Accelerating Applications at Scale Using One-Sided Communication

The lower overhead associated with one-sided messaging as compared to traditional two-sided communication has the potential to increase the performance at scale of scientific applications by increasing the effective network bandwidth and reducing synchronization overheads. In this work, we investigate both the programming effort required to modify existing MPI applications to exploit one-sided messaging using PGAS languages, and the resulting performance implications. In particular, we modify the MILC and IMPACT-T applications to use the one-sided communication features of Unified Parallel C (UPC) and coarray Fortran (CAF) languages respectively. Only modest modifications to the source code are required where fewer than 100 lines of the source code out of 70,000 need to be changed for each application. We show that performance gains of more than 50% can be achieved at scale, with the largest benefits realized at the highest concurrencies (32,768 cores).