Fault tolerance and configurability in DSM coherence protocols

With the advent of large networks and the demand to have uninterrupted service, computer systems need to be more robust and fault tolerant. There are numerous ways to implement fault tolerance and recovery. A central concept in all these methods is the requirement for replicated data for high data availability. We believe that a protocol must not only provide replication, but do so at low operation overhead. Further, the protocol must provide configurable mechanisms for varying the level of replication, so that the system may be operated at the desired overhead cost. We have developed several Distributed Shared Memory (DSM) protocols and use these with a program-driven simulation to examine the robustness, fault tolerance, and configurability of these. Our investigation compares the Write-Invalidate, Write-Invalidate with Downgrading, Write-Broadcast and several instances of the Boundary-Restricted coherence protocol class. The DSM application suite contains programs representative of various memory-access patterns and behaviors. This paper examines the performance of these protocols under different workloads and analyzes the operation costs, fault tolerance, and configurability of each.