Efficient rollback-recovery technique in distributed computing systems

We propose an approach for implementing rollback recovery in a distributed computing system. A concept of logical ring is introduced for the maintenance of information required for consistent recovery from a system crash. Message processing order of a process is kept by all other processes on its logical ring. Transmission of data messages are accompanied by the circulation of the associated order messages on the ring. The sizes of the order messages are small. In addition, redundant transmission of order information is avoided, thereby reducing the communication overhead incurred during failure free operation. Furthermore, updating of the order information and garbage collection task are simplified in the proposed mechanism. Our approach does not require information about message processing order be written to stable storage; in fact, the time consuming operations of saving information in stable storage are confined to the checkpointing activities. When failures occur, a surviving process need roll back only if some preceding order information is totally lost, which is relatively unlikely considering the ever growing speed of communication networks. It is shown that a system can recover correctly as long as there exists at least one surviving process