GPU Correctness and Debugging Workshop - Task 054

Fit to programme

This task has been identified by the working groups as part of the agenda behind WP 2.4.

The task number is 054.

Summary

Performance is not always just about how fast is a code; performant development is also crucial. Debugging an application by hand can take an incredibly long time, and hence tools exist to aid with this. For HPC, we now have access to parallel debuggers, compiler sanitisers, race detectors etc. because parallelism adds further complexity to debugging. As SHAREing is an accelerated computing hub, we propose coordinating a workshop on GPU-focused correctness and debugging. Many dRTPs are involved in porting so I think there is a significant gap in supporting these developers in appropriate tools for their porting work.

Methodology

The work will likely involve an initial exploratory analysis of available correctness and debugging tools for GPU workloads. Then for the workshop, adjustments should be made to ensure that it is as accessible as possible. In-person workshops can be highly beneficial but perhaps coordinating this event as hybrid allows for wider attendance. Similarly, avoiding school holidays can accommodate people with childcare needs. We recommend building in EDIA principles of event coordination by following guidance from CAKEbox: https://www.cake.ac.uk/CAKEbox/edia/

Outcomes

Firstly we should have a preliminary curriculum for available GPU-specific correctness and debugging tools that should be circulated through the SHAREing consortium. Then coordinate a training workshop that teaches how to get started with these tools, and any potential differences of GPU debugging than the more standard CPU techniques. The materials of this workshop should then be added to SHAREing webpages so that they are available for future workshops, or self-directed learning. Finally, after the workshop, it would be good to see a reflective document, e.g., a blog for the SHAREing website.