Parallel in Time simulations – An overview of applications of the Parareal algorithm

by Debasmita Samaddar (University of Chicago)

Monday Nov 28, 2022, 2:00 PM → 3:00 PM America/Chicago

The saturation of clock-speed has led modern supercomputers to require extreme parallel scaling. With the increasing availability of processors, parallelization is the best approach to improve performance. However, standard parallelization techniques, such as spatial parallelism, experience saturation i.e no more computational speed up is observed with increasing number of processors beyond a point.
The last two decades have witnessed an increased interest in temporal parallelization. Parallelizing time apparently violates causality and hence poses to be counter-intuitive. However, algorithms have been explored that achieve in adding another dimension, namely time, to parallelize computational codes.
This talk particularly focuses on the Parareal Algorithm. We demonstrate that the algorithm can be successfully applied to large scale complex problems including turbulence simulations. We discuss the limitations and also discuss some of the variations that lead to enhanced performance of the algorithm. These variations include modifications to the algorithm itself as well as the modes of implementation, leading to breakthroughs in significantly improving performance compared to the existing Parareal scheme.

Samaddar Presentation.mp4