A COMPARATIVE STUDY OF COMMUNICATION-INDUCED CHECKPOINTING PROTOCOLS

Abstract

This dissertation presents a comparative study of two Communication-Induced Checkpointing Protocols with Rollback-Dependency Trackability. Rollback-Dependency Trackability (RDT) is a property that states that all rollback dependencies between local checkpoints are on-line trackable by using a transitive dependency vector. These protocols are proposed for improving the fault-tolerance in distributed systems. Theoretically it can be proved that the common intuition "If a protocol forces a checkpoint only at a stronger condition, then it should take less number of forced checkpoints than a protocol based on a weaker condition", is false. This has already been proved by taking a few counterexamples [1]. Our simulation study implies that such examples are very rare and for most of the time the intuition holds true. Thus, a protocol taking checkpoints at stronger condition outperforms the protocol taking checkpoints at weaker condition as the checkpointing overhead in former protocol is less.