Implementing Asynchronous Linear Solvers Using Non-Uniform Distributions
Erik J. Jensen, Evan Coleman, Masha Sosonkina · ODU Digital Commons (Old Dominion University) · 2020
Asynchronous iterative methods may improve the time-to-solution of their synchronous counterparts on highly parallel computational platforms. This paper considers asynchronous iterative linear system solvers that employ non-uniform randomization and develops a new implementation for such methods. Experiments with a two-dimensional finite-difference discrete Laplacian problem are presented. The new finer grain implementation is compared with an existing, block-based, one and shown to be superior in terms of the convergence speed and accuracy. In general, using non-uniform distributions in selecting components to update may lead to faster convergence. In particular, the new implementation convergences up to 10% faster when it uses a non-uniform distribution.