Group communication for upgrading distributed programs

Abstract

Recently, large-scale distributed systems are being developed. Since it is expensive to newly develop the system, it is required to adopt the system in use to the changes of the user requirements and the environments. Therefore, it is essential to discuss how to upgrade the distributed programs. The system cannot be kept highly available by the conventional upgrading methods because multiple processes have to be suspended simultaneously. This paper discusses a new method which allows each process to invoke the upgrading procedure independently of the other processes. The key idea is that multiple versions of processes are allowed to co-exist temporarily. If the multiple versions of processes co-exist, such protocol errors as the unspecified receptions and the communication deadlocks can occur. Each pair of an old-version process and a new-version one is named a process group. The group communication protocol proposed in this paper supports message transmissions among the process groups. Moreover, the protocol errors can be detected and resolved by using checkpointing, timeout and rollback recovery.