Computing Science. A Non-Blocking Atomic-Multicast Service for Scalable In-Memory Transaction Systems.

摘要

Atomic Multicasting is central to managing replication. Our focus is on its deployment for enhancing scalability and performance of in-memory transactional systems where data replication is essential. When it is deployed as an external service, enhancement is demonstrably accomplished when no node within the service crashes. Since known atomic-multicast protocols block whenever a node is crashed or appears so, using them for service implementation risks the transaction system making no progress when blocking prevails. To eliminate this risk, our service simultaneously runs two protocols with distinct properties. The first one is the fastest when crashes are absent, but blocks until a crashed node is isolated. The second, newly-designed one never blocks but can fail to deliver messages with a probability that can be made vanishingly small in return for slower performance. The service uses slow protocol when the fast one is blocked and the switch-over is instantaneous. Extensive performance study confirms best crash-free performance, crashuninterrupted service and no delivery failure in practical settings.