For the past 9 years, the Gauss Centre for Supercomputing (GCS) has given out the Gauss Award during the opening session of the International Supercomputing Conference (ISC), taking place for the last several years in Frankfurt am Main, Germany.
This year, the Gauss Award committee selected a team of researchers from the Extreme Computing Research Center (ECRC) at the King Abdullah University of Science and Technology (KAUST). The multigenerational, multinational team focused on optimizing a class of solvers for data-sparse high-performance computing (HPC) applications. More specifically, the team focused their optimization work on acoustic boundary integral equations, common in a variety of engineering and fluid dynamics applications.
“It is hard to imagine a more exciting capstone for a thesis in high-performance numerics than to have it associated with the name of Gauss,” said Noha Alharthi, a recent PhD graduate from KAUST and lead author on the paper.
Due to the COVID-19 pandemic, ISC20 converted from a physical to a digital event, taking place June 22–25. The team received the award and the €5,000 prize during the first day of the conference.
The team’s work connects with an increasingly important theme across the HPC community—the need to optimize how applications use HPC systems’ memory in order to make the most out of increasingly powerful supercomputers. The ability for an application to fully exploit a systems’ raw computational capabilities rests largely in how efficiently researchers can fit highly resolved models into a given memory and not be limited by memory bandwidth.
“The transition to new systems is slow because memory management is not explicitly and routinely provided in most scientific programming languages, and most computational scientists would prefer to concentrate on other aspects of the computation,” said Dr. David Keyes, Director of the ECRC and a co-author on the paper.
To that end, the team implemented a “tile low-rank” method for performing lower-upper (LU) factorization and solution, a common method of solving linear algebra equations. The approach was implemented in the Hierarchical Computations on Manycore Architectures (HiCMA) library developed in the ECRC, and demonstrated on a variety of modern HPC processors from industry leaders such as Intel, NVIDIA, AMD, and ARM, among others. The StarPU dynamic runtime system was used to schedule the computationally intensive blocks.
The team gained an order of magnitude speedup over conventional dense implementations in solving 3D acoustic boundary integral equation (BIE) problem containing 2.5 million unknowns. While the team’s application focused specifically on acoustic scattering of a submarine, acoustic BIE problems are common in HPC applications in other fluid dynamics or physics problems.
“Exploiting data sparsity—found in many applications—leads to a renaissance in numerical linear algebra,” said Hatem Ltaief, lead architect of HiCMA. “This ultimately yields orders of magnitude for a given problem, or allows for much larger problems.”
While Keyes has been a thought leader in HPC space for 30 years, he primarily played a support role for the next generation of HPC experts coming from KAUST’s burgeoning HPC focus, which was built in to the decade-old university at its creation. (Editor’s note: David Keyes is part of the organizing committee for ISC, but had no role in the selection process for the Gauss Award—the committee consists of 7 researchers from Asia, Europe, and North America who make their independent decisions in a double blind evaluation process).
Both Alharthi and fellow PhD candidate Rabab Alomairy upended their family lives for several weeks as the submission deadline drew near. As a result the team decided to focus the award money on ensuring that these early career researchers have more access to conference and publication opportunities to support their careers with additional networking opportunities.
“Noha and I became sisters and our team became a family during the push to extend our tile low-rank Cholesky solver to full LU capability and hook it into Noha’s thesis,” said Rabab Alomairy, the paper’s second author, whose own work is in dynamic runtime systems. “We hope many others will find it useful.”