Exploration of Dynamic Job Scheduling with Network Simulations

Job scheduling policies are known to have a significant impact on the throughput of jobs running on supercomputers. With the widescale adoption of low-diameter networks, inter-job interference has become prevalent and affects the performance of individual applications, which in turn impacts the system throughput. As a result, an accurate analysis and comparison of various job scheduling policies should also account for the variations in runtimes of individual jobs due to network interference. To this end, we have added the ability to simulate dynamic job scheduling in a network simulation tool called TraceR. In this talk, I will give a brief overview of TraceR and then describe the implementation of the capability to simulate dynamic job scheduling using TraceR.