Analysis of avalanche's shared memory architecture

In this paper, we describe the design of the Avalanche multiprocessor's shared memory subsystem, evaluate its performance, and discuss problems associated with using commodity workstations and network interconnects as the building blocks of a scalable shared memory multiprocessor. Compared to other scalable shared memory architectures, Avalanchehas a number of novel features including its support for the Simple COMA memory architecture and its support for multiple coherency protocols (migratory, delayed write update, and (soon) write invalidate). We describe the performance implications of Avalanche's architecture, the impact of various novel low-level design options, and describe a number of interesting phenomena we encountered while developing a scalable multiprocessor built on the HP PA-RISC platform.