Parallel Job Scheduling Policies to Improve Fairness: A Case Study

Balancing fairness, user performance, and system performance is a critical concern when developing and installing parallel schedulers. Sandia uses a customized scheduler to manage many of their parallel machines. A primary function of the scheduler is to ensure that the machines have good utilization and that users are treated in a ``fair'' manner. A separate compute process allocator (CPA) ensures that the jobs on the machines are not too fragmented in order to maximize throughput. Until recently, there has been no established technique to measure the fairness of parallel job schedulers. This paper introduces a ``hybrid'' fairness metric that is similar to recently proposed metrics. The metric uses the Sandia version of a ``fairshare'' queuing priority as the basis for fairness. The hybrid fairness metric is used to evaluate a Sandia workload. Using these results, multiple scheduling strategies are introduced to improve performance while satisfying user and system constraints.