Afterlife: A Distributed and Recoverable File System Based on a Unified Approach to Logging

We present a distributed and recoverable file system called Afterlife that only requires a single replicated log to recover from failures in both the file system layer and the storage layer. Based on this unified log, we devise recovery procedures for both file server and block server failures that can run as background processes while the file system continues serving incoming requests. Batch flushing of log records mitigates the performance impact of logging. Our design allows the batch flushing to be integrated with the file server block cache consistency scheme and lazy lock release scheme. Multiple identical block server replicas enable high performance file system read operations and provide availability in the face of failure. We evaluate their performance on our Afterlife prototype of three block servers and two file servers, and show that logging and data replication incurs only a factor of 10 performance versus a conventional NFS file server.