Modeling Multiprocessor Cache Protocol Impact on MPI Performance

This paper presents a modeling method particularly suited to analyze interactions between message passing interface MPI library execution and distributed cache coherence protocol. The method is applied to the Ping-Pong benchmark. In addition to overall performance figures like message exchange latency, it also provides detailed analysis elements such as cache miss counts per variable. It is based on formal modeling where functional aspects and performance aspects are integrated by composition and can be refined independently. A key modeling point is that the cache coherence protocol implies that the duration of an access to a variable is not static but is state-dependent. Our Ping-Pong model allows comparison of different primitive implementations in the context of different cache coherence protocols. We applied this approach using interactive Markov chain theory and its implementation in the CADP toolbox.