Enabling Performance Efficient Runtime Support for Hybrid MPI+UPC++ Programming Models

Abstract

UPC++ is an emerging Partition Global Address Space (PGAS) library. The performance of UPC++ applications is highly dependent on the efficiency of the communication runtime. Current UPC++ runtime relies on MPI and native IB verbs based conduit implementations of GASNet communication middleware. Hybrid MPI+UPC++ programming model is seen as an attractive way to program next generation systems. However, current approach towards a hybrid model requires the use of the GASNet-MPI conduit which implements one model on top of the other, thereby significantly limiting the performance. In this paper, we propose a unified runtime based approach that natively unifies both MPI and UPC++ models to achieve better performance and deadlock free execution. We evaluated our design on a pure UPC++ based microbenchmark which we implemented on top of OSU Microbenchmark (OMB) suite. To evaluate application level performance, we implemented a Gauss-seidel based 2D-heat diffusion kernel in UPC++ and redesigned LULESH (3-D shock hydrodynamics simulation) using MPI-3 Non-blocking Collectives (NBC) for hybrid MPI+UPC++ models. At microbenchmark level, the proposed approach achieves up to 15X and 35X speedup over GASNet-MPI and GASNet-IBV conduits, respectively for dense collective operations on 128 cores. The application level evaluation of 2D-heat shows that our proposed approach achieves up to 30% improvement on 512 cores. Finally, we demonstrate the design benefits of NBC redesigned hybrid implementation of LULESH over pure MPI and pure UPC++ and show that our unified runtime based approach can achieve up to 6% improvement in total execution time on 216 (63) cores.