On checkpointing strategies in unreliable computing environments

Abstract

In this paper, we analyze performance implications of checkpointing strategies in unreliable computing environments. We show that if the appropriate checkpointing strategy is not chosen, the time to complete a job is heavy-tailed distributed. This can lead to highly-variable and long completion times. We generate asymptotics for job completion times when there is no checkpointing, a fixed number of random checkpoints, and when checkpoints occur at fixed intervals for various task time distributions. Our asymptotic results are derived using large deviation theory.