nasoq_benchmarks

The latest evaluation across problems of the SMP repository for different solvers in our benchmark:

Building benchmark

OSQP and NASOQ are already in the repository however, you need to get the license for other tools, i.e., Gurobi, Mosek, and QL.

Requirements

• NASOQ needs MKL Pardiso and METIS. CMake downloads the code from NASOQ repository.
• OSQP needs MKL library for fast solve time. CMake downloads the source from OSQP repository.
• MOSEK : Put the MOSEK license in

export MOSEKROOT=path/to/mosek/8/tools/platform/linux64x86/

• Gurobi: Put the Gurobi license in the home directory and export the path that gurobi is installed, for example:

export GUROBIROOT=/opt/gurobi811/linux64/

Build

First clone the repo and do not forget -recursive option:

git clone --recursive https://github.com/sympiler/nasoq-benchmarks

Before building the benchmark, you need to know the path to METIS and MKL library (CMake can find MKL if it is installed in the default location or is in the path). For MKL you need to set MKLROOT and for METIS you need to set METIS_ROOT_PATH. Following shows how these variables are set using cmake flags.

mkdir build
cd build
cmake -DMKL_ROOT_PATH=path/to/intel -DMETIS_ROOT_PATH=path/to/metis-5.1.0/build/Linux-x86_64/ -DCMAKE_BUILD_TYPE=Release ..
make

Running the benchmark

You first need to download the dataset and then run the all compiled solvers.

Downloading the dataset

wget https://www.dropbox.com/s/hxhelt273ajh00w/SMP_Repository_SIGGRAPH20.tgz?dl=0 

Running solvers

A script is provided that run all tools for specified dataset. To use the script emit the following command:

bash run_all path/to/dataset/ path/to/build/directory/

The script finds all sparse QP problems stored in yml and run all solvers for each of them for three different ranges of accuracy, i.e., 1e-3, 1e-6, and 1e-9.

Adding a solver

Check here for instructions.