The ability to understand the factors contributing to parallel program performance are vital for understanding the impact of machine parameters on the performance of specific applications. We propose a methodology for analyzing the performance characteristics of parallel programs based on message-passing traces of their execution on a set of processors. Using this methodology, we explore how perturbations in both single processor performance and the messaging layer impact the performance of the traced run. This analysis provides a quantitative description of the sensitivity of applications to a variety of performance parameters to better understand the range of systems upon which an application can be expected to perform well. These performance parameters include operating system interference and variability in message latencies within the interconnection network layer.
Matthew J. Sottile, Vaddadi P. Chandu, David A. Ba