A Distributed Directory Cache Coherence Scheme and Its Eeects on Network Performance List of Figures Figure Page 1 System of 8 Processors Constructed Using 2x2 Switches in The

This paper proposes a new protocol for maintaining cache coherence in a large multiprocessor system. The study focuses on systems with private caching that use a multistage network to interconnect the processors. The protocol, called the Distributed Directory Scheme, distributes part of the directory information into the interconnection network switches. The redistribution of the directory information allows the delivery of coherency commands to be streamlined, thus eliminating a bottleneck inherent in the traditional Full Vector Scheme proposed by Censier and Feautrier. When evaluating the coherence protocol performance, the operations of both the cache coherence protocol and interconnection network are considered simultaneously. It is discovered that both the Distributed Directory Scheme and the Full Vector Scheme cause undesirable traac patterns (hot spots) inside the network. However, the hot spots that arise from the coherence traac in the Distributed Directory Scheme respond quite well to (simpliied) network combining. It is found that the Distributed Directory Scheme access latencies are as much as 34% improved over the Full Vector Scheme access latencies. 5 Diierence in read access latency between the Distributed Directory and Full Vector Scheme without combining for write probability of 0. 6 Non-uniform percentage of the Forward network traac created by the Acknowledgement messages sent by the processors to the memory 7 Average message delay in the Forward network for the Full Vector Scheme for write probabilities 0.2, 0.1, 0.05, and 0. 8 Diierence in message delay in the Forward network for the Full Vector Scheme (switch size = 4). 10 Read access latencies with combining for write probability = 0.1 and 0. 11 Percent decrease in read access latency in Full Vector and Distributed Directory Schemes due to the combing for write probability of 0. 12 Diierence in read access latency between the Distributed Directory and Full Vector Scheme with combining for write probability of 0.