ºÚÁÏÍø

Title of the thesis: Graphics Processors in High-Performance Computing: Practice and Experience in the Simulation of Astrophysical Magnetohydrodynamics and Turbulence

Thesis defender: Johannes Pekkilä
Opponent: Professor James McLellan Stone, Institute for Advanced Study, School of Natural Sciences, New York, USA 
Custos: Professor Maarit Korpi-Lagg, Aalto University School of Science

The study examines techniques to employ graphics processors in extreme-scale simulations of astrophysical magnetohydrodynamics on supercomputers. 

Current state-of-the-art computer models can reproduce real-world observations of the Sun only in rough detail. One of the reasons for the discrepancy is believed to be the lack of computing power needed to resolve physical features across an extremely wide range of scales, even with the fastest supercomputers. The key contribution of the thesis, "Graphics Processors in High-Performance Computing: Practice and Experience in the Simulation of Astrophysical Magnetohydrodynamics and Turbulence", is its extensive examination on how graphics processors can be leveraged to boost the accuracy, throughput, and energy efficiency of simulations of the Sun. 

The thesis focuses on the computational aspects of such simulations, proposing new algorithms for efficient computation and data movement in a network connecting thousands of graphics processors. It puts special focus on mitigating data movement bottlenecks, which is universally important for data-heavy computations because of the growing performance gap between the compute and memory systems of modern computers. The work also proposes a new programming language for a class of tasks prevalent in computational sciences, namely, stencil computations. This generalization makes the contributions usable also for other applications, such as earthquake simulations and AI-tools based on convolutional neural networks. 

In addition to introducing these methods in an open-source software framework, the thesis proposes a novel implementation for the test-field method for electrically conducting fluids, obtaining 10× speedup per supercomputing node compared to previous state of the art. It also demonstrates efficient techniques for communicating between thousands of graphics processors, obtaining ≥ 90% weak scaling efficiency to 4096 processors on the LUMI supercomputer. By enabling a leap in the throughput and energy efficiency of plasma simulations, the work presents a performance milestone on the path toward more accurate predictions of space weather, solar cycles, fusion simulations, and other phenomena involving electrically conducting fluids.

Thesis available for public display 7 days prior to the defence at .