Professional Documents
Culture Documents
Abstract Cluster computing involves connection between a a TCP/IP socket communication, the send and receive buffer
server and multiple hosts in one single environment and multi- sizes define the receive window that specifies the amount of
cluster in inter-cluster environments. One of the crucial factors data that can be sent and not received before the send is
of a clustering is its network communication. Relying for support interrupted. If too much data is sent, it overruns the buffer and
of a server-client internetworking, cluster computing uses the
interrupts the transfer. The mechanism that controls data
TCP/IP socket communication mechanism extensively.
Mechanism that controls the data transfer interruptions between transfer interruptions is referred to as flow control. If the
a server and client is referred to as a flow control. Theoretically, receive window size for TCP/IP buffers is too small, the
if the receive window size for TCP/IP buffers is too small, the receive window buffer is frequently overrun, and the flow
receive window buffer is frequently overrun, and the flow control control mechanism stops the data transfer until the received
mechanism stops the data transfer until the receive buffer is buffer is empty.
empty. Thus, as a result flow control activity contributes to This flow control activities can consume a significant
unnecessary additional network latency which is unacceptable in amount of CPU time and result in additional network latency
any server-client networking especially in clustering environment as a result of data transfer interruptions. Increasing buffer size
where it involves highly reliable data transfer. Among the
is one of the options available to avoid flow control under
solutions to this problem is by adjusting buffer size to avoid or
reduce the potential for flow control to occur. This paper normal operating conditions. A larger buffer size reduces the
describes the analysis of TCP/IP socket buffer length in Local potential for flow control to occur. However, larger buffer size
Area Network (LAN) and Wide Area Network (WAN) clustering. can also have negative impact to the network performance
It details the process and flow of setting up the environment, such as the increasing of paging if the applications failed to
implementing as well as analyzing the throughput performance process fast data. Optimal buffer size dependable on several
of the measured buffering. network environment factors such as network topology,
memory size and data transfer size, among others. Specifying
Index Terms Throughput Analysis, Network Performance, an adequate value of buffer length for socket connection is
Cluster Computing, TCP/IP Buffer Length.
crucial in avoiding flow control, but not so large to be
processed by the computer system.
I. INTRODUCTION Having an intolerant cluster connection to data
communication failure, the network performance is seen to be
Cluster computing is a collection of several interconnected an important criterion to be measured in this research. This
computers that cooperatively work together as single research is focusing on the analysis of socket buffer size as
resources processing in parallel. Jobs of the clusters are being throughput factor monitored in order to analyze the network
scheduled according to the runtime and also the resource of performance. The experiment involves a Local Area Network
the interconnected machines. Nature of cluster interconnection (LAN) clusters as well as a Wide Area Network (WAN)
involves communication between server and client that are clusters network performance analysis. Specification of the
highly dependable on the network facilities offered by the cluster has to be considered in the performance analysis. JPerf
environment. Guarantee of data transfer between a server and performance analysis tool has been used to generate and
cluster nodes further moved to inter-cluster connections are measure the performance of the transmission control protocol
vitally important and not tolerant to any data transfer (TCP) in this network. JPerf is a Java-based graphic user
interruption. Basic cluster communication implies the network interface (GUI) tools that used to perform the function of
protocol of TCP/IP. In TCP/IP, socket communications measuring network performance. The data manipulated
mechanism is extensively implemented this this operation. For between Faculty of Electrical Engineering (FKE) and Faculty
of Applied Science (FSG) of Universiti Teknologi MARA is
categorized under Local Area Network interconnection which III. NETWORK ARCHITECTURE
also known as intra-cluster while the inter-cluster connected is
on different geographical area between FKE and Faculty of The cluster network architecture that had been analysed is
Pharmacy (FF) of Universiti Teknologi MARA which shown in Figure 1. The Linux distribution that used for this
categorized as inter-cluster connection under Wide Area cluster network is Centos 6.3 which it is a modern version of
Network. the Rock Cluster distribution. Centos provide a free
This paper is organized in the following way. It starts with computing platform which consists of binary compatibility
brief related work in part two, followed by part three which with its software development source.
describes full implementation of the system which contains
setting up the system requirement and configuration, network
architecture and defining the system flow. Network
performance analysis is being discussed in part four. The
paper ends with result and discussion, and conclusion.
622
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
Source Command
Code
Iperf 2.0.5 #wget Fig. 4. Iperf Command for FSG Server Setup.
http://sourceforge.net/projects/iperf/files/ip
erf-2.0.5.tar.gz/download
#tar -xzvf iperf-2.0.5.tar.gz
# cd iperf-2.0.5
# ./configure
# make
#make install
623
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
IV. NETWORK PERFORMANCE ANALYSIS factors that affected the performance of the network. The
distance between FKE, FSG and FF is seen to significantly
There are many types of parameters that can be used in produce an impact to the overall performance. This may
monitoring the performance of the network. The throughput caused by the losses or any distraction while the transmitting.
parameter that is being used to monitor this cluster network is Table V shows detail values of throughput for LAN while
buffer length. Socket buffers size or length are storage or well Table VI shows value of throughput for WAN for different
known as memory that allocated to a network that used to value of buffer length size. Each test is run 3 times and the
store the data that transmit between application and driver. By average is calculated.
adjusting the buffer length, it can affect the network utilization
and give an impact to the network throughput as well as on the
latency. Sufficient amount of buffer length is important to
obtain satisfactory network utilization. The buffer length that
is too small constitute to the longer waiting time for packets
transmission and make the network under utilization. On the
other hand, if the buffer length is too large, the network
commit to be delayed because apparently the buffer is filled
but still have empty space. Figure 6 shows example of the
TCP/IP buffer size architecture.
Fig. 9. Graphical Output Generated for LAN Using 16KB Buffer Length.
Sample for data that generated from JPerf for LAN and
WAN connection was illustrated as an example in graph
shown in Figure 7 of the impact on the throughput for
different buffer length. File size of the transmission is
generated randomly by the JPerf software. The time interval of
data taken is 20 seconds. It is shown that, in LAN network, the
throughput of buffer length 64KB produced higher throughput
in comparison to the WAN network.
Fig. 10. Graphical Output Generated for WAN using 16KB Buffer Length.
624
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
V. RESULT AND DISCUSSION optimization values for both network cluster implementation.
JPerf is the GUI tools that had been used in this analysis to
The network analysis are focusing on two different cluster monitor the performance of the network.
network topologies connecting cluster nodes as inter-cluster
and the other as an intra-cluster between different graphical ACKNOWLEDGMENT
locations. Tested nodes are connected by the presented
hardware and software specifications. A substantial amount of This research is part of the main initiative of cluster
data moves between the server and clients as well as between computing project at Faculty of Electrical Engineering,
one server to another server intra-cluster measured within a Universiti Teknologi MARA funded by the university. We
time interval, sufficient to produce statistical data for network would like to thank Universiti Teknologi MARA for the
measurement and analysis. opportunity granted.
Detailed throughput information was collected on both
topologies with different sizes of socket buffer length as REFERENCES
presented in both Table V and VI. The following was
concluded based on output observation: [1] Rodrigues, M. A., Saksena, V. R., Performance Analysis of a
Different network topology contributes to different LAN/WAN Bridging Architecture, IEEE Journal on Selected
network throughput for the same duration of interval Areas in Communications, vol. 9, no. 2, pp. 265 - 270, February
monitoring, buffer size and file size. The network 1991.
throughput produced by LAN is highly more stable than it [2] Buyya, R., High Performance Cluster Computing:
is for WAN network. This is also concluded that WAN Architectures and Systems, vol. 1, Prentice Hall, Upper Saddle
network requires additional time to stabilize its River, N. J., USA, 1999.
performance in given interval time as it needs to establish [3] Sadashiv, N., Kumar, S. M. D., Cluster, Grid and Cloud
the longer distance of internet connection as compared to Computing: A Detailed Comparison, 6th International
LAN. Conference on Computer Science & Education, pp. 477 - 482, 3-
5 August 2011.
Frequent greater delay produced as the buffer length
increased. For WAN network, the bandwidth decreases as [4] Frogner, B., Cannara, A. B., Monitoring and Prediction of
Network Performance, International Conference on Advance
the time interval increase.
Issues of E-Commerce and Web-Based Information Systems,
The optimal network throughput for given specification of pp. 122 - 129, August 1999.
LAN network was produced by 256KB size of buffer [5] Cluster Computing, [online],
length. http://en.wikipedia.org/wiki/Computer_cluster (Accessed: 30
The optimal network throughput for given specification August 2013).
WAN network was produced by 256MB size of buffer [6] Chan, A. C. F., Tsang, D. H. K., Gupta, S., Performance
length. Analysis of TCP in the Presence of Random Losses/
The statistical data summarize that with 20 seconds of Errors, IEEE Conference on Global Telecommunications, vol.
time interval monitoring, higher socket buffer size 1, pp. 513 518, 1998.
resulting to lower optimization of network throughput and [7] Huang, B., Bauer, M., Katchabaw, M., Hpcbench - A Linux-
higher data transfer latency. The result of an adequate Based Network Benchmark for High Performance
amount of buffer size for each network has been Networks, 19th International Conference on High Performance
presented in previous points. Computing Systems and Applications, pp. 65 - 71, 15 - 18 May
2005.
These values of buffer size in correspond to each network
proposed an optimization value for the cluster servers and [8] Nomura, A., Matsuba, H., Ishikawa, Y., Network Performance
Model for TCP/IP based Cluster Computing, IEEE Conference
nodes to work effectively in its data communication.
on Cluster Computing, pp. 194 - 203, 17-20 September 2007.
[9] Takefusa, A., Matsuoka, S., Ogawa, H., Nakada, H., Takagi, H.,
VI. CONCLUSION
Sato, M., Sekiguchi, S., Nagashima, U., Multi-Client
LAN/WAN Performance Analysis of Ninf: A High-Performance
The throughput is tested with various size of buffer length
Global Computing System, ACM/IEEE Conference on
in the purpose of monitoring the network performance. Supercomputing, 1997.
Analysis on buffer length presents that selection of buffer
[10] Buffers, [online], http://msdn.microsoft.com/en-
length is important in getting the optimize network data us/library/ms711014(v=vs.85).aspx (Accessed: 30 August
transfer thus avoiding network delay. Buffer length that is too 2013).
small can make the network underutilized while the buffer
length that is too large will contribute to network delay due to
the emptied buffer and need to be filled in. Sufficient amount
of socket buffer length is important in getting an optimum
performance of the network throughput. Statistical
characterization also shown that different network topology
contributes to the network performance thus suggesting
625