A Case for Adaptive Redundancy for HPC Resilience

Abstract

Redundancy both in space and time has been widely used to detect and in some cases correct errors in High Performance Computing (HPC) systems. With the HPC community seeking exascale class supercomputers by the end of the decade, unrealistic expectations for correct system behavior will result in exorbitant costs in terms of performance lost and energy expended. Resilience strategies will need to find balance between fault coverage and the overheads incurred. In this work, we propose an adaptive approach that factors in application level knowledge together with runtime inference about the fault tolerance state of the system to dynamically enable redundant multithreading (RMT). Our approach is based on simple programming language extensions, tightly integrated with a compiler infrastructure and a runtime framework that enables managing the performance overheads of redundant computation.