Optimizing HPC Workflows with Versatile Data Movement for AI

Hyungro Lee, Rutgers University
AI for Science report

Abstract:  Scientific HPC workflow often suffers from significant I/O overheads while sub-processes exchange data throughout the workflow and further increase with iterations. He demonstrates the performance and optimization of HPC workflow with scientific use cases which drives his work on user-level data management. The talk will include technical tips and implementations to expose how much speedup can achieve and what changes need to enable scalable applications over different HPC platforms, including ORNL Summit, LLNL Lassen, PSC Bridges, and SDSC Comet. Although the original work focused on specific applications in processing molecular dynamics pipelines, he finds that the experimental results show the advantages and challenges of versatile data movement to distributed ML frameworks. 


Please use this link to attend the virtual seminar:


Meeting ID: 756878381 / Participant passcode: 8182