Programmer-directed partial redundancy for resilient HPC

Abstract

In this work we propose partial task replication and checkpointing for task-parallel HPC applications to mitigate silent data corruption (SDC) errors. As the complete replication of all application tasks can be prohibitive due to resource costs, we introduce programmer-directed selective replication mechanism to provide fault-tolerance while decreasing costs. Results show that our scheme detects and corrects around 65% of SDC errors with only 4% overhead on average.