Cache coherence using local knowledge

• Published in: Proceedings of the 1993 ACM/IEEE conference on Supercomputing pp. 720 - 729
• Main Authors: Darnell, E., Kennedy, K.
• Format: Conference Proceeding
• Language: English
• Published: New York, NY, USA ACM 01.12.1993; IEEE
• Series: ACM Conferences
• Subjects: Cache storage; Computer science; Computer systems organization > Architectures > Parallel architectures; Condition monitoring; Contracts; Costs; Global communication; Hardware; Hardware > Integrated circuits > Semiconductor memory > Dynamic memory; Large-scale systems; Runtime; Scalability; Software and its engineering > Software notations and tools > Compilers
• ISBN: 0818643404; 9780818643408
• ISSN: 1063-9535
• DOI: 10.1145/169627.169821

Typically, commercially available shared memory machines have addressed the cache coherence problem with hardware strategies based on global inter-cache communication. However, global communication limits scalability and efficiency. "Local knowledge" coherence strategies, which avoid global communication at run-time, offer better scalability, at the cost of some additional cache misses. The most effective local knowledge strategies described in the literature are those based on generation time-stamps (TS). The authors propose a new strategy, TS1, that requires less extra storage than TS, only one extra bit per cache line, and can produce more cache hits by exploiting sophisticated compiler analysis. TS1 handles common synchronization paradigms including DOALL, DOACROSS, and critical sections. Early results show TS1 is, worst case, slightly slower than TS. Best case, TS1's flexibility allows for significant improvements.