Performance analysis of an IBM supercluster

Cluster computing has emerged as an alternative approach to deliver high performance to HPC applications. Cluster machines, built on top of commodity off the shelf components and based on open source software are becoming very popular. Moreover, state of the art cluster machines rank among the top most powerful machines, thanks to specialized hardware and software components. It is then important to characterize the actual performance achieved by these machines. We present a performance characterization of a large Linux cluster, that is, the IBM NetFinity located at the Maui High Performance Computing Center. The novelty of this study is that we have analyzed the times spent by the allocated processors to accomplish the various activities required by the application. Our performance characterization follows a bottom up approach. We initially focus on basic communication performance. Analytical models of the times spent in sending and receiving messages, as a junction of their size, are presented. Then, we discuss the behavior of a few widely used numerical algorithms when executed over both a standard Ethernet and on high performance Myrinet interconnection networks. A statistical clustering analysis of these behaviors will complete our performance study.