ABSTRACT
With the progressive march toward ever larger and faster HPC platforms, the HPC community is seeing a discrepancy between the aggregate bandwidth of the compute node’s capability to send data to storage and the bandwidth available on traditional parallel file systems, utilizing hard drives, to ingest this data. Rather than purchasing additional disks to increase bandwidth (beyond what is required for capacity requirements), the concept of a burst buffer has been proposed to impedance match the compute nodes to the parallel file system [4]. A burst buffer is an allocation of solid-state storage that is capable of absorbing a burst of I/O activity that can then be slowly drained to a parallel file system while computation resumes within the running application. The original design intent of such a system was to handle the I/O workload commonly seen in the checkpoint/restart process, which many current HPC applications use to handle faults. In checkpoint/restart, the application pauses operation and then proceeds to write the state of the entire application’s memory space to a formatted file on disk such that the current application state can be reconstructed should a future failure occur. This process is periodic in nature, but places considerable strain on the I/O subsystem for the several seconds to minutes it takes to complete.
