Reduced Overhead Logging for Rollback Recovery in Distributed Shared Memory

Rollback techniques that use message logging and deterministic replay can be used in parallel systems to recover a failed node without involving other nodes. Distributed shared memory (DSM) systems cannot directly apply message-passing logging techniques because they use inherently nondeterministic asynchronous communication. This paper presents new logging schemes that reduce the typically high overhead for logging in DSM. Our algorithm for sequentially consistent systems tracks rather than logs accesses to shared memory. In an extension of this method to lazy release consistency, the per-access overhead of tracking has been completely eliminated. Measurements with parallel applications show a significant reduction in failure-free overhead. 1 Introduction Distributed shared memory (DSM) provides the programming advantages of a shared memory image in a system with physically distributed processing nodes. DSM maintains consistency between processing nodes in software, using the virtual..