Formal verification of programs that use MPI one-sided communication

We used formal-verification methods based on model checking to analyze the correctness properties of one existing and two new distributed-locking algorithms implemented by using MPI's one-sided communication. Model checking exposed an overlooked correctness issue with the first algorithm, which had been developed by relying only on manual reasoning. Model checking helped confirm the basic correctness properties of the two new algorithms, while also identifying the remaining problems in them. Our experience is that MPI-based programming, especially the tricky and relatively poorly understood one-sided communication features, stand to gain immensely from model checking. Considering that many other areas of concurrent hardware and software design now routinely employ model checking, our experience confirms that the MPI community can benefit greatly from the use of formal verification.