Levenberg-Marquardt sparse solver scaling

Notes on using Strumpack within CCTBX

Here are the documented results of using Strumpack on a single node for a variety of data set sizes (StrumpackSolverMPI_1K,StrumpackSolverMPI_5K,StrumpackSolverMPI_10K). All tests were performed on dials.lbl.gov, and allow the tests to be repeated at the user’s discretion. Example matrices for a variety of different refinement parameters are listed in the given paths, and the times represent a single solution.

Setting up and running STRUMPACK

To build STRUMPACK alongside a conda cctbx.xfel build, follow the instructions given here/here with the following additional conda packages before building cctbx:

conda install -y IPython mysql-python matplotlib scipy mpi4py jupyter ipyparallel;

Building is now carried out as normal:

wget https://raw.githubusercontent.com/cctbx/cctbx_project/master/libtbx/auto_build/bootstrap.py;
python bootstrap.py hot update --builder=xfel --sfuser=<USERNAME> --cciuser=<USERNAME>;
python bootstrap.py build --builder=xfel --with-python=`which python` --nproc=<NUM_CORES>;
source ./build/setpaths.sh;

The MPI-enabled work requires the cctbx_project git repository be checked-out into the strumpack_solver_backend branch (eventually all changes will be merged with upstream master).

cd modules/cctbx_project
git checkout strumpack_solver_backend
cd ../..

We can now proceed with building STRUMPACK, and it’s required dependencies. The script located at github.com/exafel/exafel_project/master/strumpack/STRUMPACK_installer_shared.sh will acquire all dependencies and build the libraries against the conda MPI installation, or build OpenMPI if this is not available (such as may be true under MacOS).

wget https://raw.githubusercontent.com/ExaFEL/exafel_project/master/strumpack/STRUMPACK_installer_shared.sh
chmod +x STRUMPACK_installer_shared.sh

The STRUMPACK (and dependencies) binaries, libraries and headers will be installed into strumpack_build/{bin,lib,include}, of which will be added to the dispatcher environment given successful completion of the installation script. With the presence of the new libraries, refreshing the dispatcher and rebuilding the packages will allow any STRUMPACK-enabled Boost.Python extension modules to be built.

cd build && libtbx.refresh

To test the new libraries several test scripts are available, for a provided A matrix and b vector solution. The solver will compare both the original Eigen-based solver and the new STRUMPACK solvers for both the OpenMP backend (cctbx_project/scitbx/examples/bevington/strumpack_eigen_solver.py) and the distributed MPI-enabled solver (cctbx_project/scitbx/examples/bevington/strumpack_eigen_solver_mpi_dist.py).

Sample run commands are given in the provided notebooks.

Full integration with Samosa will be provided shortly to allow the solvers to be used therein. Initial work to provide selective choice of the Eigen solver backend is available in the eigen_solver_algo branch of https://github.com/cctbx/cctbx_project/.

Scalability tests on Cori

Scalability testing of the OpenMP and MPI backends are available in Jupyter notebook StrumpackSolverMPI_dist_Cori.ipynb

Experimental spack build

An experimental set of commands to build STRUMPACK using spack is given here. This is not supported, and not guaranteed to work. It was used as an example environment to build STRUMPACK dependencies.

Lee J. O'Riordan
Senior Quantum Software Developer I PennyLane Performance Team Lead @ Xanadu