Finding Deadlocks in Cache-Coherent Distributed Shared Memory Machines

The design of a cache-coherent distributed shared memory (CCDSM) system is complex and prone to errors due to its concurrent, distributed nature. We examine the problem of deadlocks due to resource sharing between different memory requests that are occuring concurrrently, an issue which has not been dealt with extensively in the literature. We discuss a tool for illuminating all possible deadlock scenarios due to resource sharing. 1 Introduction A CCDSM system supports cache-coherent global shared memory on distributed-memory multiprocessor hardware. Coherence is maintained in a non-broadcast manner across sites, typically using a directory-based approach. Many research[6, 5, 1, 4] and a few commerical[3, 2] CCDSM systems have been built or are being built. Figure 1 shows a typical CCDSM implemention, consisting of processor sites and memory sites connected by a scalable interconnection network. Each processor site has one or more cache-equiped processors connected by a snoopy bus. ..