Enhancing Asynchronous Iterative Linear Solvers Through Randomization
Masha Sosonkina
21 May 2019, 10h30 Salle/Bat : 465/PCRI-N
Contact :
Activités de recherche : Calcul à haute performance
Résumé :
Asynchronous implementations present a mechanism to improve the parallel performance of iterative linear system solution methods on highly parallel computing platforms by removing the synchronization overhead. In this talk, after a brief overview of asynchronous computations, we consider a class of asynchronous iterative linear system solvers that employ randomization to determine the component update orders, specifically focusing on the effects of non-uniform distributions. Results show that using such distributions may lead to a faster convergence than that when components are selected for updates uniformly. Finally, we note on the performance of hybrid parallel implementations of asynchronous methods.
