Verifying IO Synchronization from MPI Traces

The paper addresses the following question: Are IO operations of HPC applications properly synchronized? We focus on parallel file systems that satisfy POSIX semantics. The outcome of I/O operations is well-defined provided that conflicting accesses to a file location are not concurrent, but are ordered. Accesses to distinct processes are ordered by the executed MPI communication. We derive the "happens-before" relation between I/O calls of HPC runs by analyzing traces collected during program execution. Various optimizations reduce the analysis overhead. We collected traces from 17 representative HPC applications. We found that 10 of them do not perform conflicting I/O accesses and, hence, are properly synchronized by default. The remaining 7 applications properly synchronize the conflicting I/O accesses.