Data-driven job dispatching in HPC systems

As High Performance Computing (HPC) systems get closer to exascale performance, job dispatching strategies become critical for keeping system utilization high while keeping waiting times low for jobs competing for HPC system resources. In this paper, we take a data-driven approach and investigate whether better dispatching decisions can be made by transforming the log data produced by an HPC system into useful knowledge about its workload. In particular, we focus on job duration, develop a data-driven approach to job duration prediction, and analyze the effect of different prediction approaches in making dispatching decisions using a real workload dataset collected from Eurora, a hybrid HPC system. Experiments on various dispatching methods show promising results