D4.3 Planned MPI-related optimizations

Edited by Hugo Taboada (CEA)


Gilles Moreau (CEA), Hugo Taboada (CEA), Marc Pérache (CEA), Simon Pickartz (ParTec), Carsten Clauss (ParTec)

Executive summary

This document presents the two contributions from Commissariat à l’énergie atomique et aux énergies alternatives (CEA) and ParTec on the development of Message-Passing Interface (MPI) optimizations on their respective platforms Multi-Processor Computing (MPC) and ParaStation MPI. This work has been conducted in the context of WP 4, Task 4.5. The objective of this intermediary deliverable is the implementation of first prototypes of MPI-related optimisations for BullSequana eXascale Interconnect (BXI) networks.

CEA developed a first prototype of a feature called multirail to take advantage of new node infrastructure containing multiple Network Interface Cards (NICs). Developments in the meantime also layed the ground for potential collaboration with the DEEP-SEA project on the integration of a gateway feature used in supercomputers with Modular Supercomputer Architecture (MSA), WP3 Task 3.3. In this context, ParTec worked on the extension of ParaStation MPI by support for BXI networks. This does not only enable efficient communication over BXI using the ParaStation communication stack but also benefits from MSA-aware features such as gateway communication. Both contributions show promising results that pave the way for upcoming studies and optimisations.