Performance Analysis of Multi-Containerized MD Simulations for Low-Level Resource Allocation

This study discusses scheduling strategies to maximize ensemble throughput, which is the total throughput of multiple containers running simultaneously. Such a strategy is useful, for example, in ensemble runs of molecular dynamics (MD) simulations. To design the strategies, we need to tackle two major challenges: (1) how many containers and how many threads per container we should allocate, and (2) which low-level resources we should allocate to reflect workload characteristics. In particular, the latter challenge is important and inevitable for performance-sensitive applications because they effectively utilize low-level hardware such as simultaneous multi-threading (SMT) to maximize performance, while most container platforms do not handle the challenge. In this paper, as a preliminary experiment to implement scheduling strategies related to SMT, we examined whether ensemble throughput of MD simulations can be improved by deploying containers on separate logical cores even when they share the same physical cores. As a result, we obtained a 2.22-fold ensemble throughput compared with a one-container execution with 10 physical cores.