Roofline on NVIDIA GPUs Hackathon, July 8, 2020

Introduction

The Roofline performance model provides an intuitive and insightful way to understand application performance, identify bottlenecks and perform optimization for HPC applications. In preparation for the upcoming supercomputer Perlmutter at NERSC, this hackathon invites all NERSC users to join us in some informative talks, lively discussions and engaging hands-on sessions, all related to Roofline analysis on NVIDIA GPUs!

This hackathon will be a full-day event, with the morning dedicated to a few talks and demos, and the afternoon hands-on 'hacking' for attendees. The talks will cover some fundamentals of the Roofline model, the mechanism behind Roofline data collection on NVIDIA GPUs, and the newly released fully automated Roofline feature in Nsight Compute. A demo will then illustrate how Roofline analysis is done on a real-life HPC application, showcasing a few optimization techniques that attendees may find useful for their own applications. With NERSC, CRD and NVIDIA staff onsite (virtually), the afternoon is designed for attendees to try out the Roofline methodology and the latest features in Nsight Compute on their own applications, and before closing out, there will be a brief 'scrum' session where attendees each spend a few minutes summarizing their findings and/or problems during this hackathon. Your feedback would be extremely valuable to both the Roofline team at LBL and the tools team at NVIDIA!

Tentative Agenda 

Registration

Please register here. Registration will close at COB on June 30 PST.

If you have any questions, please feel free to contact Charlene Yang at cjyang@lbl.gov.

During and after the event (till Aug 8), please provide your feedback regarding the event itself, the Roofline model or the Nsight tools here.

The recordings of the talks can be found here.