Network performance isolation for latency-sensitive cloud applications

摘要

Network performance isolation is the key to virtualization-based cloud services. For latency-sensitive cloud applications like media streaming, both predictable network bandwidth and low-jittered network latency are desirable. The current resource sharing methods for virtual machines (VMs) mainly focus on resource proportional sharing such as CPU amount, memory size and I/O bandwidth, whereas they ignore the fact that I/O latency in VM-hosted platforms is mostly related to the resource provisioning rate. Even if the VM is allocated with adequate resources, network jitter can still be very serious if the resources are not provided in a timely manner. This paper systematically analyzes the causes of unpredictable network latency and proposes a combined solution to guarantee network performance isolation: (1) in the hypervisor, we design a proportional share CPU scheduling with soft real-time support to reduce scheduling delay for network packets; (2) in network traffic shaper, we introduce the concept of smooth window with feedback control to smooth the packet delay. 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.