Fair Resource Allocation for Running HPC Workloads Simultaneously

Abstract

In high performance computing (HPC) job schedulers usually divide resources of computing nodes into slots. Each slot can be assigned to execute only a single job from the queue. In some cases, jobs do not fully utilize all available resources from the slot which leads to internal fragmentation, wasted resources and to an increase of queue wait time. In this paper, we propose fair resource allocation strategies that can be applied in job schedulers for resource allocation. We cover such resources as CPU time, residential memory and network bandwidth.