Section: New Results

Optimization of the Sparse Direct Linear Solver PaStiX

Participant : Mathieu Faverge.

In the context of distributed NUMA architectures, a work has recently begun, in collaboration with the INRIA ScAlApplix team-project, on studying optimization strategies and improving the scheduling of communications, threads and I/O on the sparse direct linear solver PaStiX . The solver provides the NewMadeleine and Marcel libraries with an advanced application to validate those strategies. Mathieu Faverge studies these aspects in the context of the NUMASIS ANR project. It has been proved that NUMA allocation can significantly improve the efficiency. In the out-of-core context, new problems related to the scheduling and the management of the computational tasks may arise (processors may be slowed down by I/O operations). Thus, specific algorithms for this particular context have to be designed and implemented.