Reinit[Formula: see text]: Evaluating the Performance of Global-Restart Recovery Methods for MPI Fault Tolerance

Scaling supercomputers comes with an increase in failure rates due to the increasing number of hardware components. In standard practice, applications are made resilient through checkpointing data and restarting execution after a failure occurs to resume from the latest checkpoint. However, re-deploying an application incurs overhead by tearing down and re-instating execution, and possibly limiting checkpointing retrieval from slow permanent storage. In this paper we present Reinit[Formula: see text], a new design and implementation of the Reinit approach for global-restart recovery, which avoids application re-deployment. We extensively evaluate Reinit[Formula: see text] contrasted with the leading MPI fault-tolerance approach of ULFM, implementing global-restart recovery, and the typical practice of restarting an application to derive new insight on performance. Experimentation with three different HPC proxy applications made resilient to withstand process and node failures shows that Reinit[Formula: see text] recovers much faster than restarting, up to 6[Formula: see text], or ULFM, up to 3[Formula: see text], and that it scales excellently as the number of MPI processes grows.