At-most-once Algorithm for Linearizable RPC in Distributed Systems

In a distributed system, like RAMCloud, supporting automatic retry of RPC, it is tricky to achieve linearizability correct along with durability and liveness. Automatic retry may cause duplicate executions when a server fails after durably recording the original execution and before sending its response. We can cope with the problem by implementing at-most-once mechanism for RPC system. But in a massively distributed system, the implementation of the algorithm is not simple; correct recovery of RPC information and garbage collection for the information in cluster level are pretty challenging. In this project, I am presenting solutions to the problems and implementation in a distributed storage system, called RAMCloud.