Prioritized multiprocessor networks: design and performance

Abstract

This paper proposes and evaluates prioritized direct shared-memory multiprocessor networks. We use three components to implement prioritized networks, namely, priority-based link arbitration, priority inheritance, and dynamic virtual channels. The two major results from our study are: (i) adding priorities to direct shared-memory multiprocessor networks can lead to reduced average transaction latencies and increased system throughput when running traditional parallel applications, and (ii) a prioritized multiprocessor network can be used to reduce the worst-case latencies of time-constrained traffic when it co-exists with best-effort traffic, without penalizing the average performance of best-effort traffic.