Extending the scalable coherent interface for large-scale shared-memory multiprocessors

摘要: Massively parallel machines promise to provide enormous computing power using an amalgamation of low-cost parts. We believe many these will be shared-memory machines, since they do not burden the programmer with data placement and nonuniform access semantics. However, efficient kiloprocessor solution for paradigm has proven elusive due bottlenecks associated accesses rapidly changing data. The Scalable Coherent Interface (SCI) is IEEE ANSI standard multiprocessors, specifying a topology-independent network cache-coherence protocol. The goal this dissertation investigate ways efficiently share frequently among thousands processors. SCI platform in which methods are investigated. Before investigating protocols, we demonstrate that arbitrary topology can constructed from set interwoven rings, such as rings. This result important because it would impossible realize performance advantages our new protocols without network. Our investigation rings leads scheme deadlock avoidance does require resource partitioning. also compare various topologies. Next, two employ trees cache lines. first adds shortcuts, called temporary pointers, list created by SCI. several protocol variations request combining temporary-pointer structures. give lower bounds on latency numerous sets assumptions against variations. second transforms into probabilistically balanced, binary tree. Again, variations, comparing derive, extend support fetch-and-add severe constraints other mechanisms. show both compatible networks correctly interoperate SCI's protocol. We have small increase traffic. They similar or than SCI, significantly when there even amount global sharing.