Section: New Results

Distributed StarPU on top of a High Performance Communication Library

A new implementation of the StarPU's distributed engine is being currently developed on top of the NewMadeleine library. The first version of this engine had been written directly on top of MPI. The performance were not as good as expected when dealing with applications exchanging huge number of messages, and we had to implement within StarPU mechanisms to control the memory subscription  [14].

NewMadeleine is a high performance communication library for clusters developed in the Tadaam team. It applies optimization strategy on data flows through dynamic packet scheduling, and is usable on various high performance networks. The new implementation of the StarPU's distributed engine no longer has to deal with communication-related issues, and provides a better reactivity as the communications progress is dealt with by NewMadeleine itself. First experiments with the Chameleon solver show promising results.