VCPU Prioritization Interface for Improving the Performance of Latency-Critical Tasks

摘要

Virtualization supports multi-tenancy in the form of virtual machine (VM) and tends to oversubscribe hardware resources to improve resource utilization. However, this server consolidation comes at performance degradation, especially for latency-critical tasks. In such environments, a physical CPU (PCPU) is often time-shared by multiple virtual CPUs (VCPUs), and this time-sharing of PCPU limits guest OS-level supports for latency-critical (LC) tasks. To address this problem, this paper presents VCPU prioritization interface (VPI) that directly prioritizes the VCPU executing an LC task over other VCPUs hosting background ones. Our evaluation on real-world applications shows that VPI significantly reduces the average and 99th percentile latency of LC tasks by up to 66.6% and 60.5%, respectively.