Over the past few years, parallel sparse direct solvers made significant progress and are now able to solve efficiently industrial three-dimensional problems with several millions of unknowns. To solve efficiently these problems, PaStiX and WSMP solvers for example, provide an hybrid MPI-thread implementation well suited for SMP nodes or multi-core architectures. It enables to drastically reduce the memory overhead of the factorization and improve the scalability of the algorithms. However, today’s modern architectures introduce new hierarchical memory accesses that are not handle in these solvers. We present in this paper three improvements on PaStiX solver to improve the performance on modern architectures : memory allocation, communication overlap and dynamic scheduling and some results on numerical test cases will be presented to prove the efficiency of the approach on NUMA architectures.