Volume 55, pp. 687-705, 2022.

Porting an aggregation-based algebraic multigrid method to GPUs

Abdeselam El Haman Abdeselam, Artem Napov, and Yvan Notay

Abstract

We present a hybrid GPU-CPU version of the AGMG software, a popular algebraic multigrid (AMG) solver which implements an aggregation-based AMG method. With the new implementation, the solution stage runs on a GPU, except operations on the coarsest grid, which are executed on a CPU. To maximize the speedup, two novel features are introduced. On the one hand, $\ell_1$-Jacobi smoothing is combined with polynomial acceleration (or polynomial smoothing), leading to improved performance compared with standard $\ell_1$-Jacobi smoothing, while not requiring to compute eigenvalue estimates as standard polynomial smoothing does. On the other hand, besides the K-cycle used in standard AGMG, we introduce the relaxed W-cycle, which tends to combine the advantages of the K-cycle and the standard W-cycle. Numerical results show that the new implementation inherits the robustness of the original AGMG software, while bringing significant speedups on GPUs. A comparison with AmgX, a reference AMG solver from NVIDIA, suggests that the presented hybrid GPU-CPU version of AGMG is more robust and often significantly faster in the solution stage.

Key words

multigrid, linear systems, iterative methods, AMG, preconditioning, parallel computing, GPU

AMS subject classifications

65F10, 65N22, 65Y05, 65Y10

Links to the cited ETNA articles