Reducing execution waste in priority scheduling: A hybrid approach

Abstract

Guaranteeing quality for differentiated services while ensuring resource efficiency is an important and yet challenging problem in large computing clusters. Priority scheduling is commonly adopted in production systems to minimize the response time of high-priority workload by means of preempting the execution of low-priority workload when faced with limited resources. As a result, the system performance may not only suffer from the long queueing time of low-priority workload due to resource starvation, but also non-negligible execution waste owing to repetitive evictions. In this paper, we propose a scheduler, HYBRID, which allows the scheduler to switch between being preemptive and non-preemptive by providing a fixed number of computing resources - sticky slots providing uninterruptible task executions. In addition, to preserve performance advantages of high-priority workload by conventional preemptive priority scheduling, HYBRID also aims to reduce repetitive evictions, response times, and wasted executions caused by the low-priority workload. Trace driven simulation analysis shows that our proposed HYBRID scheduler outperforms conventional preemptive priority scheduling by improving response time of low-priority workload by 15% and reducing wasted executions by 85%.