Aggregate Memory as an Intermediate Checkpoint Storage Device

Abstract

Applications that generate bursty I/O load, like checkpointing, require additional support to perform efficiently on next generation petascale supercomputers. Tens of thousands of processors, generating terabytes of snapshot data at once at each timestep, can easily overwhelm a storage system. Further, even at the current peak I/O bandwidth rates, offered by parallel file system deployments at leadership class facilities, an application is likely to spend a significant portion of its runtime checkpointing. To address these issues, we propose a checkpoint storage device, built from memory resources, that acts as an intermediary to the central parallel file system. Our system comprises of a dedicated manager that aggregates memory resources from processors (benefactors) and makes it available as a collective space for checkpointing clients, using a standard POSIX file system interface. We argue that such a system has the potential to alleviate the I/O bandwidth bottleneck for bursty I/O operations like checkpointing by aggregating memory and interprocessor bandwidth.