Using VM replication to build a consistent, fast HA system

Abstract

Virtualization supplies a straightforward approach to high availability through iterative replications of virtual machine (VM) checkpoints that encapsulate the protected services. Unfortunately, traditional VM replication solutions suffer from deficiencies in either response latency or state recovery consistency, which constrains the adoption of VM replication in production. In this work, we redesign the typical consistency model and network architecture for virtual machine replication. In doing this, we extend the functionality of the secondary host to be the primary recipient of network requests so that the state of the primary VM (PVM) is retained by the secondary host in the form of network packets. Specifically, the secondary host is set for network redirection and packets recording. Should the primary host fail, the recorded packets are used to recreate the state on the secondary host. We name the system reverse replication of virtual machines (rRVM). Experiments demonstrate the simultaneous strong recovery consistency and low response latency of our real-time disaster recovery system. To be specific, we achieved 100% consistency without additional latency imposed.