeaw-giscience.org/.



                                                            

GIScience goes way beyond this view of time and space (considering time as function) because it allows for much more complex queries and analyses.

i



UbiCC Journal – Volume 4 No. 3



617