Section: New Results

Thread placement and memory allocation on NUMA machines

Participant : Emmanuel Jeannot.

We have worked on optimizing the tiled Cholesky factorization on NUMA machine. We have designed a new symbolic technnique for allocating task and tiles at the same time called SMA (Symbolic Mapping and Allocation). SMA provide an optimal allocation in terms of point-to-point communication for the Cholesky factorization. We have studied some performance issues regarding the way threads are grouped and tiles are allocated in the memory. We have shown how to optimize thread placement and data placement in order to achieve performance gain up to 50% compared to state-of-the-art libraries such as Plasma or MKL. This work has been published in PAAP 2012 [25] .