Node Sharing for Increased Throughput and Shorter Runtimes : an Industrial Co-Scheduling Case Study

The allocation of jobs to nodes and cores in industrial clusters is often based on queue-system standard settings, guesses or perceived fairness between different users and projects. Unfortunately, hard empirical data is often lacking and jobs are scheduled and co-scheduled for no apparent reason. In this case-study, we evaluate the performance impact of co-scheduling jobs using three types of applications and an existing 450+ node cluster at a company doing large-scale parallel industrial simulations. We measure the speedup when co-scheduling two applications together, sharing two nodes, compared to running the applications on separate nodes. Our results and analyses show that by enabling co-scheduling we improve performance in the order of 20% both in throughput and in execution times, and improve the execution times even more if the cluster is running with low utilization. We also find that a simple reconfiguration of the number of threads used in one of the applications can lead to a performance increase of 35-48% showing that there is a potentially large performance increase to gain by changing current practice in industry