When truth is efficient: analysing concurrency

Abstract

Concurrent systems are hard to develop and are even harder to analyse. The usual way to analyse concurrent systems is to give them interleaving semantics and exploit automata-based methods to investigate the resultant interleaved model. Such an approach is often hard to scale without additional tools to curb the interleaving-induced state space explosion. In this work we make an alternate case: for directly capturing the behaviour of concurrent systems using true concurrency. We show how to build composable, truly concurrent models for real-world programs written using one of the most widely adopted paradigms for developing massively parallel systems, the Message Passing Interface Standard (MPI). Our method employs general event structures to symbolically capture executions of MPI programs and uses this truly concurrent model, combined with our novel deadlock characterisation, to formulate a precise, scalable decision procedure that finds communication deadlocks in large MPI programs. We show that our analysis scales to systems with hundreds of processes and strongly outperforms state of the art interleaving semantics based approaches.