Porting scientific libraries to PGAS in XSEDE resources

Abstract

The next generation of supercomputers presents new and complex challenges that might require a change in the current paradigm of how parallel applications are developed. Hybrid programming is usually described as the best approach for exascale computers. PGAS programming models are considered an interesting alternative to work together with MPI in this hybrid model to achieve good performance in those machines. This is a very promising approach especially for one-sided and irregular communication patterns. However, this is still an emerging technology and there is not much previous experience on how to port existing MPI applications to the PGAS model. Due to the promising relevance of this approach for the next generation of devices, it is relevant to have early experience on porting applications as well as knowledge on the issues that might be faced in this new paradigm. In this paper we present two different scientific applications that are currently implemented in MPI and that are promising candidates for this PGAS paradigm. We describe how these applications have been ported, the challenges faced and some of the solutions that we found. We also show how PGAS models can achieve great performance when compared to MPI.