View article

Memory usage is crucial for sparse direct solvers and particularly for the parallel multifrontal scheme. The active memory size depends on the assembly tree associated to the factorization process and on the distribution of the computation; it can be large compared to the factors. We study in details the impact of state-of-the-art sparse matrix reordering techniques on the assembly tree and on the memory occupation of the MUMPS solver. Our main observation is that the active memory of parallel multifrontal solvers does not scale well if dynamic scheduling strategies based only on the balance of the workload is used.