Section: New Results

Correct and efficient runtime systems

Participants : Nhat Minh Lê, Robin Morisset, Adrien Guatto, Albert Cohen.

Complementing our different compilation efforts for synchronous and task-parallel data-flow languages, we studied the implementation of Kahn process networks, a deterministic parallel programming model, on shared memory multiprocessors. This model is based on a familiar abstraction: blocking communication through bounded, in-order, single-producer single-consumer queues.

We proposed two novel algorithms that construct such blocking queues on top of concurrent ring buffers and user-land scheduling components. We implemented our algorithms in C11, taking advantage of the relaxed memory model of the language, and prove the correctness of this implementation.

We used these algorithms in a complete runtime system for Kahn process networks with applications ranging from linear algebra kernels to stream computing. In particular, our implementations of the Cholesky and LU factorizations outperform state-of-the-art parallel linear algebra libraries on commodity x86 hardware.