Introduction of shared-memory parallelism in a distributed-memory multifrontal solver

We study the adaptation of a parallel distributed-memory solver towards a shared-memory code, targeting multi-core architectures. The advantage of adapting the code over a new design is to fully benefit from its numerical kernels, range of functionalities and internal features. Although the studied code is a direct solver for sparse systems of linear equations, the approaches described in this paper are general and could be useful to a wide range of applications. We show how existing parallel algorithms can be adapted to an OpenMP environment while, at the same time, also relying on third-party optimized multithreaded libraries. We propose simple approaches to take advantage of NUMA architectures, and original optimizations to limit thread synchronization costs. For each point, the performance gains are analyzed in detail on test problems from various application areas.