Fine Grain Distributed Implementation of a Dataflow Language with Provable Performances

Efficient execution of multithreaded iterative numerical computations requires to carefully take into account data dependencies. This paper presents an original way to express and schedule general dataflow multithreaded computations. We propose a distributed dataflow stack implementation which efficiently supports work stealing and achieves provable performances on heterogeneous grids. It exhibits properties such as non-blocking local stack accesses and generation at runtime of optimized one-sided data communications.