SCOUT: Scheduling Core Utilization to Optimize the Performance of Scientific Computing Applications on CPU/Coprocessor-Based Cluster

Today’s scientific computing applications require many different kinds of task and computational resource. The success of scientific computing hinges on the development of High Performance Computing (HPC) system in the role of decreasing execution time. Remarkably, the support is more enhanced with the advent of accelerators like Graphics Processing Unit (GPU) or Intel Xeon Phi (MIC) coprocessor. However, problems related to coprocessor underutilization of MIC can lead to the thread and memory over-subscription. Based on logging the runtime behaviors of scientific applications, scheduling jobs usually has constraints on the completion time of jobs as deadline or due date assignment. These problems can be solved to improve the performance by a suitable method such as scheduling or assigning priorities to job submission. In this paper, we propose a scheduling module named SCOUT by exploiting factors from the view of the application’s performance to improve the scheduler on a CPU/Coprocessor-based cluster. SCOUT focuses on the performance of applications as well as reducing their execution time on Xeon Phi accelerator. Furthermore, our scheduling module decides the order of job execution to increase the throughput and minimize the delay time. Given a set of popular scientific applications, the experimental results show that the performance and throughput of SCOUT are better than others compared policies. Especially, we implement the entire module as a seamless plug-in to an HPC workload manager named PBS Professional and show the efficiency in practice.