Abstract

In shared-memory multiprocessors sequential consistency offers a natural tradeoff between the flexibility afforded to the implementor and the complexity of the programmer's view of the memory. Sequential consistency requires that some interleaving of the local temporal orders of read/write events at different processors be a trace of serial memory. We develop a systematic methodology for proving sequential consistency for memory systems with three parameters -- number of processors, number of memory locations, and number of data values. From the definition of sequential consistency it suffices to construct a non-interfering observer that watches and reorders read/write events so that a trace of serial memory is obtained. While in general such an observer must be unbounded even for fixed values of the parameters -- checking sequential consistency is undecidable! -- we show that for two paradigmatic protocol classes -- lazy caching and snoopy cache coherence -- there exist finite-state...

Citations