Acceleration of MPI mechanisms for sustainable HPC applications

Abstract

Ultrascale computing systems are meant to reach a growth of two or three orders of magnitude of today computing systems. However, to achieve the performances required, we will need to design and implement more sustainable solutions for ultra-scale computing systems, understanding sustainability in a holistic manner to address challenges as economy-of-scale, agile elastic scalability, heterogeneity, programmability, fault resilience, energy efficiency, and scalable storage. Some of those solutions could be provided into MPI, but other should be devised as higher level concepts, less generalists, but adapted to applicative domains, possibly as programming patterns or or libraries. In this paper, we show some proposals to extend MPI trying to cover major domains that are relevant towards sustainability: MPI programming optimizations and programming models, resilience, data management, and their usage from applications.