Parallel Programming with Macro-dataflow and Tasks for Scientific Computing

Romain will present his work

Speaker

Romain Pereira

When

Monday, December 08, 13:00 pm

Where

room Grace Hopper 2

Title

Parallel Programming with Macro-dataflow and Tasks for Scientific Computing

Abstract

Exascale and AI-factory systems introduce extreme architectural heterogeneity in processors, memory, and interconnects, making parallel programming for scientific computing increasingly complex. Macro-dataflow and task-based programming can strengthen reliability, performance, and portability of applications. The talk briefly introduces previous work on tooling for reliability and performance analysis, followed by on-going developments in macro-dataflow programming, including XKBlas and OpenMP extensions. It concludes with future directions and opportunities for collaboration.