High Availability in Cyber-physical Systems by Self-determined Virtual Machine Replication

摘要

Present approaches to high availability via virtual machine replication either rely on parallel VM-execution combined with a voting mechanism or periodically transfer state modifications to an inactive backup VM. While redundant execution is expensive in terms of CPU load, periodic checkpointing leads to higher network load and the response times of protected VMs increase and get more jitter. Both alternatives are unacceptable in domains with resource constraints and realtime requirements such as cyber-physical systems. In this work we present a “self-determined” virtual machine replication model for high availability solutions that avoids the drawbacks of both established techniques at the cost of transparency. We argue that for some emerging application domains, such as cyber-physical systems, our model is more suitable. Results show that high availability via self-determined replication can reduce response latencies by an order of magnitude when compared with its periodic counterpart. Especially when used in combination with a lightweight unikernel as guest OS, the overhead of high availability becomes acceptable, even for resource-constrained cyber-physical systems.