SRC14. The Search for Missing Parallel I/O Performance on the Cori Supercomputer
Student: Matt Bryson (Lawrence Berkeley National Laboratory)
Supervisor: Suren Byna (Lawrence Berkeley National Laboratory)
Abstract: With scientific computing approaching exascale rapidly, I/O is becoming the main cause of bottlenecks in High Performance Computing (HPC). To solve this, next generation systems such as the NERSC Cori Supercomputer are equipped with I/O nodes equipped with NVRAM, otherwise known as burst buffers (BB). BB systems are designed to provide more I/O throughput than traditional Parallel File Systems (PFS) that rely on magnetic storage. This has not been the case with the Cray Burst Buffer (CBB) which is performing at 11.07% of peak performance when confronted with the popular HDF5 file library. We use the Vector Particle in Cell (VPIC) I/O kernel to benchmark the system and find potential optimization strategies. By changing the I/O access pattern of VPIC I/O kernel we are able to improve performance up to 4.6 times in some configurations.
Two-page extended abstract: pdf