Simulation Framework for Asynchronous Iterative Methods

Abstract

As high-performance computing (HPC) platforms progress towards exascale, computational methods must be revamped to successfully leverage them. In particular, (1) asynchronous methods become of great importance because synchronization becomes prohibitively expensive and (2) resilience of computations must be achieved, e.g., using checkpointing selectively which may otherwise become prohibitively expensive due to the sheer scale of the computing environment. In this work, a simulation framework is proposed and tested to examine the potential benefit of asynchronous iteration for various HPC accelerator architectures (which typically admit different granularities of computations). Additionally, an example of a case study using the simulation framework is presented to examine the efficacy of different checkpointing schemes for asynchronous relaxation methods.