Armus: dynamic deadlock verification for barriers

This paper presents a graph-based dynamic verification algorithm for deadlock detection and avoidance specialised in barrier synchronisation. Barriers are used to coordinate the execution of groups of tasks, and serve as a building block of parallel computing. The synchronisation patterns enabled by current barrier-based abstractions can introduce deadlocks, a major issue in getting parallel applications correct. Barrier deadlocks arise from a cyclic-dependency amongst tasks that participate on multiple barriers. We propose the use of a synchronisation-centric model instead a task-centric model, traditionally used in dynamic deadlock verification. We present Armus, a platform-agnostic framework for deadlock verification and introduce two applications of this framework: Armus-X10 monitors any unchanged X10 program for deadlocks; JArmus is a library to verify Java programs. To evaluate Armus, we benchmark the runtime execution against the NAS Parallel Benchmark suite and Java Grade Forum Benchmark suite. Results show that the performance overhead of our deadlock detection technique is negligible for usual parallel applications and is independent of the number of the tasks.