'An innovative synergy of some old ideas from the parallel computing community and some new low cost technologies from digital consumer electronics industry' has resulted in the parallel architectural alternative known as High Performance Clusters. 'The legacy concepts are the physical clustering of general purpose hardware resources and the abstract message passing model of distributed computing'. Together these basic capabilities and founding principles have emerged as the new domain of high performance of computing according to Thomas Sterling of Jet Propulsion Laboratories,Caltech.
Thus a cluster is a type of parallel or distributed processing system ,which may consist a collection of interconnected stand alone PCs working together as a single , integrated computing resource. The prominent componets of a cluster architecture include :
Multiple High Performance Nodes
State-of-the-art Operating System
Parallel Development Environment.
The cluster middlware is responsible for offering a unified system image and availability out of a collection of independent but interconnected computers.Parallel Development Environment that includes message passing libraries,debuggers and profiles can offer portable , efficient and easy-to-use tools for development of applications.
An important factor that has made the usage of clusters a practical proposition is the standardization and availability of of tools and utilities under General Public License. In fact this is a boon for developing countries like India , which needs systems to address grand challenges pertaining to the specific national needs.
Designing a Cluster System is a big challenge undoubtedly , but analysing the performance is a bigger one. For anlysing the performance of High Performance Clusters IEEE TFCC has made specific recommendations.The recommended suites are architecture specific , comunication specific and applications specific.
A 8 Node GNU/Linux cluster was recently benchmarked at the Asia Pacific Institute of Information Technolgy[Staffordshire University], Panipat,India , using the IEEE TFCC recommended suites. The results have been found to be extremely useful in designing a 128+ node cluster system with a desirable sustained performance.
The harware/software componets of the system are :Node hardware Intel PIII 1.0 GHz processors,128 MB RAM, L1 Cache 16 KB , L2 cache 256 KB, VRAM 16MB, Chipset Inte i815 , NIC D-LINK DFE 538 TX, HDD 80 GB, HUB: DLINK DES 1024 R+ Ethernet Switch .Node Software : RedHat7.2[Kernel 2.4.7],Resource management & Scheduling Software Nagios 1.0,PDE mpich-1.2.4. IEEE TFCC suites : Parkbench , NASA NPB2.3.
Though Theoretical Peak Performance of the system is 8 GFLOPS , the sustained performance measured as the Total Mop/s for NPB LU Class C suite , is 217.69. The system has also withstood the maximum problem size i.e. 2.86 TFLOP in NASA NPB2.3 suite.
Thus the problem size is also an important factor in designing GNU/Linux cluster systems.
Dr H K Sahjwani ,Director of the Institute commenting on the performance analysis results opined that the clusters are best able to track improvements in processor/network technology and are thus best suited for infrastructural developments in India.