Defeating network jitter for virtual machines

Abstract

Virtualization based cloud computing hosts networked applications in virtual machines (VMs), and provides each VM the desired degree of performance isolation using resource isolation mechanisms. Existing isolation solutions address heavily on resource proportionality such as CPU, memory and I/O bandwidth, but seldom focus on resource provisioning rate. Even the VM is allocated with adequate resources, if they can not be provided in a timely manner, problems such as network jitter will be very serious and significantly affect the performance of cloud applications like internet audio/video streaming. This paper systematically analyzes and illustrates the causes of unpredictable network latency in virtualized execution environments. We decouple the design goals of resource proportionality from resource provisioning rate, and adopt divide-and-conquer strategy to defeat network jitter for VMs: (1) in VMM CPU scheduling, we differentiate self-initiated I/O from event-triggered I/O, and individually map them to periodic and aperiodic real-time domains to schedule them together; (2) in network traffic shaping of VMs, we introduce the concept of smooth window to smooth network latency and apply closed-loop feedback control to maintain network resource consumption. We implement our solutions in Xen 4.1.0 and Linux 2.6.32.13. The experimental results with both real-life applications and low-level benchmarks show that our solutions can significantly reduce network jitter, and meanwhile effectively maintain resource proportionality.