Influence of Runtime Parameters on the Performance of Parallel Applications

Most parallel and sequential applications achieve a low percentage of the theoretical peak performance of a given processor and platform. While some of the reasons for this performance problem, such as cache utilization or the impact of the numerical algorithms used, have been explored, the influence of the communication library on the performance of a parallel application is not well understood. The goal of this thesis is to perform an extensive parameter study to evaluate the impact of runtime parameters of the Open MPI communication library on application performance and explore various approaches for tuning the parameters in order to minimize the execution time of an application. Among the parameters tuned in this work were point-to-point parameters of the InfiniBand network, shared memory commmunication parameters, and algorithms used for collective communication operations. Two NAS Parallel Benchmarks were used to evaluate the overall performance gain in each scenario. The results indicate that by changing the values of some parameters of the Open MPI communication library, the execution time of the applications could be reduced by 8 to 40% compared to the default configuration.