Output Driven Distributed Optimistic Message Logging and Checkpointing

Although optimistic fault tolerance methods using message logging and checkpointing have the potential to provide highly e cient transparent fault tolerance in distributed systems existing methods are limited by several factors Coordinating the asynchronous message logging progress among all processes of the system may cause signi cant over head limiting their ability to scale to large systems and o setting some of the per formance gains over simpler pessimistic methods Furthermore logging all messages received by each process may place a substantial load on the network and le server in systems with high communication rates Finally existing methods do not support nondeterministic process execution such as occurs in multithreaded processes and those that handle asynchronous interrupts This paper presents a new method using optimistic message logging and checkpointing that addresses these limitations Any fault tolerance method must delay output from the system to the outside world until it can guarantee that no future failure can force the system to roll back to a state before the output was sent With this new method only this need to commit output forces any process to log received messages or to checkpoint Each process commits its own output with the cooperation of the minimum number of other processes and any messages not needed to allow pending output to be committed need not be logged Individual processes may also dynamically switch to checkpointing without message logging to avoid the expense of logging a large number of messages or to support their own nondeterministic execution