A 'cool' way of improving the reliability of HPC machines
SC '13 Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis
| Hi-index | 0.00 |
Faults have become the norm rather than the exception for high-end computing on clusters with 10s/100s of thousands of cores, and this situation will only become more dire as we reach exascale computing. Exacerbating this situation, some of these faults will not be detected, manifesting themselves as silent errors that will corrupt memory while applications continue to operate but report incorrect results. This paper introduces RedMPI, an MPI library residing in the profiling layer of any standards-compliant MPI implementation. RedMPI is capable of both online detection and correction of soft errors that occur in MPI applications without requiring code changes to application source code. By providing redundancy, RedMPI is capable of transparently detecting corrupt messages from MPI processes that become faulted during execution. Furthermore, with triple redundancy RedMPI "votes'' out MPI messages of a faulted process by replacing corrupted results with corrected results from unfaulted processes. We present an evaluation of RedMPI on an assortment of applications to demonstrate the effectiveness and assess associated overheads. Fault injection experiments establish that RedMPI is not only capable of successfully detecting injected faults, but can also correct these faults while carrying a corrupted application to successful completion without propagating invalid data.