Do We Need a Crystal Ball for Task Migration?

For communication-intensive applications on distributed memory systems, performance is bounded by remote memory accesses. Task migration is a potential candidate for reducing network traffic in such applications, thereby improving performance. We seek to answer the question: can a runtime profitably predict when it is better to move the task to the data than move the data to the task? Using a simple model where local work is free and data transferred over the network is costly, we show that a best case task migration schedule can achieve up to 3.5x less total data transferred than no migration for some benchmarks. Given this observation, we develop and evaluate two online task migration policies: Stream Predictor, which uses only immediate remote access history, and Hindsight Migrate, which tracks instruction addresses where task migration is predicted to be beneficial. These predictor policies are able to provide benefit over execution with no migration for small or moderate size tasks on our tested applications.