cbcbox

cbcbox is a high-performance, self-contained Python distribution of the CBC MILP solver (COIN-OR Branch and Cut), built from the latest COIN-OR next branch.

On x86_64 (Linux, macOS, Windows) the wheel ships both a Haswell-optimised binary (AVX2/FMA) for maximum speed and a generic build with runtime CPU dispatch for compatibility with any x86_64 machine — selected automatically. All dynamic dependencies (OpenBLAS, libgfortran, etc.) are bundled; no system libraries or separate installation steps are needed.

Highlights

• Haswell-optimised & generic builds — on x86_64 Linux, macOS, and Windows the wheel ships two complete solver stacks: a Haswell build (OpenBLAS AVX2/FMA kernel) for maximum throughput, and a generic build (DYNAMIC_ARCH runtime dispatch) for compatibility with any x86_64 CPU. The best available variant is selected automatically at import time (see Build variants).

• Parallel branch-and-cut — built with --enable-cbc-parallel. Use -threads=N to distribute the search tree across N threads, giving significant speedups on multi-core machines for hard MIP instances.

• AMD fill-reducing ordering — SuiteSparse AMD is compiled in, enabling the high-quality UniversityOfFlorida Cholesky factorization for Clp's barrier (interior point) solver. AMD reordering produces much less fill-in on large sparse problems than the built-in native Cholesky, making barrier substantially faster. Activate with -cholesky UniversityOfFlorida -barrier (see barrier usage).

Performance (x86_64)

Auto-updated by CI after each successful workflow run. Single-threaded solve time — lower is better.

The AVX2/Haswell build is ~2.8× faster than the generic build on average (geometric mean across 30 instances, 3 x86_64 platforms: Darwin x86_64, Linux x86_64, Windows AMD64).

Single-threaded solve time across benchmark instances on Linux x86_64, sorted by solve time. Speedup factor shown above each pair. Lower is better.

See also: Windows AMD64 + macOS x86_64 summary

Build variants

On x86_64 Linux, macOS, and Windows, the wheel ships three complete sets of binaries:

Non-x86_64 platforms (Linux aarch64, macOS arm64) ship generic and debug only.

At import time cbcbox automatically selects avx2 when available and the running CPU supports AVX2; otherwise it falls back to generic.

You can override the selection with the CBCBOX_BUILD environment variable:

# Force generic (portable) build
CBCBOX_BUILD=generic cbc mymodel.mps -solve -quit

# Force AVX2-optimised build (raises an error if not available on this platform/CPU)
CBCBOX_BUILD=avx2 cbc mymodel.mps -solve -quit

# Force debug build (full symbols, no optimisation — useful for bug reports and GDB/LLDB)
CBCBOX_BUILD=debug cbc mymodel.mps -solve -quit

When CBCBOX_BUILD is set, a short summary of the selected build is printed to stdout on every call — useful for tagging experiment results:

[cbcbox] CBCBOX_BUILD=avx2
[cbcbox]   binary  : .../cbcbox/cbc_dist_avx2/bin/cbc
[cbcbox]   lib dir : .../cbcbox/cbc_dist_avx2/lib
[cbcbox]   libs    : libCbc.so.3, libClp.so.3, libopenblas.so.0

Set CBCBOX_VERBOSE=1 to always print this dispatch summary regardless of whether CBCBOX_BUILD is set — useful to confirm which binary is actually being invoked.

Supported platforms

Installation

pip install cbcbox

Usage

Command line

After installation, CBC is available directly as the cbc command (pip installs the entry point into the environment's bin/ on Linux/macOS or Scripts/ on Windows, which is already on PATH):

cbc mymodel.lp -solve -quit
cbc mymodel.mps.gz -solve -quit
cbc mymodel.mps -seconds 60 -timem elapsed -solve -quit
cbc mymodel.mps -dualp pesteep -solve -quit

Alternatively, invoke via the Python module entry point:

python -m cbcbox mymodel.lp -solve -quit

CBC accepts LP, MPS and compressed MPS (.mps.gz) files. Pass -help for the full list of options, or -quit to exit after solving.

Parallel branch-and-cut

This build includes parallel branch-and-cut (--enable-cbc-parallel). Use -threads=N to distribute the search tree across N threads:

cbc mymodel.mps -threads=4 -solve -quit

Barrier (interior-point) solver

Clp's barrier solver can be faster than simplex for large LP relaxations. This build includes SuiteSparse AMD, which enables the high-quality UniversityOfFlorida Cholesky factorization — significantly reducing fill-in compared to the built-in native Cholesky:

# Solve LP relaxation with barrier + AMD Cholesky, then crossover to simplex basis
cbc mymodel.mps -cholesky UniversityOfFlorida -barrier -solve -quit

# Useful as a root-node strategy inside MIP (let CBC use simplex for B&B):
cbc mymodel.mps -cholesky UniversityOfFlorida -barrier -solve -quit

Without AMD, only -cholesky native (less efficient) is available.

Python API

The package exposes helpers to locate the installed files:

import cbcbox
import subprocess

# Path to the cbc binary (cbc.exe on Windows).
cbcbox.cbc_bin_path()
# e.g. '/home/user/.venv/lib/python3.13/site-packages/cbcbox/cbc_dist/bin/cbc'

# Directory containing the shared libraries.
cbcbox.cbc_lib_dir()
# e.g. '.../cbcbox/cbc_dist/lib'

# Directory containing the COIN-OR C/C++ headers.
cbcbox.cbc_include_dir()
# e.g. '.../cbcbox/cbc_dist/include/coin'

# Run CBC programmatically.
result = subprocess.run(
    [cbcbox.cbc_bin_path(), "mymodel.mps", "-solve", "-quit"],
    capture_output=True, text=True,
)
print(result.stdout)

What is built

The build pipeline compiles all components from source inside the CI runner, in the following order:

On x86_64 Linux, macOS, and Windows the entire stack is compiled twice: once for the generic variant (OpenBLAS DYNAMIC_ARCH=1 with a broad set of x86_64 targets for runtime dispatch) and once for the avx2 variant (OpenBLAS DYNAMIC_ARCH=1 restricted to Haswell/Skylake targets via DYNAMIC_LIST, COIN-OR compiled with -march=haswell -DCOIN_AVX2=4). Both variants use NO_CBLAS=1 (COIN-OR only calls the Fortran BLAS interface). AMD is built only once (it is pure combinatorial code with no BLAS dependency) and reused by both COIN-OR variants.

The COIN-OR stack (CoinUtils, Osi, Clp, Cgl, Cbc) is always compiled with -ffp-contract=off, which prevents the compiler from fusing separate multiply/add operations into FMA instructions. FMA computes with extra intermediate precision, which can introduce tiny (last-bit) numerical differences that make CBC's branch-and-cut behave inconsistently across toolchains/architectures; disabling contraction keeps results reproducible. This flag is not applied to OpenBLAS or AMD, whose own numerics are unaffected by this concern.

Symmetry detection via Nauty is currently disabled (--without-nauty) and is not part of this build.

All COIN-OR components are built as shared (.so / .dylib / .dll) libraries. The shared libraries are patched with self-relative RPATHs and bundled inside the wheel, making them directly usable via cffi or ctypes without any system installation.

Wheel contents

The wheel installs under cbcbox/ inside the site-packages directory. On x86_64 Linux, macOS, and Windows it contains two dist trees; other platforms contain only cbc_dist/:

cbc_dist/           ← generic build (all platforms)
cbc_dist_avx2/      ← AVX2-optimised build (x86_64 Linux/macOS/Windows)
├── bin/
│   ├── cbc           # CBC MIP solver binary  (cbc.exe on Windows)
│   └── clp           # Clp LP solver binary   (clp.exe on Windows)
├── lib/
│   ├── libCbc.so / libCbc.dylib / libCbc.dll  # CBC solver
│   ├── libCbcSolver.so ...
│   ├── libClp.so ...                          # Clp LP solver
│   ├── libCgl.so ...                          # Cut generation
│   ├── libOsi.so ...                          # Solver interface
│   ├── libOsiClp.so ...                       # Clp OSI binding
│   ├── libOsiCbc.so ...                       # CBC OSI binding (where available)
│   ├── libCoinUtils.so ...
│   ├── libopenblas.so / .dylib / .dll         # OpenBLAS BLAS/LAPACK
│   ├── pkgconfig/                             # .pc files for all libraries
│   └── <bundled runtime shared libs>          # Platform-specific — see below
└── include/
    ├── coin/      # COIN-OR headers (CoinUtils, Osi, Clp, Cgl, Cbc)
    └── *.h        # SuiteSparse / AMD headers

Bundled dynamic libraries

Because OpenBLAS links to the Fortran runtime, the following shared libraries are bundled inside the wheel and their paths are rewritten so no system installation is required.

Linux (lib/ directory, RPATH set to $ORIGIN)

macOS (lib/ directory, install names rewritten to @rpath/)

Windows (bin/ directory, DLLs placed next to the executable)

CI / build pipeline

Wheels are built and tested automatically via GitHub Actions using cibuildwheel. The workflow (.github/workflows/wheel.yml) runs independent compile jobs in parallel, then packages each platform:

Each platform's compile jobs run in parallel. Once all compile jobs for a platform finish, the corresponding package-* job assembles the wheel via cibuildwheel and runs the test suite against the installed wheel.

A final combine_reports job collects per-platform performance results and commits the updated README.md to the repository.

Integration tests

The test suite (pytest) solves 24 MIP instances and checks the optimal objective values, in both single-threaded and parallel (3-thread) modes. On x86_64 Linux, macOS, and Windows each test is run twice — once against the generic binary and once against the avx2 binary — and a side-by-side performance comparison is recorded:

Time limits are generous to avoid false failures on slow CI runners.

Performance results

Auto-updated by CI after each successful workflow run.

Summary

Geometric mean solve time (seconds) across all test instances.

1 thread

3 threads

Per-instance results

pp08a

sprint_hidden06_j

air03

air04

air05

nw04

mzzv11

trd445c

nursesched-sprint02

stein45

neos-810286

neos-1281048

j3050_8

qiu

gesa2-o

pk1

mas76

app1-1

eil33-2

fiber

neos-2987310-joes

neos-827175

neos-3083819-nubu

markshare_4_0

irp

qap10

swath1

physiciansched6-2

mzzv42z

neos-860300

NAQ — Never Asked Questions

Why not benchmark on the full MIPLIB 2017 library?

Several practical constraints shape the benchmark set:

1. CI time limits. GitHub Actions enforces a 6-hour wall-clock limit per job. The full MIPLIB 2017 collection contains ~240 instances, many of which take hours even on fast hardware. Including all of them would make every CI run time out before producing any useful measurements.

2. Comparing apples to apples requires instances solved to optimality. If some instances are only solved within a time limit (i.e., a gap > 0 %), a meaningful performance comparison must account for both solve time and solution quality simultaneously. This greatly complicates analysis and makes plots harder to interpret. Restricting to instances that CBC reliably solves to proven optimality keeps the comparison clean: a single elapsed-time number per instance is all that is needed.

3. The instance set is intentionally biased toward set packing / covering / partitioning structure. Most instances in the benchmark (pp08a, sprint_hidden06_j, nw04, mzzv11, nursesched-sprint02, air0x, trd445c) contain large blocks of set packing, covering, or partitioning constraints. This structure arises naturally in applications such as crew scheduling, nurse scheduling, vehicle routing, and cutting stock — exactly the domain where column generation is most valuable. Since the benchmark focuses on this problem class rather than providing a general-purpose solver survey, it is a specially interesting use case.

Local debug builds

The released wheels include an optimised build and a debug build (see Build variants). For most debugging needs, CBCBOX_BUILD=debug is all you need. If you want to rebuild with a sanitizer or need exact parity with the CI container, use the scripts in scripts/.

Quick start

Linux / macOS (native build):

# x86_64 → debug + AVX2 → cbc_dist_debug_avx2/bin/cbc
# ARM64  → debug only  → cbc_dist_debug/bin/cbc
./scripts/build_debug.sh

# With AddressSanitizer (Linux/macOS only):
./scripts/build_debug.sh --asan

# With ThreadSanitizer:
./scripts/build_debug.sh --tsan

# Force a clean rebuild from scratch (required when switching sanitizers):
./scripts/build_debug.sh --asan --clean

Linux (manylinux_2_28 container — matches CI exactly):

# Requires Docker; the script prints install instructions if it is missing.
./scripts/build_debug_manylinux.sh
./scripts/build_debug_manylinux.sh --asan
./scripts/build_debug_manylinux.sh --tsan

Windows (PowerShell):

# Requires MSYS2 at C:\msys64.  Note: sanitizers are not supported on Windows/MinGW.
.\scripts\build_debug_windows.ps1
.\scripts\build_debug_windows.ps1 -Clean   # force full rebuild

Debugging

# GDB (Linux):
gdb cbc_dist_debug_avx2/bin/cbc
(gdb) run mymodel.mps -solve -quit

# LLDB (macOS):
lldb cbc_dist_debug/bin/cbc
(lldb) run mymodel.mps -solve -quit

Sanitizer tips

Note: The debug build shipped in the wheel does not include a sanitizer. Use the local build scripts above (--asan / --tsan) on your development machine to enable sanitizer instrumentation.

ASan and TSan are mutually exclusive. Neither is available on Windows/MinGW. Always pass --clean when switching from one sanitizer to another to avoid linking mismatched object files.

OpenBLAS is always built without sanitizer flags to avoid false positives from hand-optimised BLAS assembly; only the COIN-OR stack is instrumented.

License

CBC and all COIN-OR components are distributed under the Eclipse Public License 2.0. OpenBLAS is distributed under the BSD 3-Clause licence. SuiteSparse AMD is distributed under the BSD 3-Clause licence.