ExPBB : A framework to explore the performance of Burst Buffer

ShaheenII supercomputer provides 268 Burst Buffer nodes based on Cray DataWarp technology. Thus, there is an extra layer between the compute nodes and the parallel filesystem by using SSDs. However, this technology is new, and many scientists try to understand and gain the maximum performance. We present an auto-tuning I/O framework called Explore the Performance of Burst Buffer. The purpose of this project is to determine the optimum parameters to acquire the maximum performance of the executed applications on the Burst Buffer. We study the number of the used Burst Buffer nodes, MPI aggregators, striping unit of files, and MPI/OpenMP processes. The framework aggregates I/O performance data from the Darshan tool and MPI I/O statistics provided by Cray MPICH, then it proceeds to the study of the parameters, depending on many criteria, till it concludes to the maximum performance. We report results, where in some cases we achieved speedup up to 4.52 times when we used this framework. Keywords-DataWarp, Burst Buffer, I/O, performance, optimization