Application-level Fault Tolerance and Resilience in HPC Applications

Abstract

The rapid increase in the computational demands of science has lead to a pronounced growth in the performance offered by supercomputers. As High Performance Computing (HPC) systems grow larger, including more hardware components of different types, the system's failure rate becomes higher. Efficient fault tolerance techniques are essential not only to ensure the execution completion but also to save energy. Checkpoint/restart is one of the most popular fault tolerance techniques. However, most of the research in this field is focused on stop-and-restart strategies for distributed-memory applications in the event of fail-stop failures. Thís thesis focuses on the implementation of application-level checkpoint/restart solutions for the most popular parallel programming models used in HPC. Hence, we have implemented checkpointing solutions to cope with fail-stop failures in hybrid MPI-OpenMP applications and OpenCL-based programs. Both strategies maximize the restart portability and malleability, ie., the recovery can take place on machines with different CPU / accelerator architectures, and/ or operating systems, and can be adapted to the available resources (number of cores/accelerators). Regarding distributed-memory applications, we propose a resilience solution that can be generally applied to SPMD MPI programs. Resilient applications can detect and react to failures without aborting their execution upon fail-stop failures. Instead, failed processes are re-spawned, and the application state is recovered through a global rollback. Moreover, we have optimized this resilience proposal by implementing a local rollback protocol, in which only failed processes rollback to a previous state, while message logging enables global consistency and further progress of the computation. Finally, we have extended a checkpointing library to facilitate the implementation of ad hoc recovery strategies in the event of soft errors) caused by memory corruptions. Many times, these errors can be handled at the software-Ievel, tIms, avoiding fail-stop failures and enabling a more efficient recovery.