A Julia Interface to MUMPS

OSX and Linux:

MUMPS is a library for the solution of large linear systems using a factorization. Structure can be exploited, such as symmetry, or symmetry and definiteness. The factorization and solve phases can be performed in parallel.

How to Install

Prerequisites

Currently, MUMPS must be installed outside of Julia. On macOS, we recommend using Homebrew. On Linux, we recommend using Linuxbrew. Please follow the installation instructions of each package manager.

In both cases, the commands to install MUMPS are the same:

$ brew tap brewsci/num
$ brew install brewsci-mumps  # use brew options brewsci-mumps for build options

Note: on Linux, apt-get install libmumps-dev installs a version of OpenMPI that is too old for MPI.jl, and installation will fail. See the Troubleshooting section below.

All examples above install OpenMPI. If you wish to use MPICH, you will have to build MPICH, SCALAPACK and MUMPS by hand.

Building MUMPS.jl

If MUMPS and SCALAPACK are not in standard locations, you can help by setting the environment variables MUMPS_PREFIX and SCALAPACK_PREFIX.

The Homebrew and Linuxbrew methods above install MUMPS and SCALAPACK in nonstandard locations. You can define

julia> ENV["MUMPS_PREFIX"] = "/usr/local/opt/brewsci-mumps"
julia> ENV["SCALAPACK_PREFIX"] = "/usr/local/opt/brewsci-scalapack"

on macOS, and something of the form

julia> ENV["MUMPS_PREFIX"] = "/home/linuxbrew/.linuxbrew/opt/brewsci-mumps"
julia> ENV["SCALAPACK_PREFIX"] = "/home/linuxbrew/.linuxbrew/opt/brewsci-scalapack"

on Linux.

At the Julia prompt, type

julia> using Pkg
julia> Pkg.clone("https://github.com/JuliaSmoothOptimizers/MUMPS.jl.git")
julia> Pkg.build("MUMPS")
julia> Pkg.test("MUMPS")

Troubleshooting

On macOS or Linux, if you see the error message

[ 11%] Building Fortran object CMakeFiles/gen_constants.dir/gen_constants.f90.o
│ /home/ubuntu/.julia/packages/MPI/U5ujD/deps/gen_constants.f90:43:43:
│
│    call output("MPI_NO_OP       ", MPI_NO_OP)
│                                            1
│ Error: Symbol ‘mpi_no_op’ at (1) has no IMPLICIT type

your OpenMPI library is too old.

If you are running macOS and see error messages of the form

PMIx has detected a temporary directory name that results in a path that is too long for the Unix domain socket:

  Temp dir:
  /var/folders/rq/p5nq9tv17p5drlk49755jjz80000gn/T/openmpi-sessions-501@your_computer_name_0/44473

Try setting your TMPDIR environmental variable to point to something shorter in length

simply exit Julia and set the environment variable TMPDIR to, e.g., \tmp:

$ export TMPDIR=/tmp

The issue has to do with OpenMPI and is documented in their faq.

How to Use

The main data type holding information on a factorization is Mumps. Remember to initialize MPI before attempting to create a Mumps object. A simple session is as follows:

julia> using MUMPS
julia> using MPI
julia> MPI.Init()
julia> A = sprand(10, 10, .2) + speye(10); rhs = rand(10)
julia> x = solve(A, rhs)  # Mumps object is created and destroyed
julia> norm(x - A \ rhs) / norm(x)
2.640677159735313e-16
julia> MPI.Finalize()     # if you're finished

It is possible to separate the initialization, the analysis/factorization, and the solve phases. It is also possible to access the information reported by MUMPS after the factorization and solve phases, and to modify this information (e.g., to perform iterative refinement).

When creating an instance of a Mumps object explicitly, it is important to specify in advance what arithmetic should be used. Single and double precision real (Float32 and Float64) and complex (Complex64 and Complex128) arithmetics are supported.

For instance,

julia> MPI.Init()
julia> mumps = Mumps{Float64}(mumps_unsymmetric, default_icntl, default_cntl64)  # Real, general unsymmetric
julia> A = sparse(rand(4,4)); rhs = rand(4)       # Happens on all cores
julia> associate_matrix!(mumps, A)
julia> factorize!(mumps)
julia> associate_rhs!(mumps, rhs)
julia> solve!(mumps)
julia> x = get_solution(mumps)
julia> finalize(mumps)
julia> MPI.Finalize()

Once the arithmetic of the Mumps instance has been specified, it cannot be changed. The module is flexible in that various data types may be used to define the matrix to be factorized and the right-hand side, and appropriate conversions will take place. Dense matrices may be used, and they will be converted to sparse format.

For intance,

julia> mumps = Mumps{Complex128}(mumps_unsymmetric, default_icntl, default_cntl64)
julia> A = rand(Int16, 4, 4); rhs = rand(Float32, 4)
julia> associate_matrix!(mumps, A)  # A is converted to a sparse Complex128 matrix
julia> associate_rhs!(mumps, rhs)   # rhs is converted to a Complex128 array

See test for more examples.

Constants and Methods Exposed

Constants

The following convenience constants may be used when initializing a Mumps object:

See Sections 5.1 and 5.2 of the MUMPS User's Manual for a description of the integer and real control arrays.

Methods

A Mumps object is created using the default constructor, which must be supplied with:

• the data type for the arithmetic to be used, as a type parameter, i.e., Mumps{Float64}(...) or Mumps{Complex128}(...)
• sym: one of the constants mumps_unsymmetric, mumps_definite or mumps_symmetric. Note that there is no support for Hermitian complex matrices in MUMPS. Therefore, we recommend to always use mumps_unsymmetric for complex data.
• icntl: an integer parameters array (see the MUMPS Users's Manual)
• cntl: a real parameters array (see the MUMPS Users's Manual)

The convenience function get_icntl() returns an array of integer parameters corresponding to certain commonly-used options. Its arguments are all optional:

• det: a boolean indicating whether the determinant should be computed (default: false)
• verbose: a boolean (default: false)
• ooc: a boolean indicating whether factors should be stored out of core (default: false)
• itref: the number of iterative refinement steps (default: 0).

A Mumps object is destroyed by calling the finalize() method. Because finalize still issues MPI commands, it is important to call finalize() before calling MPI.Finalize().

Parallel Execution

MPI is controled by way of MPI.jl. Look for the lines that say NUMBER OF WORKING PROCESSES in the output of

mpirun -np 4 julia examples/mumps_mpi.jl

To Do (Pull Requests Welcome!)

• Support double precision complex arithmetic (in 99c23fe)
• Support single precision real and complex arithmetic (in 654814a)
• Support distributed matrices / vectors
• User-selected permutation
• Out-of-core option (in 73e829b)
• Determinant (in 73e829b)
• Compute entries of the inverse
• Control iterative refinement (in 73e829b)
• Obtain a Schur complement
• Solve with sparse right-hand sides
• Sequential, version with no MPI requirement