Accelerating sparse iterative solvers and preconditioners using RACE

Dr. Christie Alappat
Erlangen National High Performance Computing Center
Friedrich Alexander University

"Accelerating sparse iterative solvers and preconditioners using RACE"
Thursday, July 25, 2024, 3:00-3:40 pm UTC (30min talk + 10min questions)
8 am PDT / 10 am CDT / 11 am EDT / 3 pm UTC / 5 pm CEST / 12 am JST

Watch this webinar here

Supercomputing Spotlights is a new webinar series featuring short presentations that highlight the impact and successes of high-performance computing (HPC) throughout our world. Presentations, emphasizing achievements and opportunities in HPC, are intended for the broad international community, especially students and newcomers to the field. Supercomputing Spotlights is an outreach initiative of SIAG/Supercomputing (https://siag-sc.org) … Join us!

Abstract: 

The sparse matrix-vector multiplication (SpMV) kernel is a key performance-limiting component of numerous algorithms in computational science. Despite the kernel's apparent simplicity, the sparse and potentially irregular data access patterns of SpMV and its intrinsically low computational intensity have been challenging the development of high-performance implementations of sparse algorithms over decades. In this talk, we present methods to increase the computational intensity and thereby accelerate the performance of SpMV kernels. The method is based on the concept of levels as developed in the context of our RACE library framework. We demonstrate that one can typically achieve a speedup of 1.5-4x on a single modern Intel or AMD multicore chip for symmetric SpMV and matrix power kernels using this level-based approach.

After briefly introducing the optimization strategy, we apply these optimized kernels in iterative solvers. To this end, we discuss the coupling of the RACE library with the Trilinos framework and address the application to communication-avoiding s-step Krylov solvers, polynomial preconditioners, and algebraic multigrid (AMG) preconditioners. We then dive into the performance benefits and challenges of the RACE integration and show that our optimization produces numerically identical results and improves the total solver time by 1.3x - 2x.

Bio: Christie Louis Alappat received a master's degree with honors from the Bavarian Graduate School of Computational Engineering at the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is currently working as a research assistant at Erlangen National High-Performance Computing Center and is in the final stages of completing his doctoral studies under the guidance of Dr. Gerhard Wellein. His research interests include performance engineering, sparse matrix and graph algorithms, iterative linear solvers, and eigenvalue computations. He has received numerous awards including the 2017 Software for Exascale Computing Best Master Thesis Award, the 2018 Supercomputing ACM Student Research Competition (SRC) Award, second place in the 2019 ACM SRC grand finals, and the 2020 International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems Best Short Paper Award.

Christie Alappat, Friedrich Alexander University