Enhancing Load-Balancing of MPI Applications with Workshare

Some high-performance parallel applications (e.g., simulation codes) are, by nature, prone to computational imbalance. With various elements, such as particles or multiple materials, evolving in a fixed space (with different boundary conditions), an MPI process can easily end up with more operations to perform than its neighbors. This computational imbalance causes performance loss. Load-balancing methods are used to limit such negative impacts. However, most load-balancing schemes rely on shared-memory models, and those handling MPI load-balancing use too much heavy machinery for efficient intra-node load-balancing. In this paper, we present the MPI Workshare concept. With MPI Workshare, we propose a programming interface based on directives, and the associated implementation, to leverage light MPI intranode load-balancing. In this work, we focus on loop worksharing. The similarity of our directives with OpenMP ones makes our interface easy to understand and to use. We provide an implementation of both the runtime and compiler directive support. Experimental results on well-known mini-applications (MiniFE, LULESH) show that MPI Workshare succeeds in maintaining the same level of performance as well-balanced workloads even with high imbalance parameter values.