Replication and recovery in a distributed file server

Abstract

This project presents a replicated system that takes advantage of the properties provided by Group Communication Systems to support the crash-recovery model. Our main goal is to point out some key issues that must be taken into account when dealing with recovery in a replicated system. As far as the type of replicated service is concerned, special attention has been paid to transactional systems, such as databases [4, 24, 25, 26]. With the aim to study problems that may arise when the operation is not performed inside the boundaries of a transaction, we have focused on non-transactional services. In particular, we have implemented a replicated file server that allows clients to remotely execute basic operations over a structure of directories and files (reads, writes, creations and deletions), as well as to block and unblock files to prevent other clients from accessing them, by managing a lock system. Apart from formalizing the system specification for several representative examples of replication and recovery techniques, we evaluate the cost of replication and compare the performance of active and passive replication protocols for the file server, depending on the workload and rate of update operations. This work also assesses the overhead introduced by the recovery process, analyzing different solutions for total and partial recovery in a variety of reconfiguration settings. We intend to determine the circumstances in which partial recovery performs better than total recovery, and discuss the advantages of a combination of both approaches.