MUMPS 5.8.2 - Sparse Direct Solver

MUltifrontal Massively Parallel sparse direct Solver

MUMPS solves large sparse systems of linear equations A x = b using Gaussian elimination with multifrontal techniques. It supports all four arithmetic precisions (single/double real, single/double complex) and runs on sequential, multithreaded, or distributed-memory (MPI) architectures.

This enhanced distribution adds a Python/ctypes interface, a Typer CLI, a FastAPI webapp dashboard, and a comprehensive benchmarking suite with pre-generated test matrices.

Upstream: mumps-solver.org | License: CeCILL-C

---

Quick Start

# 1. Build all precisions (shared libraries)
make allshared

# 2. Run the benchmark with a test matrix
LD_LIBRARY_PATH=lib:src/PORD/lib \
  examples/dmumps_bench --matrix-file benchmarks/matrices/test_large_100x100_laplacian.coo

# 3. Install Python interface & CLI
pip install -e ".[cli,test]"

# 4. Use the CLI
mumps build all --shared
mumps test pytest
mumps vendor detect
mumps benchmark sparse

# 5. Start the webapp dashboard
mumps webapp start --port 9002

---

Directory Layout

MUMPS/
├── src/                 Fortran/C source (80 templates + 23 hand-written files)
│   ├── templates/       Template files (.F.in) for precision generation
│   ├── PORD/            PORD ordering package
│   ├── libseq/          Sequential MPI stub library
│   ├── MATLAB/          MATLAB interface
│   └── SCILAB/          SCILAB interface
├── include/             C/Fortran header files (xmumps_c.h, etc.)
├── lib/                 Built libraries (libxmumps.so, libmumps_common.so)
├── examples/            Example programs and benchmark drivers
├── python/              Python/ctypes interface (pymumps)
├── cli/                 Typer CLI (mumps command)
├── webapp/              FastAPI dashboard with Jinja2 templates
├── tests/               pytest test suite with SQLite result storage
├── scripts/             Build, benchmark, and vendor testing scripts
├── benchmarks/          Test matrices (.coo) and benchmark fixtures
├── docs/                User guide (PDF) and license texts
├── Make.inc/            Platform-specific Makefile configurations
└── cmake/               CMake build support

---

Building

See INSTALL for detailed instructions. The basic workflow:

# Copy a platform config
cp Make.inc/Makefile.debian.SEQ Makefile.inc

# Edit Makefile.inc to set compiler flags, BLAS library, orderings, etc.
# Then build:
make allshared        # All 4 precisions, shared libraries
make dshared          # Double precision only
make all              # Static libraries

BLAS Vendor Detection

The CLI can auto-detect available BLAS libraries:

mumps vendor detect           # Show all detected BLAS libraries
mumps vendor rebuild          # Build MUMPS against each vendor (skips existing)
mumps vendor rebuild --force  # Force rebuild all vendors

Detection priority: pkg-config > ldconfig > environment ($MKLROOT) > filesystem scan.

---

Python Interface

The pymumps package provides a ctypes wrapper for all 4 MUMPS precisions:

from pymumps import MumpsSolver

solver = MumpsSolver(precision='d', sym=0)
solver.set_matrix(n=1000, irn=rows, jcn=cols, a=values)
solver.set_rhs(rhs)
solver.solve()               # JOB=6: analysis + factorization + solve
x = solver.get_solution()
solver.finalize()

Install with: pip install -e . or uv pip install -e .

---

CLI

The mumps command wraps all build, test, benchmark, and vendor operations:

Command	Description
mumps build library -p d	Build double precision library
mumps build all --shared	Build all 4 precisions (shared)
mumps build examples -p d	Build example programs
mumps test pytest	Run Python test suite
mumps test vendors	Test all BLAS vendors
mumps test fortran	Run Fortran example tests
mumps benchmark sparse	Sparse solver benchmark (JOB=6)
mumps benchmark dense	Dense GEMM benchmark
mumps benchmark all	Full suite with SQLite storage
mumps vendor detect	Detect BLAS libraries
mumps vendor rebuild	Build per-vendor libraries
mumps webapp start	Start dashboard at localhost:9002
mumps info system	Show compilers, BLAS, memory
mumps clean	Clean build artifacts

Install CLI dependencies: pip install -e ".[cli]"

---

Benchmarks

The benchmark drivers (examples/[sdcz]mumps_bench) support both auto-generated Laplacian matrices and external .coo matrix files:

# Auto-generated 220x220 Laplacian (default)
examples/dmumps_bench --ngrid 220 --nrhs 4

# Load from .coo file
examples/dmumps_bench --matrix-file benchmarks/matrices/test_large_100x100_laplacian.coo

# With symmetry flag (0=unsym, 1=SPD, 2=general sym)
examples/dmumps_bench --matrix-file matrix.coo --sym 1 --ordering 5

Pre-generated Test Matrices

10 SPD matrices in MUMPS coordinate format (.coo, 1-based indexing):

Category	Matrix	Dimension	Non-zeros
Small	test_tiny_10x10_diagonal.coo	10	10
Small	test_small_10x10_laplacian.coo	100	460
Small	test_small_100_tridiagonal.coo	100	298
Medium	test_medium_50x50_laplacian.coo	2,500	12,300
Medium	test_medium_1000_tridiagonal.coo	1,000	2,998
Medium	test_medium_1000x1000_diagonal.coo	1,000	1,000
Large	test_large_100x100_laplacian.coo	10,000	49,600
Large	test_large_10000_tridiagonal.coo	10,000	29,998
Huge	test_huge_200x200_laplacian.coo	40,000	199,200
Huge	test_huge_50000_tridiagonal.coo	50,000	149,998

Generate custom matrices:

python3 scripts/generate_test_matrices.py --size all
python3 scripts/generate_test_matrices.py --type laplacian --ngrid 300

---

Testing

# Python tests (pytest)
mumps test pytest
# or directly:
uv run pytest tests/ -v

# Multi-vendor testing (builds & tests against each BLAS)
./scripts/test_all_vendors.sh

# Compare vendor results
./scripts/compare_vendors.sh

# BLAS correctness checks
./scripts/check_blas_correctness.sh

Test results are stored in SQLite (mumps.sqlite) and viewable in the webapp.

---

Webapp Dashboard

A FastAPI web interface for monitoring benchmarks, test results, and system information:

mumps webapp start --port 9002 --reload
# Visit http://localhost:9002

Screenshots

Dashboard -- Overview with navigation to all sections, benchmark results summary, and sparse matrix modules reference.

Benchmarks -- Available benchmark suite with execution history, configuration details, and test matrix catalog.

Migration Status -- Live KPIs tracking Fortran 2023 and C23 modernization progress across all source files.

CLI Dashboard -- Interactive command execution with Graphviz use-case flowchart and job output panel.

Pytest Results -- Python test suite results with per-vendor breakdown, session history, and failure analysis.

---

Fortran Modernization

This distribution includes a fully modernized codebase:

• 103/103 source files converted to free-form Fortran (100%)
• 80 template files generating 320 precision-specific files via scripts/generate_from_template.sh
• 13 PORD C files migrated to C23 (ANSI function prototypes)
• Build system with auto-detection of C23 (-std=gnu23) and Fortran 2023 (-std=f2023)

See MODERNIZATION_STATUS.md for full details.

---

References

• MUMPS Users' Guide: docs/userguide_5.8.2.pdf
• MUMPS website: mumps-solver.org
• License: CeCILL-C v1

Copyright 1991-2026 CERFACS, CNRS, ENS Lyon, INP Toulouse, Inria, Mumps Technologies, University of Bordeaux.