TRUSS: A Reliable, Scalable Server Architecture

Traditional reliable servers require costly design changes to the processor, use custom system or application software, or cannot scale beyond a few processing elements. We present TRUSS, a family of server architectures providing reliable, scalable computation from distributed shared-memory hardware while requiring no changes to software. The TRUSS paradigm centers around a logical division of computation and memory that isolates errors in processing from memory storage and vice versa. In this paper, we present the key mechanisms that enable this separation and use full-system simulation to evaluate the impact on a range of commercial and scientific workloads.