Docstoc

CHARACTERIZATION AND PREDICTION OF RESOURCE AVAILABILITY IN GRIDS

Document Sample
CHARACTERIZATION AND PREDICTION OF RESOURCE AVAILABILITY IN GRIDS Powered By Docstoc
					 INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING &
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
                             TECHNOLOGY (IJCET)

ISSN 0976 – 6367(Print)
ISSN 0976 – 6375(Online)                                                        IJCET
Volume 4, Issue 4, July-August (2013), pp. 91-99
© IAEME: www.iaeme.com/ijcet.asp
Journal Impact Factor (2013): 6.1302 (Calculated by GISI)                     ©IAEME
www.jifactor.com




CHARACTERIZATION AND PREDICTION OF RESOURCE AVAILABILITY
                       IN GRIDS

                                   Chitra S1, Dr.Prashanth C.S.R2
                    1
                   Research Scholar, NHCE, Bangalore, VTU, Karnataka, India
  2
      HOD, Department of Computer Science and Engineering, NHCE, Bangalore, VTU, Karnataka,
                                             India



ABSTRACT

        Computational grids comprise heterogeneous collections of resources. Resources exhibit
different availability properties mainly due to administrators' policies for resource availability in the
Grid, and their failure/ unavailability properties. These make resources' availability predictions for
optimized resource selection for scheduling tasks, a challenging problem. Resources are highly
dynamic, unreliable and unpredictable in high performance computing environment, due to various
factors such as load of the system, available network bandwidth, system failure, network failure,
introducing new resources, performance degradation, contention among remote Grid task and the
local tasks of resources. Machines may fail at any time without advance notice. New machines can
join the system at any time. A node that is up at one time might be down at other time, and that is
busy at one time might be idle at another time. So there is temporal availability of resources. There
must be a mechanism to evaluate and manage the availability levels of grid resources while
scheduling the tasks. In this paper, a survey of some of the methods of characterization and
prediction of resource availability in grids are described.

Keywords: Multi-state Resource Availability Model, Prediction accuracy, Prediction Product Score,
Resource Utilization, Transitional Weighting Schemes.

I. INTRODUCTION

        A large number of heterogeneous resources such as computational machines and data storage
nodes, are managed by different owners such as companies, universities or other business or
scientific organisations. These resources are interconnected by high speed communication network.
The coordinated collaborative environment of resources located potentially at global scale, is utilized
for processing of tasks. Resource is identified with a network node which can be either an individual
machine or a cluster. Owners apply different policies for making the machines available; some will
be completely dedicated to the grid, others may only be donated when they would otherwise be idle,

                                                   91
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

and other policies may be between these extremes. The machines themselves will range from
individual laptops and desktop machines, to high performance clusters and supercomputers, which
have vastly different usage patterns and failure characteristics. The functional heterogeneity of
resources, machine failure characteristics, individual and organizational policies for resource usage,
and nodes joining or leaving the grid, will increasingly vary. Hence resources are highly dynamic,
unreliable and unpredictable in high performance computing environment, due to various factors
such as load of the system, available network bandwidth, system failure, network failure, introducing
new resources, performance degradation, contention among remote Grid task and the local tasks of
resources, etc. Availability of resource, mainly processor can be defined as the total time a
computing resource is functional during a given interval. There is temporal availability of resources
in case of a heterogeneous computing environment. Machines may fail at any time without advance
notice. New machines can join the system at any time. Machines may not be dedicated. A node that
is up at one time might be down at other time, and that is busy at one time might be idle at other
time. Resources in non-dedicated environment oscillate between being available and unavailable
based on resource and network failure characteristics, owner administrative policies, scheduling
mechanisms and application offered load. Also percentage of Resource Bandwidth allocated to local
and global jobs varies which again contributes to the temporal availability.
        There must be a mechanism to evaluate and manage the availability levels of grid resources
while scheduling the tasks. So the grid schedulers and resource brokers need information about
resource availability properties and predictions about their future availability. The goal is to
dynamically assign resources to tasks so as to maximize the probability of completing the entire task
execution within a desired total response time.

II. LITERATURE REVIEW

2.1 Multi-State Grid Resource Availability Characterization[1][3]
        In computational grids, resources availability is highly dynamic. Resources will transition
between availability modes. Different resources will exhibit varying unavailability characteristics.
Site autonomy is an important attribute of grids where local jobs take precedence over remote jobs.
Schedulers should know about resources failure and also application characteristics. Resources
may be highly available, gracefully unavailable, in machine down state, or intermittently available.
Different types of unavailability and patterns have to be predicted. Availability predictions are based
on failure characteristics of the machines , unavailability characteristics , application characteristics,
policies of resource owners, the scheduling policies and mechanism of the grid middleware, and the
characteristics of the grid’s offered application load.
        Resources can be classified into five availability states. The multi-state model has 5
availability types. They are Available - when the CPU load is less than 50%; CPU Threshold
exceeded - when the local load is greater than 50%; User Present ; Job Eviction and Unavailable
state. Three of the availability states are based on the CPU load level. The two other states indicate
memory thrashing and resource unavailability. If a job is running or suspended, and enters the Job
Eviction state, it is a graceful transition to the Unavailable state. A transition directly to Unavailable
is ungraceful.
        Grid resources will have different characteristics in terms of how long they stay in each
availability state, how often they transition between the states, and which states they transition to.
Different applications will behave differently on different resources. If local job is present, the job
can be suspended and restarted after some time interval. If the local job is present for longer, then
the job can be terminated on that resource and scheduled on the next priority resource. If a standard
universal job is suspended and then eventually gracefully evicted, it could checkpoint and resume on
another machine. If there is an ungraceful transition, then the most recent periodic checkpoint is
required. If the host resource transitions gracefully to Unavailable, then a checkpointable application

                                                   92
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

need not be restarted from the beginning. A job that is not checkpointable must restart from the
beginning, even when gracefully evicted. Previous N days of a resource’s history is used to produce
a Markov chain for state transition predictions. Confidence interval predictions for availability
duration based on model-fitting can be done. Time-stamped CPU load as a percentage and idle time
in seconds can be recorded. Idle times of zero imply user presence, and the absence of data indicates
that the machine was down. if its local (i.e. non-Condor) load is above 50%, it is said to be CPU
Threshold Exceeded state Otherwise, a machine that is online with no user present and CPU load
below 50% is considered Available.
        The average availability duration determines whether the application is likely to have
completed, given by probability that an application is likely to get executed completely. How a
machine transitions to Unavailable (gracefully or ungracefully) determines whether the job can
checkpoint before migrating. The machines are classified based on average availability duration
such as high, medium high, medium low and low, and then the number of graceful and ungraceful
transitions to Unavailable per day. Jobs may be completed jobs, gracefully evicted jobs or
ungracefully evicted jobs. Availability during a certain time interval, the checkpointability and
expected runtime of a job along with predicted resource availability can suggest better mappings of
tasks to resources in a task scheduling problem in the heterogeneous computing environment.

2.2 Resource Availability in Enterprise Desktop Grids [2]
        Desktop grids use the idle cycles of many desktop PC’s. There are three types of availability
in the availability trace of desktop grids.
(i) Host availability: This is a binary value that indicates whether a host is reachable. Causes of host
unavailability include power failure, or a machine shutdown, reboot, or crash.
(ii) Task execution availability: This is binary value that indicates whether a task can execute on the
host or not, according to a desktop grid worker’s recruitment policy. The recruitment policy consists
of a CPU threshold, a suspension time, and a waiting time. The CPU threshold is some percentage of
total CPU use, for determining when a machine is considered idle. The suspension time refers to the
duration that a task is suspended when the host becomes non-idle. When a busy host becomes
available again, the worker waits for a fixed period of time called the waiting time before starting a
task. Causes of execution unavailability include prolonged user keyboard/mouse activity, or a user
compute-bound process.
(iii) CPU availability: This is percentage value that quantifies the fraction of the CPU that can be
exploited by a desktop grid application. Factors that affect CPU availability include system and user
level compute-intensive processes.
        Host unavailability implies execution unavailability, which implies CPU unavailability. In
which case no new task may begin execution and any executing task would fail. If there is a period
of task execution unavailability (e.g., due to keyboard/mouse activity), then the desktop grid worker
will stop the execution of any task, causing it to fail, and disallow a task to begin execution implies
CPU unavailability. However CPU unavailability does not imply execution unavailability. For
example, a task could be suspended and therefore have zero CPU availability, but since the task can
resume and continue execution, the host is available in terms of task execution. Similarly, execution
unavailability does not imply host unavailability. For example, a task could be terminated due to user
mouse/keyboard activity, but the host itself could be still up. The trace would log CPU availability
in terms of the CPU time a real application would receive if it were executing on that host. The trace
would record execution availability when failures occur. The trace would determine the cause of the
failures (e.g., mouse/user activity, machine reboot or crash). From this, one can find the temporal
structure of availability intervals. Availability interval is the interval of time in between two
consecutive periods of exec unavailability. Given two hosts with identical availability interval
lengths, one would prefer the host with smaller unavailability intervals. Using availability and
unavailability interval data, one can predict whether a host has a high chance of completing a task by

                                                  93
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

a certain time. Based on the characterization of the temporal structure of resource availability, it is
possible to derive the expected task failure rate, that is the probability that a host will become
unavailable before a task completes, from the distribution of number of operations performed in
between periods of unavailability based on random incidence. The expected task failure rate is
strongly dependent on the task lengths.

2.3 On the Dynamic Resource Availability in Grids [4]
         It is important to know about the characteristics of the grid resource availability, and of the
impact of resource unavailability on the performance of grids. System resources may be at any time
out of reach due to distributed resource ownership, scheduled maintenance, or unpredicted failures.
Many of the grids comprise computing resources grouped in clusters, whose owners may share them
only for limited periods of time. Often, many of a grid’s resources are removed by their owner from
the system, either individually or as complete clusters, to serve other tasks and projects.
         The grid resource availability model, considers the time when resource failures occur, the
duration of a failure, the number of nodes affected by a failure, and the distribution of failures per
grid cluster.
         The statistical distributions fitted onto the availability data are Log normal, Normal, Weibull,
Exponential and Gamma distribution these distributions are fitted using Maximum Likelihood
Estimation. Then goodness of fit tests are performed to assess the quality of the fitting for each
distribution, and to establish a best fit for each of the model parameters. For each distribution d, the
hypothesis that the Grid availability data comes from the distribution d, whose parameters are found
during the fitting process (the null-hypothesis of the goodness-of-fit test) is formulated. The
Kolmogorov-Smirnov test (KS-test) is used for testing the null-hypothesis. The KS-test statistic, D,
estimates the maximal distance between the Cumulative Distribution Function(CDF) of the empirical
distribution of the input data, and that of the theoretical distribution. The null-hypothesis is rejected
if the D is greater than the critical value obtained from the KS-test table. A lower value of D
indicates better similarity and a higher degree of similarity between the input data and data sampled
from the theoretical distributions. This latter property is used to select the best fits. It is found that the
best fits for the inter-arrival time between failures, the duration of a failure, and the number of nodes
affected by a failure, are the Weibull, Log-Normal, and Weibull, respectively. The results for the
inter-arrival time between failures are alarming: the shape parameter of the Weibull distribution, our
best fit, indicates an increasing hazard rate with strong effects on the ability of grids to execute long
jobs (even single-processor).
The five availability-aware performance metrics, each adapted from a traditional metric considered
are:
(i) Utilization -defined as the percentage of resources consumed by the system users, from the total
resources present in the system, over a period of time. The ideal utilization value is 100%. However,
due to resource fragmentation and other reasons, a utilization of 60-70% is considered high for
systems that run parallel jobs.
(ii) Wait time
(iii) Response time
(iv)Normalized Throughput - throughput traditionally characterizes the number of jobs finished
during a time interval, divided by number of processors in the system.
(v) Normalized Goodput - The goodput characterizes the amount of resources consumed by all jobs
towards their completion (this excludes the amount of time spent waiting in queues or for data to
arrive), divided by the number of processors in the system.
Four models of grid availability information are
1. Systems with Steady Availability (SA): This model assumes that all resources are online at all time.
Many resource management results are readily available for these steady systems.


                                                     94
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

2. Systems with Known Availability (KA): This model assumes a system with dynamic resource
availability. However, the information regarding availability is perfect (complete and on-time).
3. Systems Automated Monitoring of Availability (AMA): This model assumes a system with dynamic
resource availability. It also assumes that the most recent resource availability information is
available from a monitoring system, which samples periodically the grid for individual computing
nodes’ availability. If the monitoring period is high, the monitoring information can be stale; if it is
low, the monitoring overhead is unbearable for the grid.
4. Systems with Human Monitoring of Availability (HMA): This model is similar to the AMA model,
but assumes that the availability information is provided by the (human) system administrator at
fixed, but relatively large intervals: 1 week or 1 month for instance.
When the scheduler wrongly considers that there are enough number of idle processors for the job,
the job execution fails due to a situation called job submission failure. When a job fails when due to
at least one processor used by this job crashing, the situation is called job execution failure.

2.4 Characterizing, Modelling and Predicting Dynamic Resource Availability in a Large Scale
Multi-Purpose Grid [5][7]
        The resource's components, Grid middleware maturity, varying resource maintenance and
failure properties make resource availability characterization, modelling and prediction important.
An analysis of availability of resources in a large-scale, multi-cluster and multipurpose Austrian Grid
is described.
Three main classes of resources in Austrian Grid are :
(i) Dedicated resources : The dedicated resources are meant to be always available to the Grid users.
(ii) Temporal resources : Resources from university labs, referred as temporal resources are
available in the Grid as long as they are turned on.
(iii) On-demand resources: They are resources that are made available to the Grid only on demand
from the Grid users for large scale jobs.
Availability analysis is done over time, in a year, hour of the day and day of the week.
Two classes of availability are:
(i) Lower availability at weekends, and higher during the working week days. This class has an
increasing availability till Tuesday or Wednesday and then decreasing till weekend.
(ii) Higher availability on weekends and lower during the working weekdays. This class has a
decreasing availability till Wednesday and then increasing availability till Saturday.
Resource (un)availability is modelled at individual classes level to find mathematical model of their
(un)availability over time. Different models are found best for different resource classes, which
collectively model the availability of the whole Grid. Lognormal, Pareto and Weibull were found the
distributions most closely satisfying the availability duration distributions in dedicated, temporal and
on demand resources respectively. Resource survival analysis with PDF, survival function, and
hazard function helps in analyzing these availability models.
Kolmogorov-Smirnov test (KStest) is used to test the null hypothesis for goodness-of-fit of a
distribution di, where di represents a distribution model that is to be tested. Null hypothesis is: The
resource class availability data comes from the distribution di., whose parameters were estimated in
the last step. The KS-test compares the empirical distribution function of availability data with that
of the distribution specified in null hypothesis. Kolmogorov-Smirnov statistic D, for a given test,
estimates the maximum distance at any point between the two distributions. The null hypothesis is
rejected if the test statistic D is greater than the critical value obtained from a KS-table. The higher
the value of D, the higher the degree of dissimilarity between the availability data and the data from
the tested distribution. This property is used to select the best fit, and find that the best fit for
availability durations are Lognormal, Pareto and Weibull for dedicated, temporal and on-demand
classes of resources respectively, whereas best fit model for overall Austrian Grid is found to be
Lognormal. The overall probability distribution of Grid model of (un)availability is shaped from the

                                                  95
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

probability distribution models of underlying different classes of resources, which are based on their
different policies of availability to the Grid. Hence, lognormal dominates for the three resource
classes, which is due to dominating size of dedicated resources class in the Austrian Grid. The hazard
rate gives the instantaneous probability of failure/unavailability given the survival to a given time. It
is the Probability Distribution Function divided by survival function. The survival function gives the
probability of survival as a function of time. It is simply 1 minus cumulative distribution function (1-
CDF). In the selected Best fit models the hazard rate turns out to be increasing in temporal and on-
demand resources, which means that the resources are more susceptible to failure/unavailability as
time passes due to aging. The hazard rate in dedicated resources decreases which means that
resources get more stable as the time increases. The survival function of dedicated resources
increases, indicating that resources have higher probability of availability for the longer durations.
Contrarily, decreasing survival functions in temporal and on demand resources show a lower
probability of availability for longer durations. Resource unavailability includes resource
maintenance and failure properties, and their policies for unavailability in the Grid. Resource
(un)availability properties are modelled as Mean Time to Reboot(MTR). For dedicated resources,
Weibull, Lognormal and Pareto seem to closely model the real trace data and for temporal and on-
demand resources it is Weibull and Lognormal.
        Considering resource dynamic availability properties and patterns over time and its current
availability behaviour, two methods from pattern recognition and classification the Bayes' Rule and
Nearest Neighbor Rule are employed to serve resource instance or point availability and duration
availability predictions. Instance or point availability describes resource availability at a certain
point of time. It refers to the next monitoring instance. Duration availability describes resource
availability over a certain time duration. It refers to the immediate next duration of a certain time
span. These two methods utilize knowledge of resource past and current availability behaviour,
gained in the characterization phase, to serve its future availability predictions.
Bayes' rule exploits resource availability characteristics from past traces and current behaviour.
According to Bayes' rule, the probability of class ω when a feature x is selected is:



where p(x) is the PDF of x and for our case of two classes                              )
Decide w1 if                                 ; otherwise w2.
Nearest Neighbour rule is a well known pattern classification technique, which gives the nearest
neighbour value as a prediction. The last monitored status is taken as the nearest neighbour. This
method typically suits the machines with high Mean Time Between Failures (MTBF) and Mean
Time to Reboot (MTR) , and has a theoretical per day error rate of:




where fd denotes number of transitions from available to unavailable state per day, and Md represent
number of resource states monitored per day.
Prediction accuracy = (No. of true predictions)/(Total number of prediction queries) represented as
percentage.
 For Instance availability prediction, a prediction is treated true if the resource immediate next status
is the same as predicted, otherwise it is considered false. In case of duration availability prediction,
prediction is treated as true if a resource is predicted to be available for a certain immediate duration

                                                   96
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

and resource is available for that duration or longer, otherwise it is considered false. In case when
resource is available for a duration lesser than the predicted, the accuracy of prediction is calculated
as the ratio of actual available duration to the predicted. The accuracy was evaluated for all the
resources in every resource class and accuracy at class level is presented by taking their average.
Different resources are compared based on their availability properties like daily availability, hourly
availability, MTBF, MTR etc. Resources are also compared based on their availability distribution in
different durations and distribution of these durations over time.

2.5 Resource Availability Prediction for Improved Grid Scheduling [6]
         A multi-state availability model can help improve scheduling performance by capturing the
various ways a resource may be available or unavailable to the grid. Several prediction techniques to
forecast resource transitions into the model’s states are described. Computational grids comprise
heterogeneous collections of resources. A failure predictor, which forecasts the availability of
candidate resources for the period of time that an application is likely to run on them, can help
schedulers make better application placement decisions. If some particular resource is expected to
become unavailable because its owner reclaims the machine for local use (as opposed to the machine
failing unexpectedly), then that machine may be a suitable candidate for scheduling an application
that can take an on-demand checkpoint.
         A machine is said to be in the Available state if it is currently running with network
connectivity, has no user present, and a local CPU load of below the CPU threshold, and can be used
by the grid. A resource may transition to the User Present state if the keyboard or mouse is touched
or to the CPU Threshold Exceeded state if the local CPU load becomes too high. If the resource
remains in either of these suspension states too long, if the job is evicted while running for any other
reason, or if the machine is shut down, it transitions to the Job Eviction state. Finally, if a machine
fails or becomes unreachable, it directly transitions to the Unavailable state.
         Grid resources will have different characteristics in terms of how long they reside in each
availability state, how often they transition between the states, and the states to which they transition.
Different applications will behave differently on different resources. Grid applications are diverse. A
checkpointable application need not be restarted from the beginning if its host resource transitions
gracefully to Unavailable, but not all jobs are checkpointable. Longer jobs will experience more
faults. If a checkpointable job is suspended and then eventually gracefully evicted, it can checkpoint
and resume on another machine. An ungraceful transition requires using the most recent periodic
checkpoint. A job that is not checkpointable must restart from the beginning, even when gracefully
evicted. Characteristics of applications, such as checkpointability and expected runtime, can
influence the effectiveness of scheduling those applications on resources that behave differently
according to their transitions between the availability states.
         A multi-state prediction algorithm takes as input a length of estimated application execution
time and uses a resource’s availability history to predict the probabilities of that resource next exiting
the available state into each of the non-available states, and to remain available throughout the
interval. These probabilities sum upto 100%. The availability predictor outputs four probabilities,
one each for entering User Present, CPU Threshold Exceeded and Unavailable states next, and one
for the probability of completing the interval without leaving the Available state (Completion). There
are two approaches for predictions. The first approach examines a resource’s past N days of
availability behaviour during the interval being predicted (N-Day or Day). Another approach
examines a resource’s most recent N hours of activity immediately preceding the prediction
(NRecent). The predictors analyze the segments of resource availability history in two ways. The
first makes predictions based on the resource’s state Durational behaviour by calculating the
percentage of time spent in each state during the analysis period. The state that is occupied majority
of the time could be the most likely next state. The second and more successful approach considers a
resource’s Transitional behaviour by counting the number of transitions from available to each of the

                                                   97
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

other states. In this approach, when examining a section of a resource’s availability history, the
number of transitions from available to each of the other states is counted. Also, for every interval of
time a resource was available, the number of times the requested job duration (e.g. a ten hour
availability interval and a two hour job means five completions) could be completed is counted. The
probabilities for each exit state as well as the completion state are calculated by summing each
state’s transition count and dividing the total number of transitions for each state by the total number
of transitions. For both the Durational and Transitional approaches, when combining different
sections of availability (e.g. when combining the N days’ probabilities together), the probabilities for
each state are averaged. The Transitional schemes weigh all transitions equally (Equal weighting
scheme). The predictors weigh transitions according to the time of day (Time). This means that for
the Time weighting scheme, the closer a transition time is to the time of day at which the prediction
occurs, the higher the weight of that transition. Freshness weighting gives a higher weight to
transitions that occur closer to the time of the prediction and likewise, a smaller weight to events that
occurred further in the past. Predictor accuracy is defined as the ratio of correct predictions to the
total number of predictions. A correct prediction is one for which the machine is predicted to exit on
a certain non-available state and it does, or for which the machine is predicted to remain available
throughout the interval, and it does.
The Prediction Product Score (PPS) algorithm calculates each resource i’s score as



where ci is the resource’s predicted completion probability during the application’s expected
execution interval as given by the predictor, mi is it’s MIPS score, and li is the CPU load currently on
resource i. At job scheduling time, the predictor calculates this score for each available resource, and
chooses the resource with the highest score.

III. CONCLUSION

        The availability and performance of the resources in grids, which are already heterogeneous,
vary dynamically, even during the course of task execution. Schedulers and resource brokers need
information about resource availability, properties and predictions about future availability. An
availability predictor, which forecasts the availability of candidate resources for the period of time
that an application is likely to run on them, can help schedulers make better application placement
decisions. Resource categorization into multi-states helps in distinguishing between graceful and
ungraceful transitions to unavailability. The accuracy of prediction is improved through transition
weighting schemes. Resource (un)availability is modelled at individual classes level to find
mathematical model of their (un)availability over time. Different models are found best for different
resource classes, which collectively model the availability of the whole Grid. Selecting resources that
will remain available for the duration of application runtimes would improve the performance of
both individual applications and of the heterogeneous computing environment as a whole.

REFERENCES

 [1]   Brent Rood and Michael J. Lewis," Multi-State Grid Resource Availability Characterization",
       IEEE 8th Grid Computing Conference, 2007
 [2]   Derrick Kondo, Gilles Fedak, Franck Cappello, Andrew A. Chien, Henri
       Casanova,"Resource Availability in Enterprise Desktop Grids"
 [3]   D. Kondo, M. Taufer, C.. Brooks III, H. Casanova and A. Chien, "Characterizing and
       Evaluating Desktop Grids: An Empirical Study," IPDPS'04, 2004, p. 26b.


                                                   98
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME

 [4]  Alexandru Iosup, Mathieu Jan, Ozan Sonmez, Dick H.J. Epema," On the Dynamic Resource
      Availability in Grids",IEEE 8th Grid Computing Conference, 2007
 [5] Farrukh Nadeem, Radu Prodan, Thomas Fahringer, "Characterizing, Modeling and Predicting
      Dynamic Resource Availability in a Large Scale Multi-Purpose Grid", Eighth IEEE
      International Symposium on Cluster Computing and the Grid,2008
 [6] Brent Rood and Michael J. Lewis," Resource Availability Prediction for Improved Grid
      Scheduling"
 [7] D. Kondo et al. Characterizing and evaluating desktop grids: An empirical study, 2004.
 [8] J. Nabrzyski, J. M. Schopf, and J. Weglarz, editors. Grid Resource Management: State of the
      Art and Future Trends, Chapter 26. Kluwer Publishing, 2003.
 [9] Dr. Damanjeet Kaur, “Smart Grids and India”, International Journal of Electrical Engineering
      & Technology (IJEET), Volume 1, Issue 1, 2010, pp. 157 - 164, ISSN Print : 0976-6545,
      ISSN Online: 0976-6553.
 [10] D.Asir, Shamila Ebenezer and Daniel.D, “Adaptive Load Balancing Techniques in Global
      Scale Grid Environment”, International Journal of Computer Engineering & Technology
      (IJCET), Volume 1, Issue 2, 2010, pp. 85 - 96, ISSN Print: 0976 – 6367, ISSN Online:
      0976 – 6375.
 [11] Arjun Singh, Dr. Prasun Chakrabarti and Surbhi Chauhan, “Grid Resource Discovery: A
      Survey and Comparative Analysis”, International Journal of Computer Engineering &
      Technology (IJCET), Volume 4, Issue 3, 2013, pp. 550 - 559, ISSN Print: 0976 – 6367,
      ISSN Online: 0976 – 6375.




                                              99

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:0
posted:7/26/2013
language:
pages:9