Dynamic Data Replication for Tolerating Single Node Failures in Shared Virtual Memory Clusters of Workstations

In this paper we investigate how shared memory clusters can take advantage of replication to tolerate single system failures. We start from a shared virtual memory protocol (GeNIMA) that has been optimized for low-latency, highbandwidth system area networks. We propose a set of extensions that maintain shared data consistent in the presence of failures and support SMP nodes. Our scheme uses dynamic data replication to guarantee that no shared data is lost when a failure occurs. A failing node is removed from the system and the rest of the nodes recover dynamically and can continue with application execution. We deal both with data consistency and lock synchronization issues. Our approach leverages the low initiation overhead operations provided by modern system area networks as well as the availability of network bandwidth to guarantee data consistency in the presence of failures, and the low-latency operations for dealing with lock synchronization issues. We have implemented the proposed scheme on a cluster of 32, Intel-based dual-processor systems interconnected with a Myrinet network. We are currently evaluating the performance implications of our protocol extensions.