Docstoc

Comparative Study and Performance Enhancement of Scaled out Cloud Architecture of Amazon web Services

Document Sample
Comparative Study and Performance Enhancement of Scaled out Cloud Architecture of Amazon web Services Powered By Docstoc
					   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 1, Issue 4, November – December 2012                                    ISSN 2278-6856



          Comparative Study and Performance
      Enhancement of Scaled out Cloud Architecture
               of Amazon web Services
                                     D.S. Chauhan1, Manjeet Gupta2 and Brijesh Kumar3
                                      1
                                       Vice Chancellor, Uttrakhand Technical University, Uttrakhand
                         2
                             Assistant professor, Department of Computer Science and Engineering, JMIT, Radaur
                     3
                         Assistant professor, Department of Computer Science and Engineering, Lingaya’s University,
                                                              Faridabad
Abstract: The introduction of affordable infrastructure on           of cluster systems, we have proposed a virtual machine
demand, specifically Amazon’s Elastic Compute Cloud                  PC cluster in which virtualization is applied to a PC
(EC2), has had a significant impact in the business IT               cluster that uses a general-purpose personal computer for
community and provides reasonable and attractive                     each node [4].In addition, we have used IP-SAN for
alternatives to locally-owned infrastructure. For scientific         storage access which can realize long-distance connection
computation however, the viability of EC2 has come into              at low cost over a high latency network [3].In this paper,
question due to its use of virtualization and network shaping        we have confirmed that the execution time becomes
and the performance impacts of both. Several works have              shorter with the migration of virtual machines even the
shown that EC2 cannot compete with a dedicated HPC cluster           cost of migration is taken into account, compared with the
utilizing high-performance interconnects, but how does EC2           case of accessing data over a network when an I/O-
compare with smaller departmental and lab-sized commodity            intensive application is executed [5]. Moreover, we have
clusters that are often the primary computational resource
                                                                     measured basic performance on Amazon Elastic Compute
for scientists? Because of the increasing volume of
                                                                     Cloud (EC2) [1], throughput of network and iSCSI disk
information distribution and data accumulation by means of
                                                                     access, as basic data of a real cloud computing system.
broadband networks, the amount of available information is
                                                                     VMware [2] and Virtual PC [3] are known to realize a
increasing in recent years. In this study, we aim to acquire
insufficient resources dynamically from cloud computing
                                                                     virtual computing environment. They enable us to install
systems while basic computation is performed on its own
                                                                     guest OS on top of host OS, and therefore, they have a
local clusters. Scalable resource management is achieved by          problem of processing performance degradation compared
monitoring resource usage of own local cluster and                   with the case of real machine because guest OS works as
insufficient resources are acquired dynamically from cloud           an application on host OS. On the other hand, Xen [4]
computing. We have confirmed that the execution time                 provides a basic platform of virtualization called Virtual
becomes shorter with the migration of virtual machines even          Machine Monitor, on which multiple OSes operate as
the cost of migration is taken into account, compared with the       shown in Figure1[4]. Since the overhead of virtual
case of accessing data over a network when an I/O-intensive          machine is reduced, Xen can achieve higher performance
application is executed. Moreover, we have measured basic            which is close to the case of real machine. Xen is used
performance on Amazon Elastic Compute Cloud (EC2) as                 even in a business field recently, because Xen achieves
basic data of a real cloud computing system.                         remarkably high performance as open source software
                                                                     [5].In the architecture of Xen, Virtual Machine Monitor
Keywords: Cloud computing, Amazon                   cloud, EC2,
                                                                     is foundation for virtualization [2] and virtual machines
Virtualization.
                                                                     called domain are allocated on top of it. Domain0 behaves
                                                                     as a host OS and DomainU behaves as guest OS.
1. INTRODUCTION                                                      Domain0 has a privilege to access physical hardware
In this study, we aim to acquire insufficient resources              resources and to manage other domains [1].
dynamically from cloud computing systems while basic
computation [3] is performed on its own local clusters.
Because of the increasing volume of information
distribution and data accumulation by means of
broadband networks [2], the amount of available
information is increasing in recent years. Scalable
resource management is achieved by monitoring resource
usage of own local cluster and insufficient resources are                     Figure 1 Architecture of XEN[4]
acquired dynamically from cloud computing[1]. In terms

Volume 1, Issue 4 November - December 2012                                                                            Page 24
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 1, Issue 4, November – December 2012                                    ISSN 2278-6856


Using virtualization software, a system can be constructed      addition, iSCSI is expected for the realization of the long-
as multiple computers that operate virtually on a single        distance remote access because IP network infrastructure
computer [7]. In this study, we have used virtual machine       is widely deployed and maintained in wide area networks.
in worker nodes of virtual machine PC cluster [8]. As we        IP-SAN is expected to be used not only remote storage
use virtual machine as worker nodes, flexible                   like data center but also outsourcing computer resources
management of infrastructure can be introduced                  in a cloud computing framework. Figure 2 shows [5] the
depending on system load and service demand [9]. In             layered structure of iSCSI. iSCSI encapsulates a SCSI
addition, the migration mechanism can be introduced that        command within a TCP/IP packet and transmits the
migrates virtual machine to another node while                  volume of data between server (Initiator) and storage
maintaining the state of running applications.                  (Target). In the future, since the gigabit/ 10 gigabit class
                                                                line is expected to be more popular by development of the
2. BACKGROUND TECHNOLOGIES                                      Internet, iSCSI will be effective furthermore. In this
                                                                paper, we have used iSCSI protocol and constructed a
   2.1 Cloud Computing                                          virtual machine PC cluster assuming wide area network
Cloud computing is an idea that users are allowed to use        environment communication.
computing resources including server machines and
storage across the Internet without [6] being conscious of      3. EXECUTION OF DATA-INTENSIVE APPS
its neither existence nor inner structure. The usage of
remote data center and cloud computing should increase             3.1 Evaluation of PC cluster consolidated with
in the near future, because it is expected to reduce the        IPSAN
introduction and management cost of information                 PC cluster consolidated with IP-SAN is introduced in [6]
systems, which is better than the case that all computing       which consolidate back-end SAN that connects between a
resources must be purchased at the user site, like              node (server) and storage and front-end LAN that
Hardware as a service (Haas)[8].Amazon EC2 is one of            connects among nodes by using iSCSI. In the case of PC
typical cloud computing systems, which is a rental service      cluster consolidated with IP-SAN, both the back-end SAN
of virtual machines on Internet. In the back-end of             and front-end LAN can be unified into a single
Amazon EC2 [2][4][5], server virtualization technology is       commoditized network built with TCP/IP and Ethernet by
used and users can back up whole image of own OS.               using iSCSI, as shown in Figure 3. Therefore, the
Flexible system operation is possible because a new             reduction [6] of network construction cost and the
machine can start up in minutes when a system load              increase in efficiency of operational management can be
increases suddenly.                                             expected. However, since both back-end and front-end
                                                                networks use the same network resources, it is concerned
                                                                the communication packets transmitted between nodes
                                                                collide with packets of storage access on the same
                                                                network, and performance deteriorates as a result. On PC
                                                                cluster consolidated with IP-SAN, we have executed Hash
                                                                Partitioned Priory (HPA) [7] and Parallelized FP-growth
                                                                (PFP) which are parallel association rule mining, and
                                                                mpiBLAST [8] which is a sort of parallelized large-scale
                                                                scientific computation in bioinformatics. According to the
                                                                result of above experiments, the feature of total
                                                                performance of the system becomes CPU-bound or I/O-
                                                                bound, not network-bound in these cases even though
                                                                iSCSI network is consolidated.


             Figure 2 Configuration of iSCSI

Recently, since the amount of data processed in an
information system has become huge, SAN is introduced
to connect storage, and it has come to be widely used.
SAN unifies the distributed storage held by each node,
and realizes efficient practical use with central control for
disk resources [6]. IP-SAN is expected as SAN of the next
generation built with IP network. Internet Small
Computer System Interface (iSCSI) [5] is the most
popular protocol of IPSAN and we can build SAN with                    Figure 3 PC cluster consolidated with IP-SAN
inexpensive Ethernet and TCP/IP using iSCSI [3]. In

Volume 1, Issue 4 November - December 2012                                                                        Page 25
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 1, Issue 4, November – December 2012                                    ISSN 2278-6856


  3.2 Benchmark Execution with remote storage                simulates a decision support system, and this is an I/O-
    access                                                   bound application that inserts and deletes data to
In [9], the behavior of a virtual machine PC clusters        databases and queries are executed repeatedly. Remote
executing an I/O-bound application over high latency         storage access has caused I/O-bound in the application
networks has been analyzed. Specification of each node of    execution, and therefore significant execution time
the cluster is shown in Table 4. A virtual machine           difference was observed as the delay becomes longer.
(DomainU) has been created for each worker node. We
have monitored a virtual machine PC




                                                                     Figure 5 Execution time of OSDL-DBT3 with
                                                                                remote storage access
     Figure 4 Experimental environment

cluster using a monitoring tool Ganglia [10] including
iSCSI communication to access remote storage. We have
built a virtual machine PC cluster like Figure 4 that
contains four servers (Initiator) and two iSCSI storage
nodes (Target). We have executed a database benchmark
Open Source Development Labs Database Test3
(OSDLDBT3) [12] and compared the performance of two
cases in which job is given to Domain0 and DomainU,
respectively. A virtual machine (DomainU) is created for
each worker nodes. For storage access we have used
iSCSI. Two servers access the storage in a local area and
other two servers access the remote storage. To construct
a remote access environment, we have inserted Dummy                   Figure 6 Execution time of HPA with remote
net [11] that simulates delay artificially between a local                          storage access
site and a remote site, and we have inserted delay as RTT
of remote iSCSI access is 1msec, 2msec, 4msec, 8msec,
16msec.The execution time of OSDL-DBT3 is shown in           4. EVALUATION OF CLUSTER WITH SERVER
Figure 5. According to the graph, execution time is          MIGRATION
increasing as delay becomes longer from 1msec to
16msec in iSCSI access [1][5]. Especially, severe increase     4.1 Basic Experiment
in execution time is observed when delay is longer than      First, the performance of a real machine PC cluster and a
4msec. On the contrary, Figure 6 shows execution time of     virtual machine PC cluster has been compared with a
HPA with the data size of 20 megabytes. HPA is a             basic experiment. We have built the cluster connecting
parallel application that parallelizes association rule      four servers (Initiator) and a storage node (Target) with
mining which is based on Priory algorithm using a hash       iSCSI, and executed OSDL-DBT3 on each environment
function, and generates frequent item sets from candidate    (Figure 7) [5]. Worker nodes on the real machine cluster
item sets repeatedly. Although HPA is data mining that       have 4 gigabytes memory, and we have divided it to two 2
processes huge amount of transaction data, since it          gigabytes and assigned them for Domain0 and DomainU
includes heavy CPU processing,[11] it is not I/O-bound.      respectively in the case of the virtual machine cluster.
Therefore the difference of execution time is small even     The execution time is shown in Figure 8. From this
in the case of longer delay. On the other hand, OSDL-        result, performance of DomainU (virtual machine)
DBT3 that is simplified from Transaction Processing          degrades a little compared with that of the real machine
Council Benchmark-H (TPC-H) [13] benchmark                   cluster. However, taking flexible management of

Volume 1, Issue 4 November - December 2012                                                                   Page 26
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 1, Issue 4, November – December 2012                                    ISSN 2278-6856


infrastructure into account, using a virtual machine is       RTT. The result is 21 seconds when RTT is from 0msec
considered to have greater benefits. In this graph,           to 4msec, and 52 seconds when RTT is 16msec that is the
performance of DomainU is better than that of Domain0.        longest RTT measured in this study. Figure 12 shows
This is considered to be the result of optimization of        total amount of time that sums up migration time to
DomainU in the mechanism of Xen [14] [15].                    migrate a virtual machine to a remote site and execution
                                                              time of OSDL-DBT3 at the remote site. Figure 12
             Table 1: Experimental Setup: PCS                 includes

 OS             initiator : Linux 2.6.18-
                53.1.14.el5(CentOS5.3)
 CPU            initiator : Intel (R) Xeon(TM)
                3.6GHz
                target : Intel (R) Xeon(TM)
                3.6GHz
 Main           initiator(Domain0) : 2GB
 Memory
 iSCSI          initiator : iSCSI-initiator-utils             Figure 8 Comparison of real machine cluster and virtual
 Monitorin      Ganglia                                                         machine cluster
 g Tool




                                                                Figure 9 Experimental Environment before migration



 Figure 7 Basic experiment: Experimental Environment


  4.2 Experiment including virtual machine migration
The results of previous research have shown that parallel
data processing applications that is not I/O-bound like
HPA[11] can be executed with sufficiently practical
performance even when the data is stored at a remote site.
In contrast, in the case of I/O intensive applications like
OSDL-DBT3,         we    have     confirmed     remarkable
performance decline when data exists on the site through       Figure 10 Experimental environment of after migration
high-latency network. Thus in this study, we propose a
technique to migrate virtual machine to a cloud               DomainU execution time of existing research, in which
environment over high latency networks site that stores       data at the remote site is accessed directory through
data, in order to achieve load balancing and optimization     WAN, for comparison. From this figure, the execution of
of storage access instead of iSCSI remote access[13]. We      this experiment is faster than that of existing research
have built a virtual machine PC cluster as shown in           when RTT is long. We can execute an application without
Figure 9 and 10 that contains six servers (Initiator) and     delay on iSCSI by means of migration of virtual machines
two iSCSI storage nodes (Target). Four servers and a          to a remote site. Therefore, as RTT becomes longer, it is
storage node are located in the local site, two servers and   effective to migrate a server to a remote site where storage
a storage node are located in the remote site. Assuming       is located.
remote iSCSI access, we have inserted 1msec, 2msec,
4msec, 8msec, 16msec RTT by Dummy net between the
local site and the remote site as the same with existing
research. First, Figure 11 shows migration time of a
virtual machine from a local site to a remote site on each

Volume 1, Issue 4 November - December 2012                                                                      Page 27
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 1, Issue 4, November – December 2012                                    ISSN 2278-6856


                                                               between virtual machines on the local site cluster for
                                                               comparison.




                Figure 11 Migration time                                 Figure 13 EC2 - Network Bandwidth

                                                               In network throughput, there is a difference in
                                                               performance about 4.5 times in local cluster and Amazon
                                                               EC2. We have examined network throughput between
                                                               numbers of virtual machines on Amazon EC2, and
                                                               observed that the performance of all cases is almost the
                                                               same. This indicates that all the virtual machine should
                                                               be deployed on the same real (physical) machine in this
                                                               case.
                                                                   5.3 Performance of disk access
                                                               Next, we have measured performance of disk access on
                                                               Amazon EC2. Figure 14 shows performance of Write and
                Figure 12 Execution time                       Figure 15 shows performance of Read access to disk.

5. BASIC PERFORMANCE OF AMAZON EC2
     5.1. An overview of experiments
We aim to realize the load balancing between local own
cluster and remote real cloud dynamically, estimating
application execution performance by means of network
throughput and performance of disk access in data
intensive application. Performance evaluation of each            Figure 14 EC2 - Performance of Disk access (Write)
element should be essential for performance prediction of
data intensive application. First we have measured             These graphs include performance of local disk access
throughput of network and iSCSI disk access of Amazon          and iSCSI disk access on EC2. Performance of iSCSI disk
EC2 because we are using iSCSI for storage access. We          access is significantly lower than that of local disk access.
have used Open-iSCSI [16] for iSCSI Initiator and SCSI-        This must come from the low throughput of network on
target-utils for iSCSI Target. For the installation of Open-   EC2, which is shown in Figure 13. According to this
iSCSI, it is common to compile a kernel of OS at first and     result, we should take network performance inside cloud
then compile Open iSCSI in order to make a kernel              into account,
module of Open-iSCSI. However, in Amazon EC2, users
can not boot up a kernel that users have compiled by
themselves. Thus we have prepared Amazon Machine
Image (AMI) which includes gcc version 4.0, which is
suitable for compiling both the kernel and Open-iSCSI,
and created a kernel image in which kernel modules can
be installed. Using the kernel image, we have deployed
AMI and installed iSCSI.


    5.2 Network Bandwidth inside cloud
Figure 13 shows network throughput between virtual
machines on Amazon EC2. It includes throughput
                                                                  Figure 15 EC2 - Performance of Disk access (Read)

Volume 1, Issue 4 November - December 2012                                                                        Page 28
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 1, Issue 4, November – December 2012                                    ISSN 2278-6856


including iSCSI disk access, which we outsource a job to         [9] Shiori Toyoshima, Saneyasu Yamaguchi, Masato
a cloud computing system.                                          Oguchi: “Analyzing Performance of Storage Access
                                                                   Optimization with Virtual Machine Migration,”
6. CONCLUSIONS                                                     CPSY,Vol.109,     No.296,    CPSY2009-37,pp.13-
                                                                   18,Kyoto,Nobember 2009.
In existing research, we have experimented server access
                                                                 [10] Ganglia monitoring System:http://www. Ganglia
to remote storage directly during runtime. We have
                                                                   .info/
confirmed as RTT is longer the execution time becomes
                                                                 [11]            Luigi    Rizzo:http://info.iet.unipi.it/
longer in iSCSI remote access. Since remote access is
                                                                   luigi/dummynet
bottleneck of execution of the application, we have
migrated a virtual machine on a local site to a remote site,
                                                                 [12]   OSDL-DBT3:http://ldn.linuxfoundation.org/.
and OSDL-DBT3 is executed with storage access in the
                                                                 [13]   TPC-H:http://www.tpc.org/tpch/.
remote site. As a result, total amount of migration and
                                                                 [14] Aravind Menon, Alan L.Cox, Willy Zwaenepoel:
execution time is shorter than that obtained in existing
                                                                   “Optimizing Network Virtualization in Xen,” 2006
research as RTT is longer. We have confirmed that our
                                                                   USENIX Annual Technical Conference, pp.15-28,
method is effective to relocate a local server to a remote
                                                                   May 2006.
site when it takes longer to access to a remote storage.
                                                                 [15] Jose Renate Santos, Yoshio Turner.G(John)
Moreover, we have introduced iSCSI on Amazon EC2,
                                                                   Janakiraman,Ian Pratt “Bridging the Gap between
measured throughput of network and iSCSI disk access
                                                                   Software and Hardware Techniques for I/O
for estimating execution performance of applications on
                                                                   Virtualization” 2008 USENIX Annual Technical
real cloud. As a future work, we will build a system of
                                                                   Conference, pp.29-42, June 2008.
load balancing dynamically if the load is heavy which
                                                                 [16]   Open-iSCSI: http://www.open-iscsi.org/
will migrate virtual machine to a remote site
                                                                 [17]   Tanimura Yusuke, Ogawa Hirotaka, Nakada
automatically, and analyze its behavior.
                                                                   Hidemoto, Tanaka Yoshio, Sekiguchi Satoshi:
                                                                   “Comparison of Methods for Providing an IP Storage
REFERENCES                                                         to A Virtual Cluster System,” IPSJ SIG Notes 2007,
  [1]O’Riordan, Dan. Cloud Computing Breaking                      pp.109-114, March 2007.
    through the haze. s.l.: IBM Corporation 2009.
  [2] Lijun Mei, W.K. Chan, T.H. Tse, “A Tale of
    Clouds: Paradigm Comparisons and some                      AUTHOR
    Thoughts on Research issues”, 2008 IEEE       Asia-
    Pacific Services Computing Conference.
  [3] Hong Cai, Ning Wang, Ming Jun Zhou, “A                                            1. Dr. D.S. Chauhan did his B.Sc
    transparent Approach of Enabling SaaS Multi-                                        Engineering(1972)       in     Electrical
                                                                                        Engineering at I.T. B.H.U., M.E.
    tenancy in the Cloud ’’, 2010 IEEE 6th World
                                                                                        (1978) at R.E.C. Tiruchirapalli (
    Congress on Services.                                                               Madras University ) and PH.D. (1986)
  [4] Chang Jie Guo, Wei Sun, Ying Huang, Zhi Hu                                        at IIT /Delhi . Under his supervision 24
    Wang, Bo Gao , “A Framework for Native Multi-                                       candidates has completed their PHD.
    Tenancy       Application      Development     and                                  He is an active Member of IEEE, ACM,
    Management’’ The 9th IEEE International                    SIAM. He has been nominated by UGC as chairman of
    Conference on E-Commerce Technology and The 4th            Advisory committees of four medical universities. Dr. Chauhan
    IEEE International Conference on Enterprise                got best Engineer honour of institution of Engineer in 2001 at
    Computing, E-Commerce and E-Services(CEC-EEE               Lucknow .He is currently serving as Vice Chancellor in
                                                                                      Uttrakhand technical University.
    2007) 0-7695-2913-5/07 .
                                                                                      2. Manjeet Gupta has received B.Tech
  [5]    Z. Pervez, Sung young Lee, Young-Koo Lee,
                                                                                      and M.Tech degree in CSE in 2005 and
    “Multi-Tenant, Secure, Load Disseminated SaaS
                                                                                      2008 respectively .Now she is pursuing
    Architecture’’,12th International Conference on
                                                                                      her Ph.D in Cloud Computing. She has
    Advanced Communication Technology 2010                                            worked with Jaypee university of
  [6] J Brass, “Physical Layer Network Isolation in                                   Information Technology for 1 Year as
    Multi-tenant Clouds”, International Conference on                                 Assistant Professor in the department of
    Distributed Computing Systems Workshops 2010.              computer science and engineering. Currently she is working as
  [7] Pankaj Goyal, “Computing Policy-based Event-             Assistant Professor in CSE department at JMIT, Radaur.
    driven Services-oriented Architecture for Cloud
    Services Operation & Management” presented at
    2009 IEEE International Conference on Cloud.
  [8] Guoling Liu. “Research on Independent SaaS
    Platform”,2nd IEEE conference on Information
    Management and Engineering,2010.

Volume 1, Issue 4 November - December 2012                                                                            Page 29

				
DOCUMENT INFO
Description: International Journal of Emerging Trends & Technology in Computer Science (IJETTCS) Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 1, Issue 4, November – December 2012, ISSN 2278-6856, Impact Factor of IJETTCS for year 2012: 2.524