REDAC: Distributed, Asynchronous Redundancy in Shared Memory Servers

The emergence of multi-core architectures—driven by continued technology scaling—has led to concerns about increasing soft- and hard-error rates in commodity designs. Because modern chip designs consist of multiple high-speed clock domains, conventional lockstepped redundant execution is no longer practical. Recent work suggests an asynchronous approach to redundant execution, where processor pairs independently execute an instruction stream and treat any differences like soft errors, invoking rollback recovery. Because prior designs buffer instruction results within the out-of-order instruction window, they are limited to tightly coupled redundancy within a single chip, which limits availability and serviceability in the presence of hard errors. We propose REDAC, a set of lightweight mechanisms for distributed, asynchronous redundancy within a sharedmemory multiprocessor. REDAC provides scalable buffering for unchecked state updates, permitting the distribution of redundant execution across multiple nodes of a scalable shared-memory server. The REDAC mechanisms achieve high performance by enabling speculation across common serializing instructions and mitigating the effects of input incoherence. We evaluate REDAC using cycle-accurate fullsystem simulation of common enterprise workloads and show that performance overheads average just 10% when compared to a non-redundant system. These results are comparable to the performance of a similarly configured lockstep design, but offer the substantial benefits of asynchronous redundancy.