pghysels
commented
3 years ago Hi Mike, The sparse direct solver does support distributed memory, relying on MPI and ScaLAPACK. Sierra is a GPU machine, much like summit, correct? We also have GPU support. For a single GPU, we use CUDA and cuBLAS/cuSOLVER. To run on multiple nodes, with multiple GPUs, we use SLATE (https://bitbucket.org/icl/slatehttps://bitbucket.org/icl/slate) as well as CUDA. Here is more info on the installation process: https://portal.nersc.gov/project/sparse/strumpack/master/installation.html
Apart from the sparse direct solver, we also have distributed memory preconditioners, based on approximate factorization. This uses rank-structured matrix approximations, like Hierarchically Semi-Separable, Hierarchically Off-Diagonal Low-Rank, Block Low Rank and Butterfly matrix decompositions. These preconditioners are also aimed at large sparse problems from for instance FEM codes. See for instance here: https://portal.nersc.gov/project/sparse/strumpack/master/HSS_Preconditioning.html
We'd be happy to work with you.
Hi, my name is Mike Puso and I work with a number of finite element code developers at LLNL and are interested in you sparse direct solver in particular. I noticed in your latest paper that you mention you have a version of that but the results were from a shared memory version. I was wondering what the status and availability of distributed memory solver is. We would be very interested in collaborating with you and running it on our latest sierra platform. We currently use PWSMP (leased) from the IBM group of Anshul Gupta, Super LU, MUMPS and Pastix. The PWSMP solver is what we use the most due to its superior performance but we would prefer to have a very good open source direct solver.