Russell Sparse - Sparse matrix tools and solvers

This crate is part of Russell - Rust Scientific Library

This repository contains tools for handling sparse matrices and functions to solve large sparse systems.

Documentation:

• API reference (docs.rs)

Installation

Install some libraries:

bash sudo apt-get install \ liblapacke-dev \ libmumps-seq-dev \ libopenblas-dev \ libsuitesparse-dev

Add this to your Cargo.toml (choose the right version):

toml [dependencies] russell_sparse = "*"

Optional: Use a locally compiled MUMPS library

The standard Debian libmumps-seq-dev does not come with Metis or OpenMP that may lead to faster calculations. Therefore, it may be advantageous to use a locally compiled MUMPS library.

We just need the include files in /usr/local/include/mumps and a library file named libdmumps_open_seq_omp in /usr/local/lib/mumps.

Follow the instructions from https://github.com/cpmech/script-install-mumps and then set the environment variable USE_LOCAL_MUMPS=1:

bash export USE_LOCAL_MUMPS=1

Number of threads

By default OpenBLAS will use all available threads, including Hyper-Threads that make the performance worse. Thus, it is best to set the following environment variable:

bash export OPENBLAS_NUM_THREADS=<real-core-count>

Furthermore, if working on a multi-threaded application, it is recommended to set:

bash export OPENBLAS_NUM_THREADS=1

Examples

Solve a sparse linear system

```rust use russelllab::{Matrix, Vector}; use russellsparse::{ConfigSolver, Solver, SparseTriplet, Symmetry, StrError};

fn main() -> Result<(), StrError> { // allocate a square matrix let mut trip = SparseTriplet::new(3, 3, 5, Symmetry::No)?; trip.put(0, 0, 0.2)?; trip.put(0, 1, 0.2)?; trip.put(1, 0, 0.5)?; trip.put(1, 1, -0.25)?; trip.put(2, 2, 0.25)?;

// print matrix
let (m, n) = trip.dims();
let mut a = Matrix::new(m, n);
trip.to_matrix(&mut a)?;
let correct = "┌                   ┐\n\
               │   0.2   0.2     0 │\n\
               │   0.5 -0.25     0 │\n\
               │     0     0  0.25 │\n\
               └                   ┘";
assert_eq!(format!("{}", a), correct);

// allocate rhs
let rhs1 = Vector::from(&[1.0, 1.0, 1.0]);
let rhs2 = Vector::from(&[2.0, 2.0, 2.0]);

// calculate solution
let config = ConfigSolver::new();
let (mut solver, x1) = Solver::compute(config, &trip, &rhs1)?;
let correct1 = "┌   ┐\n\
                │ 3 │\n\
                │ 2 │\n\
                │ 4 │\n\
                └   ┘";
assert_eq!(format!("{}", x1), correct1);

// solve again
let mut x2 = Vector::new(trip.dims().0);
solver.solve(&mut x2, &rhs2)?;
let correct2 = "┌   ┐\n\
                │ 6 │\n\
                │ 4 │\n\
                │ 8 │\n\
                └   ┘";
assert_eq!(format!("{}", x2), correct2);
Ok(())

} ```

Sparse solvers

We wrap two direct sparse solvers: UMFPACK (aka UMF) and MUMPS (aka MMP). The default solver is UMF; however UMF may run out of memory for large matrices, whereas MMP still may work. The MMP solver is not thread-safe and thus must be used in single-threaded applications.

Tools

This crate includes a tool named solve_mm_build to study the performance of the available sparse solvers (currently MMP and UMF). The _build suffix is to disable the coverage tool.

solve_mm_build reads a Matrix Market file and solves the linear system:

text a ⋅ x = rhs

with a right-hand-side containing only ones.

The data directory contains an example of Matrix Market file named bfwb62.mtx and you may download more matrices from https://sparse.tamu.edu/

Run the command:

bash cargo run --release --bin solve_mm_build -- data/matrix_market/bfwb62.mtx

Or

bash cargo run --release --bin solve_mm_build -- --help

for more options.