The Booster module of the JUWELS supercomputer at JSC consists of 936 compute nodes, each equipped with 4 NVIDIA A100 GPUs and providing a peak performance of over 70 PFLOP/s. Per node, 4 InfiniBand HDR200 adapters are installed, providing 800 GBit/s of network bandwidth per node. The system is in production since end of 2020 and used by many researchers for advanced simulations at large scales.

This workshop will focus on applications already enabled for a few GPUs with the goal to improve scaling and performance on many GPUs.

Dedicated lectures will present state-of-the-art tools and techniques for efficient multi-GPU computing; while the main focus is on interactive hands-on sessions together with mentors, in which the applications of the participants will be analyzed and optimized.

Applicants should ideally have an active compute time project on JUWELS and should already have a user account on the system which can be used during the workshop. Applicants planning to use JUWELS in the near future are welcome as well; access to the machine will be provided. For the most benefit from the workshop, applicants should have in-depth knowledge of their code, libraries it uses, and how to build and launch it on the system. A test/reference case needs to be provided to validate correctness during porting of the application.

An ideal prelude to this workshop is the VI-HPS Tuning Workshop, covering in detail the tools used during the Tuning and Scaling workshop.

Due to capacity restrictions, the maximum number of particpants is limited and early registration is therefore recommended.

For detailed information, agenda and registration, please visit the workshop homepage.