Automatic Inference of Task Parallelism in Task-Graph-Based Actor Models

Abstract

Automatic inference of task level parallelism is fundamental for ensuring many kinds of safety and liveness properties of parallel applications. For example, two tasks running in parallel may be involved in data races when they have conflicting memory accesses, or one is affecting the termination of another by updating shared variables. In this paper, we have considered a task-graph-based actor model, used in signal processing applications (e.g., baseband processing in wireless communication, LTE uplink processing) that are deployed on many-core platforms, in which the actors, task-graphs, and tasks are the active entities running in parallel. The actors invoke task graphs, which in turn invoke tasks, and they communicate through message passing, thus creating different kinds of dependencies and parallelism in the application. We introduce a novel May Happen in Parallel (MHP) analysis for complex parallel applications based on our computational model. The MHP analysis consists of (i) data-flow analysis applicable to parallel control-flow structures inferring MHP facts representing pairs of tasks running in parallel, (ii) identification of all direct and indirect communication by generating a context-free grammar and enumerating valid strings representing parallelism and dependencies among active entities, and (iii) inferring MHP facts when multiple task-graphs communicate. Our analysis is applicable to other computational models (e.g. Cilk or X10) too. We have fully implemented our analysis and evaluated it on signal processing applications consisting of a maximum of 36.57 million lines of code representing 232 different tasks. The analysis took approximately 7 minutes to identify all communication information and 10.5 minutes to identify 12052 executable parallel task-pairs (to analyse for concurrency bugs) proving that our analysis is scalable for industrial-sized code-bases.