PERFORMANCE ANALYSIS OF CLOUD AS A FEASIBLE

PLATFORM FOR HPC APPLICATIONS USING RADIX

SORT ALGORI THM

Md. Aiatullah, Rashid Hassani, Abbas Malekpour, Peter Luksch

Faculty of Computer Science and Electrical Engineering, University ofRostock, Rostock, Germany
md. [email protected], rashid. [email protected], [email protected],
[email protected]

Keywords: Cloud Computing, Parallel Radix Sort, Open application is running on the Cloud [ I ]. The software re­
MPI, EC2. sources are accessed through a thin client over the Internet;
normally its functions are performed remotely as a Web­
Abstract based. Platform as a Service (PaaS) is a way to rent hard­
ware, operating systems, storage and network capacity
Cloud Computing has become a significant technology over the internet. It also refers to deploy the applications
trend. It is going to reshape IT processes and IT market­ created by the development languages and tools suchas
place in recent years. It is grabbing the people's attention PHP, Ruby, Java, Python and .NET provided by service
providers to the Cloud infrastructure. There are many
rapidly as an opportune resource of computational power
commercial providers such as Amazon, IBM, Microsoft
as well as priced Cloud computing services save pains to
and Heroku who offer Cloud environments for developing
maintain the computational ambience. Gradually, Cloud is
and deploying applications. Infrastructure as a Service
bringing new possibilities for a scientific community to
(IaaS) is a provision model in which computing resources
build High Performance Computing platforms. Despite
provided to users on lease such as storage, processing
the wide benefits the Cloud offers, the question on every­
power, network including operating system. The client
one's mind is "Whether is Cloud a feasible platform for
typically pays on a per-use basis.
HPC applications?" This paper evaluates the performance
The Cloud deployment models are public, private,
of Amazon Cloud as well as non-Cloud using Parallel
community and hybrid Cloud. A public Cloud is hardware
Radix Sort algorithm based on Open MPI. Message Pass­
and software infrastructures perform by third parties such
ing Interface (MPI) is chosen as the efficient method to
as Google, Facebook and Amazon. Whereas a private
develop and implement the Parallel Radix Sort in the
Cloud refers to infrastructure run and preserved by a sin­
Cloud platforms. Most preceding works on High Perform­
gle organization. A community Cloud is a multi-tenant
ance Computing have been ported to the public Cloud
infrastructure that is shared among several organizations
background in various fields. The problems and chal­
from a specific group with common computing concerns.
lenges are assessed related to HPC on local and Cloud
A hybrid Cloud is a combination of both public and pri­
environments. Based on these experiment results, this vate Cloud and offers profits that the data and programs
paper will conclude the Cloud is a suitable platform for can be moved from one to another Cloud model.
proposed type of HPC applications. Most supercomputers are actually multiple computers
that perform parallel processing and near the currently
1 Introduction highest operational rate for computers. Supercomputing,
often called as High Performance Computing (HPC) is
In recent years, Cloud computing has developed rapidly as referred to a characteristic that utilizes parallel processing
a new illustration in IT industry. Recent progresses of to perform huge computations in short time. Scientists,
engineering have cut down costs of computers and net­ Business, Engineers and others require a large number of
work, and these alterations gave a huge impact on high computational powers to solve the problems on the day
performance computing environment. Cloud computing is after day basis. HPC application problems exist in various
a virtualized ambience that provides on-demand provi­ disciplines like including quantum mechanics, weather
sioning of storage and compute resources. The key bene­ forecasting, oil and gas exploration, molecular modeling
fits offered by Cloud computing compared to traditional and physical simulations etc. Usually, they being solved
IT resources are nearly infmite compute resources availa­ using massive hardware and software computing system
ble on-demand that can be deployed rapidly and scaled up known as Supercomputer or HPC clusters. HPC provides
elastically. In Cloud computing, using zero upfront in­ the capability to conduct massive simulation and analysis
vestment and pay-by-usage model reduces the costs. computations. However, Supercomputers are not a feasi­
There are three types of Cloud computing services i.e., ble solution to purchase for the normal users and small
(IaaS) Infrastructure-as-a-Service, (PaaS) Platform-as-a­ businesses intending to run their applications for a short
Service and (SaaS) Software-as-a-Service. SaaS is a soft­ period, as they are not designed for on-demand access and
ware service that provides to the user basically when an the users in many cases cannot distribute and coordinate a

577

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on October 03,2024 at 08:06:22 UTC from IEEE Xplore. Restrictions apply.
 large scale job on different machines. [6].
The Cloud is a cost effective explanation to supercom­ • Network Interconnect problem: there are also makes
puters to solve the problems by creating a cluster using the Amazon EC2 slower than the traditional Linux and
laaS, scaling up the number of virtual servers dynamically High Performance Computing cluster, limiting the per­
when needed rely on the size of the task and relinquish formance and causing variability issues [7].
when the job is performed. This is one of the main reasons • Virtualization: The Cloud Computing technology can
motivating many users and organizations to port super­
have the most difficult performance issue, which can
computing applications to the Cloud. Conventional HPC
have a great impression on HPC applications [8].
or supercomputing platforms present limited hardware
• Network overhead problem: This is a serious topic in a
access and fail to scale. Therefore, this motivates to carry
out a detail study on HPC in Cloud. virtual cluster of High Performance Computing Cloud.
Open MPI is an open source software or open source Therefore, it can be resolved by using lightweight
high performance computing implementation of the Mes­ communication protocol instead of TCP/IP in inter­
sage Passing Interface (MPI) standard. MPI defines an domain virtual instances of Cloud [9].
API that is used for a specific type of portable, high­ Based on the previous mentioned Cloud problems we
performance inter-process communication (IPC): message have to focus on the following questions:
passing. Specifically, the MPI document describes the • What has previously done on Supercomputing in the
reliable transfer of discrete, typed messages between MPI Cloud environments?
processes. • What are the various HPC applications that have been
Sorting algorithm is one of the best classic applications deployed in the Cloud?
of supercomputing. Modern supercomputers have ad­ • What are the problems and challenges faced by HPC
vanced rapidly in size and changed significantly in archi­
applications in the Cloud?
tecture, forcing new adoptions to sorting algorithms. In
• How can the Radix-Sort algorithm be implemented in
general of view, many improvements have been done on
order to obtain good speed up in a Cloud ambience?
sorting with increasing changes in multiprocessor archi­
tectures. There are many types of distributed sorting algo­
• What is the speedup of Parallel Radix-Sort in Cloud
rithms such as Counting sort, Bucket sort, Radix sort, environment compare to other parallel environments?
Quick sort, Heap sort and Merge sort. Radix sort is a well­
known fast algorithm having a time complexity of 0 (kN) 3 State of the Art
on a very good average and 0 (k+N) on worst-case per­
formance. The fundamental principle of Radix sort is that More than a few research jobs had been carried out to
it performs a random number of elements called as a list examine the performance of HPC applications deployed
significant digit and most significant digit. on Cloud. In [10] study has been performed HPC Bench­
Therefore, the reason for selecting Radix sort is that it mark on the Amazon EC2.The author has applied HPC
performs faster for longest integer datasets as compared to applications physical nodes as well as virtual machines in
other algorithms like Bucket sort, Merge sort, Quick sort
the EC2 Cloud and evaluated the performance. The out­
and so on. So, this knowledge principally motivated to
come states that EC2 faster than physical nodes.
develop a Parallel Radix Sort algorithm and perform as a
In [11] study performed parallel Quicksort on Symmet­
purpose of an open source high performance computing to
ric multi processors. The implementation method used for
test on the public Cloud platform, local platform and eva­
Quicksort is MPICH. Master-slave model has been fol­
luated performance.
lowed for division of work. The outcomes show good
speeds for Quicksort. But, MPICH I is not so efficient for
2 Problem Statement
handling data in a network capacity of I Gbps. Therefore,
in order to obtain the faster communication and better
• An important motivating test for this work is under­ traffic management MPICH2 is needed.
standing data at the big scale, where it is anticipated
In [12], the research evaluates the CPU performance
that the amount of data produced by users will far ex­
while running an HPC application on EC2 with some
ceed capability of the systems. High-performance com­
HPC benchmark tests. The writer has suggested that a new
puting is the computer systems that are able of proc­
phenomenon is required to evaluate the performance.
essing tasks of high performance computing. High Per­
In [13], the researchers have considered three different
formance Computing is faced problems in the Cloud
such as: parallel sorting algorithms and have discussed the problem
• Scaling Limitations: which different tests have been of sorting with minimal data. The tests were performed
performed on many Clouds namely EC2 [4] [3]. using MPI on Blue Gene Supercomputer and the out­
• Establishing connection among instances in a Cloud comes shows excellent scalability without any additional
cluster: this is a significant problem faced by parallel memory.
scientific applications, which results are actually in
more delay [5]. 4 Solution Approach & Method
• Greater latency and lower bandwidth: the performance
degradation in the Cloud is observed due to greater la­ The goal of this paper work is to evaluate performance
tency and lower bandwidth when measured up to whether the Cloud is feasible platforms for establish High
against dedicated Supercomputers or cluster systems Performance Computing applications and also to provide

578

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on October 03,2024 at 08:06:22 UTC from IEEE Xplore. Restrictions apply.
 the complete overview of HPC in the Cloud. The main mance computing. The entire libraries are compiled with
intentions of these researches are: GNU compiler. Cloud and non-Cloud has been configured
• To fmd the preceding jobs of High Performance Com­ with MPI and tested.
puting in a Cloud ambience. The Parallel Radix Sort is compiled and executed on the
• To identify the problems and challenges faced by HPC non-Cloud node (A physical node and our different local
in a Cloud ambience. MPI server Clusters). The execution times are tabulated
• A study is paddled out to select a Cloud well-matched for different number of input sizes. Secondly the sort also
parallel programming technique. After then selecting compiled and executed on the Master node of the Amazon
the suitable technique, a Parallel RadixSort is devel­ EC2 Cloud for different number of input sizes and similar­
oped and implemented on the public Amazon EC2 ly the Execution times are also tabulated. Execution time
Cloud and non-Cloud. or sorting time is calculated using "timeJ' function.
• To evaluate performance analysis found out using Par­ Table 1 shows the configuration of software that we

allel Radix Sort in the Cloud and non-Cloud. used in the experiments. Table 2 and Table 3 summaries
In this research, predominantly, OpenMPI, MPI version the hardware configuration of physical nodes and our Si­
Parallel Radix Sort used and test the performance over rius server. In the experiments purpose of this paper, we
different platforms. This idea is mainly motivated to cho­ used large instances of Amazon EC2 with a public image
sen Massage Passing Interface (MPI) as the efficient for all the experiments. Ubuntu server (12.10) is installed
method to develop and implement the Parallel Radix Sort on to the public image for cluster instance.
in the local system and Cloud environment. Several re­ Since, as we have seen different number of operations
searchers have been implemented parallel sorting algo­ that execution times are not stable. When the server is not
rithm on different areas but as we seen none of them im­ busy as a compare the system performances are almost
plement Parallel Radix Sort in the cloud environment, stable. So mainly it depends on servers busy that doneby
which already mentioned above in previous section. We several test. Table 4 provides system information of the
are come to implement Parallel Radix Sort in the local regarding EC2 Instance types and Table 5 provides the
system and Cloud environment. system information an instance of Amazon Elastic Com­
puting Cloud.
Then in the Figure 1 Parallel Radix Sort run on local
5 Amazon Ee2
system and by taking different numbers of integer data,
the first input size (25000) and execution time is 0.8 sec.
Amazon EC2 offers a collection of services that allows
The algorithm-parallelized execution time is done almost
business subscribers to run application programs in the
sequentially. If the number of input size increases, the
Amazon Cloud computing environment. The EC2 can
perform as a practically unlimited set of virtual machines.
Table 1. Library/Complier/Osinformation.
It is designed to make web-scale computing easier for
developers. The most important reasons for choosing Library/Compiler/OS Bold
Amazon EC2 is Quick and reliable Cloud servers, easy
management, low cost, free tier, Node.js API, scalability, Linux kernel 2.6.21.7-2.fc8xe n-ec2-v1.0

auto-scaling, market leader and it provides both OpenMPI GNU compiler 4.1.2
and mpich RPMs in their repository.
Furthermore to use the EC2, a customer creates an Sirius 2.0 .5
Amazon Machine Image (AMI) containing the operating Open MPI Vl.6.4
system, application programs and configuration settings.
After that the AMI is uploaded to the Amazon Simple
Table 2. Hardware specification of physical node.
Storage Service (Amazon S3) and registered with Amazon
EC2, generating a so-called AMI identifier (AMI ID). Platform Ubuntu
Once this has been done, the consumer can requisition
virtual machines on an as-needed basis. This introduce a Processor Intel (R) Core (TM) i3 CPU
dynamic size which can be increased or decreased in real I.3GHz
Cost Speed
time from as few as one to more than 1000 virtual ma­
chines simultaneously. With a default plan, a consumer Ram 4 GB
can run up to 20 instances at a time. Each instance varies
Type system 32 bit
depending on the cost takes place according to the com­
puting and network resources consumed.
Table 3. Hardware specification of sirius cluster.
6 Performance Evaluation
3 Nodes 8 Nodes

To evaluate the performance of Cloud and non-Cloud en­ 4CPUs Opteron 842 @ 2.0 GHz 4CPU Opteron 8350
vironmentsusing MPI based Parallel Radix Sort. As a high 51 GB disk space 160GB SATA
performance message passing library, we used Open MPI 4 GB DDR-RAM ECC 160GB DDR2-RAM
[2] which are implementation open source high perfor-

579

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on October 03,2024 at 08:06:22 UTC from IEEE Xplore. Restrictions apply.
 Table 4. CC2.8xlarge. 0.5
.gj
Cluster Compute Quadruple Extra Large g
g 04
.
"2o
60 .5 GiB of memory 0.3
88EC2 Compute Units (2 x Intel Xeon X5570 @ 2.93GHz, .�
"- _VHRSirius
quad-core) � 0.2 Cluster/Seconds
o
3370 GB of instance storage �

0.1
"

64-bit platform E
I/O Performance: Very High (10 Gigabit Ethernet). "
;z:
EBS-Optimized Available: No** o
API name: cc2.8xlarge Number of Processes(25000-34000)

Figure 2. Variation of measured performance of Sirius server

Table 5. M3.xlarge. according to number of blocks.

15 GiB memory
13 EC2 Compute Units (4 virtual cores with 3.25 EC2 Compute Units
each)
EBS storage onIy
64-bit platform
1/0 Performance: Moderate -Amazon

EBS-Optimized Available: 50 0 Mbps EC2/Seconds

API name: m3.xlarge

2
Number of Processes(25000-34000)

.gj
c::
Figure 3. Variation of measured performance according number
g 1.5 of different input size deployed on ec2 Cloud.
"2o
.�� 0. 1 5
1
00
'"
_Physical
�
o
oeds/Seconds
"
8 0.1
�
0.5
•
�c:: g •
.� 0.05 •
Z
:3 •
;z: • -
- -
• EC2 instance
-
Q. • - -
o � •
- - -
- - -
•
0
0 •
Number of Processes(25000·34000) '-0
"
�

Figure 1. Physical machine performance of different number of

E
Z Input Sizes on Different Cores
input sizes.
Figure 4. Performance comparison of different core on Amazon
computation execution time is also increased. We ob­ ec2 Cloud.

served that physical node computation execution time is

2
not uniquely stable. It varies on the busy of thenetwork
l-
00

§
'0

systems. � I-
� 1.5
Similarly, the Figure 2, Figure 3 run sort program and �
I-
the computation execution time are performed in the same
;:;
.= I==='"
I-
1=
::: 1 • EC2 Cluster
..,
-I- r-- l-
go
--

way. Therefore in the high performance computing, Sirius • Physical IVlachine

'Cl .
I-f- r-- l-
cluster server is better as a compared than physical nodes. 0.5
-

..,
I-l-I- Ir- t-c • VHR ClllsterSirius
The performance of computation execution time is done
.r>
'"

z
:;;
relatively parallel in the Amazon ec2 Cloud. Because the 0
Number of Processes
different number of operation per seconds are computed
nearly sequential. In the general view of observation, the Figure 5. Performance comparisons of a physical machine, Sir­
Amazon EC2 Cloud computation execution time is faster ius cluster and an Amazon EC2 instance. Plots are relative dif­
ferent number of operations per second with a base of the regard­
than Sirius cluster and physical node.
ing measurements of a physical machine and Sirius server.
Now here, we have compared an Amazon EC2 instance
by taking with different cores that are figured in the Fig­ 2
.....
ure 4. Soin Figure 4 the same number of input sizes, that 0 �
:;; ] I-
00

1.5
.g u
0
if the cores are increased execution time taken less as a � f-- = I-
"
:;; �
compared previous core. Likewise Figure 5 performances c:: 1

are comparison between a physical machine, our local .� .� il 0.5 1-1-1· 1- • Physical

go �L:' l- I- � t- I--
--
-.;
sirius cluster and an instance of Amazon EC2, utilizing ..::
0 I- t=- .Ee2

open source high performance Open MPI. The figure is Inl'ut Sizes

clearly shows physical machine and our local sirius cluster Figure 6. Performance comparisons between an Amazon EC2
slower than EC2 Cloud. instance, and a physical machine. Blue's are the physical and
Furthermore in the above Figure 6 is also compared the red's are the Amazon ec2 number of operation per seconds .Red
performance of an Amazon EC2 instance and a physical is the better performance.

580

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on October 03,2024 at 08:06:22 UTC from IEEE Xplore. Restrictions apply.
 machine. Figure 6, Physical nodes are 70% slower than both Cloud, non-Cloud as well as at the end summarized
Amazon EC2 nodes. Therefore Amazon EC2 is still better the whole observations of this paper, physical node is
performance for computing. At the end also compared in slower than Sirius Cluster. Similarly as a compared other
the Figure 7, which indicates, an Amazon EC2 instance, a platforms Amazon Ec2 is double faster than a physical
physical machine and VHR Sirius Cluster. This Figure 7 node. Furthermore this paper proved that in Figure 7
shows that blue bars are relative numbers of operations on Amazon EC2 Cloud nods are around 50% faster than our
average, red bars are the worst relative numbers of opera­ local our Sirius Clusters, which is based on a very fast
tions, and pink bars are the best relative numbers of opera­ infiniband local network. Our data clearly shows a strong
tions. So we have observed that min operations are the relationship between times an application spends on
relatively best performance, as a compared Amazon EC2 communication, and its overall perfonnance on EC2.
node. Still Amazon EC2 Cloud performance is better than This paper is concluded that Amazon Elastic Comput­
our local platform. Obviously this paper is compared all ing Cloud is more suitable platform for exposed HPC ap­
performance results depending on hardware configuration. plications.
That is, this paper can also supposition the performance
measured in this experiment is not almost stable whereas References
our Sirius cluster performance measured experiment is
almost stable and also physical nodes performance meas­ Peng, J., Zhang, x., Lei, Z., Zhang, B., Zhang, W.
[1]
ured result is nearly stable but not accurately. Therefore and Li, Q. (2009) Comparison of Several Cloud
this paper cannot provide a clear explanation for this phe­
Computing Platforms. 2009 Second International Sym­
nomenon immediately, and we need further investigation
posium on Information Science and Engineering
why it is not stable in the Amazon Cloud platform. As an
(lSISE), 23-27.
Amazon EC2 instances are virtual machines, one suspi­
[2] (1996) Open MPI.
cious point is a functionality of time measurement. The
http://www.open-MPI.orgFLEXChip Signal Proces­
measurement is carried out based on CPU time not wall­
sor (MC68175ID), Motorola.
clock time.
[3] Rashid, L., Hassanein, W. and Hammad, M. Analyz­
ing and Enhancing the Parallel Sort Operation on
7 Conclusions Multithreaded Architectures. The Journal of Super­
computing, 53, 293-312.
Improving efficiency for HPC is always centred to opti­ http://dx.doi.orgll0.10071s11227-009-0294-5
mization. Many research works have been done to de­ [4] Ghoshal, D., Canon, R.S. and Ramakrishnan, L. (20[[)
crease the execution time or to improve the throughput. 1/0 Performance of Virtualized Cloud Environments.
The efficient use of machines in the Cloud can signifi­ Proceedings of the Second International Workshop
cantly reduce the cost. In this work, we evaluated the per­ on Data Intensive Computing in the Clouds, New
formance of Cloud and non-Cloud services through Paral­ York, 71-80.
lel Radix Sort algorithm based on Open MPI, which is http://dx.doi.orgIl0.1145/2087522.2087535
intended for high performance computing. Whereas this He, Q., Zhou, S., Kobler, 8., Duffy, D. and Mc­
[5]
paper accessed the feasibility of the public Cloud Amazon
Glynn, T. (2010) Case Study for Running HPC Ap­
EC2, VHR Sirius cluster as well as physical nodes usages
plications in Public Clouds. Proceedings of the 19th
for running tightly coupled of MP[ programs. [n general,
ACM International Symposium on High Performance
our study expresses a picture considerably more positive
Distributed Computing, New York, 395-40 I .
for running MP[ applications on public Clouds and non­
http://dx.doi.orgIl0.11451l851476.1851535
Cloud, as a comparison to the past published evaluation
[6] Rehr, J.J., Vila, F.D., Gardner, J.P., Svec, L. and
results. In the experiments, compared the performance of
Prangepa, M. (20[0) Scientific Computing in the
an Amazon EC2 instance, Sirius cluster and a physical
Cloud. Computing in Science Engineering, 12, 34-43.
machine with Open MP[ v I .6.4 and tabulated the per­
http://dx.doi.org/ I O.[ 109/MCSE.201 0.70
formance results. [n the same time performed several tests
[7] Chiba, T., den Burger, M., Kielmann, T. and Matsu­
�
1.2 oka, S. (2010) Dynamic Load-Balanced Multicast
§
for Data-[ntensive Applications on Clouds. 20[0
1
"0

il lOth IEEEIACM 45 International Coriference on Clus­

0.8
.!(f
!!l
0 ter, Cloud and Grid Computing (CCGrid), 5-14.
..,
3 0.6 • average ops
[8] Jackson, K.R., Ramakrishnan, L., Muriki, K., Canon,
g.
""'
0
0.4 -max ops
S., Cholia, S., Shalf, J., Wasserman, H.J. and Wright,
0.2 N.J. (20[0) Perfonnance Analysis of High Perfor­
�
" -.nulops
=
.0
"
Z 0
mance Computing Applications on the Amazon Web
Sirius �c physical ee2 �e
Services Cloud. 2010 IEEE Second International Con-
4c ference on Cloud Computing Technology and Sci­
Figure 7. The average performance, the best performance, and ence (CloudCom), 159-168.
the worst performance of EC2, physical and Sirius server with 4 [9] Akioka, S. and Muraoka, Y. IEEE HPC Benchmarks
cores respectively. Min operation is better. on Amazon EC2.

581

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on October 03,2024 at 08:06:22 UTC from IEEE Xplore. Restrictions apply.
 [10] Gupta, A. and Milojicic, D. (2011) Evaluation of raIlel Implementation of Quicksort and Its Perfor­
HPC Applications on Cloud. Open Cirrus Summit mance Evaluation on SUN Enterprise 10000. Pro­
(OCS), 22-26. ceedings of the Eleventh Euromicro Conference on
[ I I ] Berriman, G.B., Juve, G., Deelman, E., Regelson, M. Parallel, Distributed and Network-Based Processing,
and Plavchan, P. (2010) The Application of Cloud 372-381.
Computing to Astronomy: A Study of Cost and Per­ [13] Tham, S. and Morris, J. (2003) Cilkvs MPI: Compar­
formance. 20 I 0 Sixth IEEE International Confe­ ing Two Very Different Parallel Programming Styles.
rence on e-Science Workshops, 1-7. Proceedings of the 2003 International Conference
[12] Tsigas, P. and Zhang, Y. (2003) A Simple, Fast Pa- on Parallel Processing, 143-152.

582

Authorized licensed use limited to: Nirma University Institute of Technology. Downloaded on October 03,2024 at 08:06:22 UTC from IEEE Xplore. Restrictions apply.