Probability Distribution Based Resource Management for Transient Resource

Abstract

Co-locating latency-critical jobs and batch jobs is an effective way to increase utilization and reduce costs in modern datacenters. The idle resources of latency-critical jobs can be used to run batch jobs, but these transient resources must be revoked whenever latency-critical jobs require them again for guaranteeing their strict quality of service (QoS). The main challenge of this approach is interference between colocated jobs on transient resources. We propose a transient resource management scheme based on probability distribution. Our resource management scheme consist of two components: Tail Latency Aware Allocation (TLAA) and Cost-Effective Resource Revocation (CERR). We character the resource consumption and the completion time of the task of batch jobs and get the probability distribution function of them. Based on the probability distribution of the task resource consumption, we design TLAA to use transient resources efficiently while ensuring tail delay requirements. Based on the probability distribution of the task completion time, we design CERR to revoke transient resources with minimal eviction costs, thereby reducing the impact of resource revocation on batch tasks. Experimental results demonstrate the efficiency of our resource management. TLAA can improve the utilization of transient resources and reduce the completion time of batch jobs to varying degrees under different tail delay requirements. CERR can reduce the impact of resource revocation on batch jobs, and perform better than existing resource revocation strategies.