One of the topics concerns the benefit of MPI Sessions for CUDA-Aware communication libraries. MPI is the de facto standard for distributed computing. CUDA-Aware libraries were introduced to ease GPU inter-node communications. However, they induce some overhead that can degrade overall performance. The MPI 4.0 Specification standard introduces the MPI Sessions model which offers the ability to initialize specific resources for specific components of the application. We present a way to totally remove the overhead induced by CUDA-aware libraries with a solution inspired by MPI Sessions.

The other topic will be the presentation of CDSS 2022. This year the CDSS will focus on hardware/software co-design. The execution efficiency of any application depends on both its algorithmic implementation and the underlying architecture. For this Co-Design project we focus on Matrix Chain Multiplication algorithms and explore the costs associated with the underlying memory capacity. The idea of CDSS this year is to evaluate the energy cost of varying the shared memory capacity of GPUs. Our goal is to design an optimal algorithm for Matrix Chain Multiplication on GPUs that minimizes computation as well as off-chip data transfers.

Presenters:

Mina Warnet is a Master student at the Université de Reims Champagne-Ardennes, studying High Performance and Visual Computing. She's attending the Co-Design Summer School 2022 and plan to carry on in Co-Design Computer Science and Physics/Chemistry.

Maxim Moraru is a PhD student in computer science at the University of Reims. Its research concerns the inter-node communication between GPUs. Maxim is studying and optimizing the performance of a French proprietary interconnect called Bull eXascale Interconnect.

Host: Julien Loiseau