qpsolvers_benchmark 1.2.0

Benchmark for quadratic programming solvers available in Python.

QP solvers benchmark

Benchmark for quadratic programming (QP) solvers available in Python.

The goal of this benchmark is to help users compare and select QP solvers. Its methodology is open to discussions. The benchmark ships standard and community test sets, as well as a qpsolvers_benchmark command-line tool to run test sets directly. The main output of the benchmark are standardized reports evaluating all metrics across all QP solvers available on the test machine. This repository also distributes results from running the benchmark on a reference computer.

New test sets are welcome! The benchmark is designed so that each test-set comes in a standalone directory. Feel free to create a new one and contribute it here so that we grow the collection over time.

Test sets

The benchmark comes with standard and community test sets to represent different use cases for QP solvers:

Test set	Problems	Brief description
Maros-Meszaros	138	Standard, designed to be difficult.
Maros-Meszaros dense	62	Subset of Maros-Meszaros restricted to smaller dense problems.
GitHub free-for-all	12	Community-built, new problems are welcome!

Solvers

Solver	Keyword	Algorithm	Matrices	License
Clarabel	clarabel	Interior point	Sparse	Apache-2.0
CVXOPT	cvxopt	Interior point	Dense	GPL-3.0
DAQP	daqp	Active set	Dense	MIT
ECOS	ecos	Interior point	Sparse	GPL-3.0
Gurobi	gurobi	Interior point	Sparse	Commercial
HiGHS	highs	Active set	Sparse	MIT
HPIPM	hpipm	Interior point	Dense	BSD-2-Clause
MOSEK	mosek	Interior point	Sparse	Commercial
NPPro	nppro	Active set	Dense	Commercial
OSQP	osqp	Douglas–Rachford	Sparse	Apache-2.0
PIQP	piqp	Proximal Interior Point	Dense & Sparse	BSD-2-Clause
ProxQP	proxqp	Augmented Lagrangian	Dense & Sparse	BSD-2-Clause
qpOASES	qpoases	Active set	Dense	LGPL-2.1
qpSWIFT	qpswift	Interior point	Sparse	GPL-3.0
quadprog	quadprog	Goldfarb-Idnani	Dense	GPL-2.0
SCS	scs	Douglas–Rachford	Sparse	MIT

Metrics

We evaluate QP solvers based on the following metrics:

• Success rate: percentage of problems a solver is able to solve on a given test set.
• Computation time: time a solver takes to solve a given problem.
• Optimality conditions: we evaluate all three optimality conditions:
  • Primal residual: maximum error on equality and inequality constraints at the returned solution.
  • Dual residual: maximum error on the dual feasibility condition at the returned solution.
  • Duality gap: value of the duality gap at the returned solution.
• Cost error: difference between the solution cost and the known optimal cost.

Shifted geometric mean

Each metric (computation time, primal and dual residuals, duality gap) produces a different ranking of solvers for each problem. To aggregate those rankings into a single metric over the whole test set, we use the shifted geometric mean (shm), which is a standard to aggregate computation times in benchmarks for optimization software. This mean has the advantage of being compromised by neither large outliers (as opposed to the arithmetic mean) nor by small outliers (in contrast to the geometric geometric mean). Check out the references below for further details.

Here are some intuitive interpretations:

• A solver with a shifted-geometric-mean runtime of $Y$ is $Y$ times slower than the best solver over the test set.
• A solver with a shifted-geometric-mean primal residual $R$ is $R$ times less accurate on equality and inequality constraints than the best solver over the test set.

Results

The outcome from running a test set is a standardized report comparing solvers against the different metrics. Here are the results obtained on a reference computer:

Test set	Results	CPU info
GitHub free-for-all	Full report	Intel(R) Core(TM) i7-6500U CPU @ 2.50GHz
Maros-Meszaros	Full report	Intel(R) Core(TM) i7-6500U CPU @ 2.50GHz
Maros-Meszaros dense	Full report	Intel(R) Core(TM) i7-6500U CPU @ 2.50GHz

You can check out results from a variety of machines, and share the reports produced by running the benchmark on your own machine, in the Results category of the discussions forum.

Limitations

Here are some known areas of improvement for this benchmark:

• Cold start only: we don't evaluate warm-start performance for now.
• CPU thermal throttling: the benchmark currently does not check the status of CPU thermal throttling. Adding this feature is a good way to start contributing to the benchmark.

Check out the issue tracker for ongoing works and future improvements.

Installation

You can install the benchmark and its dependencies in an isolated environment using conda:

conda env create -f environment.yaml
conda activate qpsolvers_benchmark

Alternatively, you can install the benchmark on your system using pip:

pip install qpsolvers_benchmark

By default, the benchmark will run all supported solvers it finds.

Running the benchmark

Once the benchmark is installed, you will be able to run the qpsolvers_benchmark command. Provide it with the script corresponding to the test set you want to run, followed by a benchmark command such as "run". For instance, let's run the "dense" subset of the Maros-Meszaros test set:

qpsolvers_benchmark maros_meszaros/maros_meszaros_dense.py run

You can also run a specific solver, problem or set of solver settings:

qpsolvers_benchmark maros_meszaros/maros_meszaros_dense.py run --solver proxqp --settings default

Check out qpsolvers_benchmark --help for a list of available commands and arguments.

Plots

The command line ships a plot command to compare solver performances over a test set for a specific metric. For instance, run:

qpsolvers_benchmark maros_meszaros/maros_meszaros_dense.py plot runtime high_accuracy

To generate the following plot:

Contributing

Contributions to improving this benchmark are welcome. You can for instance propose new problems, or share the runtimes you obtain on your machine. Check out the contribution guidelines for details.

Citation

To cite this benchmark in your scientific works, follow the Cite this repository button on GitHub (top-right of the repository page).

See also

References

• How not to lie with statistics: the correct way to summarize benchmark results: why geometric means should always be used to summarize normalized results.
• Optimality conditions and numerical tolerances in QP solvers: note written while figuring out the high_accuracy settings of this benchmark.

Other benchmarks

• Benchmarks for optimization software by Hans Mittelmann, which includes reports on the Maros-Meszaros test set.
• jrl-qp/benchmarks: benchmark of QP solvers available in C++.
• osqp_benchmark: benchmark examples for the OSQP solver.
• proxqp_benchmark: benchmark examples for the ProxQP solver.