A Method for Job Management Using MPI Profiling Interface

In High-Performance Computing (HPC), there is a requirement to run both short-term jobs and long-term jobs. For meeting this requirement, job switching is effective. Slurm, a job scheduler commonly used in HPC, can switch jobs with a gang scheduling function. However, gang scheduling is implemented by sending signals to all processes in a job at fixed intervals. This is because Slurm does not assume frequent job switching. In this paper, we propose a system that uses the Message Passing Interface (MPI) profiling interface to notify the properties and running status of jobs. This system allows estimating the overhead associated with job switching and determining its timing. To assess the usefulness of our system, we evaluated the overhead caused by job switching. The evaluation results showed a 12.8% reduction in overhead when switching when considering of the job’s running status, compared to switching without consideration.