An application-driven study of parallel system overheads and network bandwidth requirements

Evaluating and analyzing the performance of a parallel application on an ar chitecture to explain the disparity between projected and delivered perform ance is an important aspect of parallel systems research. However, conducti ng such a study is hard due to the vast design space of these systems. In t his paper, we study two important aspects related to the performance of par allel applications on shared memory parallel architectures. First, we quant ify overheads observed during the execution of these applications on three different simulated architectures. We next use these results to synthesize the bandwidth requirements for the applications with respect to different n etwork topologies. This study is performed using an execution-driven simula tion tool called SPASM, which provides a way of isolating and quantifying t he different parallel system overheads in a nonintrusive manner. The first exercise shows that in shared memory machines with private caches, as long as the applications are well-structured to exploit locality, the key determ inant that impacts performance is network connection. The second exercise q uantifies the network bandwidth needed to minimize the effect of network co nnection. specifically, it is shown that for the applications considered, a s long as the problem sizes are increased commensurate with the system size , current network technologies supporting 200-300 MBytes/sec link bandwidth are sufficient to keep the network overheads (such as latency and contenti on) within acceptable bounds.