Mimic: Fast Recovery from Data Corruption Errors in Stencil Computations

Abstract

The largest computing systems routinely run into silent data corruption (SDC) as part of its normal operation. The number of SDCs will increase drastically as computing systems approach the exascale mark, forcing a need to reconsider the resilience approach taken to counteract the effects of unmitigated data corruption errors. Yet any resilience method must be sensitive to both resource and energy requirements. In this paper we explore the propagation of data corruption errors caused in stencil computation, an iterative kernel with structured communication pattern that is found in a wide variety of scientific and engineering problems. We present a computational model, refered to as mimic replication, that provides resilience against SDC errors through dynamic reexecution of processes that are vulnerable to having their data tainted due to a detected latent error. We then provide an analytical model that allows tradeoff between resource and energy consumption and resilience.