COLO: COarse-grained LOck-stepping virtual machines for non-stop service
Proceedings of the 4th annual Symposium on Cloud Computing
| Hi-index | 0.00 |
There is an increasing need for fault tolerance capabilities in logic devices brought about by the scaling of transistors to ever smaller geometries. This paper presents a hypervisor-based replication approach that can be applied to commodity hardware to allow for virtually lockstepped execution. It offers many of the benefits of hardware-based lockstep while being cheaper and easier to implement and more flexible in the configurations supported. A novel form of processor state fingerprinting is also presented, which can significantly reduce the fault detection latency. This further improves reliability by triggering rollback recovery before errors are recorded to a checkpoint. The mechanisms are validated using a full prototype and the benchmarks considered indicate an average performance overhead of approximately 14 percent with the possibility for significant optimization. Finally, a unique method of using virtual lockstep for fault injection testing is presented and used to show that significant detection latency reduction is achievable by comparing only a small amount of data across replicas.