Improving Network I/O Performance in Virtual Machines through Workload Profile Clustering

Abstract

The unpredictability of virtual machines workload makes effective CPU resources allocation a hard assignment for the hypervisor. Even if a virtual machine is allocated with adequate CPU resources, the quality of service of those running both CPU and I/O intensive tasks (heterogeneous workload), may be seriously affected if they are not provided in a timely manner. As server consolidation grows, CPU sharing among multiple virtual machines leads to negative impact on I/O intensive tasks running within the guest OS, due to incurred scheduling latency and lack of prioritization by hypervisor's scheduler. Although a fair amount of researches have dedicated to improve I/O performance on a multi-core platform through the implementation of coexisting schedulers, heterogeneous workloads still lacks of in-depth exploration. In this paper, we systematically evaluate scheduler's interference on I/O responsiveness of heterogeneous workloads domains under server consolidation. The performance study aims to create a synthetic heterogeneous workload based on media streaming applications by varying packet sizes. Furthermore, we apply the coexisting schedulers approach in order to analyze the linearity of quality of service metrics such as throughput, jitter and packet loss over this specific scenario.