ONLINE -- MSA Seminar: Scaling Lattice QCD on many GPU Nodes of JUWELS

Modular Supercomputing Architectures Seminar

Speakers: Eric Gregory (JSC), Mathias Wagner (NVIDIA)

Lattice QCD is a first-principles approach to calculating properties of hadrons – composite particles bound by the strong force. Many advances of the past 15 years have been made using a collection of community software which incorporate state-of-the-art algorithms into hardware-specific backend libraries.

QUDA is one of these backend libraries, aiming to increase the scientific output on NVIDIA GPUs. It features state-of-the-art implementations of current algorithms and infrastructure to research into novel algorithms and techniques, specially tuned to the hardware.

This talk has two parts:
Eric Gregory (JSC) will briefly introduce Lattice QCD, including the physics background and the type of questions it can answer. He will show initial experiences running simulations on the GPU nodes of JUWELS and compare the results to results from the JURECA Booster (KNL).
Mathias Wagner (NVIDIA) will introduce QUDA and highlight key features like mixed-precision solvers and kernel-autotuning and how to make them work well over large systems by using GPUDirect RDMA and NVSHMEM. He will also touch on recent algorithmic developments like Multigrid and block-solvers that improve the scientific throughput.

This seminar is offered online via BigBlueButton at https://webconf.fz-juelich.de/b/and-2jh-qj4 .

In case of difficulties with BigBlueButton, the seminar will be switched to DFNconf streaming at https://www.conf.dfn.de/stream/nr5hm44jym4c4