Evaluating Simultaneous Multi-threading and Affinity Performance for Reproducible Parallel Stochastic Simulation

This paper investigates whether simultaneous multi-threading (SMT) can improve performance on modern computing clusters with reproducible results on four types of applications, focused on stochastic simulations with different memory bound and compute bound constraints. We manually set the affinity of processes to compare its efficiency with the computing time obtained by the automatic assignment of the operating system. To measure SMT and affinity impact on a modern multicore processor, we parallelize up to 128 processes of the four types of applications. We expect repeatable numerical results between the sequential and parallel versions of simulations. For the three applications that are not memory bound, SMT is more effective by up to 30%. This represents an interesting increase up to 10% more performance for compute bound applications when compared to the initial papers discussing the efficiency of SMT. However, for the memory-bound application, SMT is less effective and can even decrease performance. The manual setting of core affinity does not show an increase in performance compared to the automatic assignment. All code and data used in the study are available to help reproducible research.