Verification of the burst send queuing system model for parallel programs

The assumption that events are statistically independent and uncorrelated in queuing system models can lead to inaccurate estimates of performance for parallel programs. Many parallel programs have very strong interactions between the processors in the system. Programs such as the parallel versions of the NAS benchmarks partition the problem into blocks of elements. Each processor computes the values on a black then it sends a burst of messages to update the values held on neighboring processors. Each processor also expects to receive messages from ifs neighbors before it can start processing the next time step. This behavior in parallel programs lead to the development of the burst send model. The burst model send simulation's results were compared to the original data collected on a PC cluster and to an equivalent queuing model that represents a typical queuing system (and did not group the sends together). The burst send simulation's performance estimates were much more accurate than the ungrouped send model. The worst error for the burst send model was under 20% while the best case error for the typical queuing system was 40%.