1Investigating Lightweight Storage and Overlay Networks for Fault Tolerance

Abstract—The next generation of capability-class massively parallel pro-cessing (MPP) systems is expected to have tens-to-hundreds of thousands of processors, with individual applications consuming large fractions of the system. In such an environment, it is critical to have fault-tolerance mecha-nisms that allow continuous computing with minimal performance impact on the application. Unfortunately, the current “in-practice ” approaches to fault tolerance do neither. This paper analyzes the performance impact of exiting approaches on next-generation systems and describes a new project at Sandia National Laboratories to investigate the use of “lightweight ” stor-age architectures and overlay networks for fault tolerance. The combined use of these technologies has the potential to significantly reduce the I/O impact of checkpoint operations on application performance and allow ap-plications to recover from component failures without a complete restart and with minimal use of compute node resources. I