European teams across scientific disciplines come together to advance their work with GPU acceleration

Helmholtz GPU Hackathon Attendees
Authored by
Publication Date

In September 2020,  the fifth GPU hackathon took place in Germany. Like previous events, it was jointly organised by the Helmholtz Centre Dresden-Rossendorf (HZDR), the Jülich Supercomputing Centre (JSC), OpenACC and NVIDIA; however,  unlike the past, this was a fully online event because of the COVID-19 pandemic. Thanks to the experiences collected during similar online events,  this year’s participating teams showed a similar level of productivity as observed during earlier Helmholtz hackathons.

The teams came from all around Europe and spanned a broad range of natural sciences across a variety of length scales, starting from scales that are much smaller than a microscope could resolve to the macroscopic scales of hydrologic modelling. In total, eight teams were accepted with the majority using  GPUs for accelerating simulations, and two teams focusing on machine learning methods.

The OpenQCBerlin team from DESY and the Humboldt University Berlin used the hackathon to kick-start porting of the OpenQCD code to GPUs. This code is used by theoretical particle physicists for simulating Quantum Chromodynamics (QCD), a theory that is believed to describe the strong interactions between quarks and gluons. Another team addressed challenges in experimental particle physics: The FALCON team uses machine learning techniques to create very fast models for simulating events of large-scale particle physics experiments like the Compact Muon Solenoid (CMS) experiment at CERN.

 The Machine Learning for Molecular Dynamics (MLMD) team from the Institute for Computational Biomedicine at Forschungszentrum Jülich used GPUs for the training of models where the training data is obtained through simulations of molecular systems with hybrid quantum chemical (QC) and molecular mechanical (MM) potentials along very long trajectories.

Another GPU Hackathon success story is the team from the University of Göttingen who worked on the SOMA code, a massively-parallel and GPU accelerated implementation of a soft, coarse-grained polymer model. Building on work achieved at an earlier hackathon, the code allows for the exploration of polymeric materials that are of interest for a wide range of applications from batteries to fuel cells to microelectronic devices. he code can scale to 128 GPUs, and the  team is now preparing for the new JUWELS Booster system at JSC with almost 4,000 NVIDIA A100 GPUs.

Three teams--Mostafa, ParFlow and SHYFEM--came with applications used for hydrologic modelling. The Mostafa team continued earlier efforts to speed-up their Regional Inundation Model on GPUs, which can be used to model river floods. ParFlow simulates surface, soil and groundwater flow and has only recently been ported to GPUs in preparation of the aforementioned JUWELS Booster. SHYFEM is a model that mainly targets modelling of coastal regions like the Venice lagoon. Common to all three codes is the goal of generating highly relevant insights e for the increasingly likely cases of extreme weather conditions.

At the conclusion of the Hackathon, all teams reported that the event was useful for them. Whether they used the event to discover and correct computational bottlenecks to achieve speed-up or to integrate modern and efficient libraries like AmgX, the key takeaway of the Hackathon for the participants was learning and improving their skills across performance tools, libraries, and programming approaches. These are important skills many of the participants will continue to benefit from far after the event itself. In response to the question of whether it was worthwhile to participate, one team answered: “Definitely. This chance is rare. We found exactly what we wanted.”

Acknowledgements

We would like to thank the Technical University Dresden for access to the Power9 cluster accelerated by NVIDIA V100 GPUs, NVIDIA for access to DGX-1 nodes and JSC for access to the JUWELS supercomputer with its V100 and A100 GPUs. The continuous availability of the Nsight team from NVIDIA is gratefully acknowledged. Finally, very special thanks go to the15 mentors who continuously supported the teams during the 4 days of the Hackathon as well as to the organisations that allowed them to take this role: Center for Advanced Systems Understanding (CASUS), HZDR, JSC, and NVIDIA.
 

Type

Author(s)

Dirk Pleiter
Dirk Pleiter

Dirk Pleiter is research group leader at the Jülich Supercomputing Centre (JSC) and professor of theoretical physics at the University of Regensburg. At JSC he is leading the work on application-oriented technology development. Currently he is among others principal investigator of exascale labs at JSC including the NVIDIA Application Lab at Jülich. He co-organised the Helmholtz GPU Hackathon 2020 together with Guido Juckeland (HZDR).

Guido Juckeland
Guido Juckeland

Guido Juckeland is the head of the Computational Science department at Helmholtz-Zentrum Dresden-Rossendorf (HZDR). His team and he are working on research software engineering, applied AI in scientific use cases, large scale research management and GPU-accelerated high performance computing. Guido also teaches GPU programming at Technische Universität Dresden. Together, with Dirk Pleiter, Guido has been organizing GPU Hackathons since 2016.