Lustre I/O performance investigations on Hazel Hen: experiments and heuristics

Abstract With ever-increasing computational power, larger domains are employed and thus the data output grows as well. Writing this to disk can become a significant part of runtime if done serially. Even is in parallel, e.g., via MPI I/O, there many user-space parameters for tuning performance. This paper focuses on available Lustre file system Cray MPICH implementation I/O. Experiments XC40 Hazel Hen using Sonexion 2000 were conducted. In experiments, core count, block size striping configuration varied. Based these parameters, heuristics terms count determined, yielding up 32-fold improvement write rate compared default. corresponds 85 GB/s peak bandwidth 202.5 GB/s. The shown be applicable small test program well complex application.