Constrained data-driven parallelism

Abstract

In data-driven parallelism, changes to data spawn new tasks, which may change more data, spawning yet more tasks. Computation propagates until no further changes occur. Benefits include increasing opportunities for fine-grained parallelism, avoiding redundant work, and supporting incremental computations on large data sets. Nonetheless, data-driven parallelism can be problematic. For example, convergence times of data-driven singlesource shortest paths algorithms can vary by two orders of magnitude depending on task execution order. We propose constrained data-driven parallelism, in which programmers can impose ordering constraints on tasks. In particular, we propose new abstractions for defining groups of tasks and constraining the execution order of tasks within each group. We sketch an initial implementation and present experimental results demonstrating that our approach enables new efficient data-driven implementations of a variety of graph algorithms.