Gurobi

Gurobi is a commercial software suite for solving complex optimization problems. This page wiki page describes the non-commercial use of Gurobi software on our clusters and is currently a work in progress.

License Limitations

We support and provide a free license to use Gurobi on the Graham, Cedar, Béluga and Niagara clusters. The license provides a total number of 4096 simultaneous uses (tokens in use) and permits distributed optimization with up to 100 nodes. A single user can run multiple simultaneous jobs. In order to use Gurobi you must agree to certain conditions. Please contact support and include a copy of the following completed Academic Usage Agreement. You will then be added into our license file as a permitted user within a few days.

Academic Usage Agreement

My Compute Canada username is "_______" and I am a member of the academic institution "_____________________".  This message confirms that I will only use the Compute Canada Gurobi license provided on Compute Canada systems for the purpose of non-commercial research project(s) to be published in publicly available article(s).

Configuring your account

You do NOT need to create a ~/.licenses/gurobi.lic file. The required settings to use our Gurobi license are configured by default when you load a Gurobi module on any cluster. To verify your username has been added to our Gurobi license and is working properly, run the following command:

$ module load gurobi
$ gurobi_cl 1> /dev/null && echo Success || echo Fail

If it returns "Success" you can begin using Gurobi immediately. If the test returns "Fail" then check whether a file named ~/.license/gurobi exists; if so, rename or remove this file, reload the module and try the test again. If it still returns "Fail", check whether there are any environment variables containing GUROBI defined in either of our your ~/.bashrc or ~/.bash_profile files. If you find any, comment or remove those lines, log out and log in again, reload the Gurobi module and try the test again. If you still get "Fail", contact contact support for help.

Minimizing License Checkouts

Note that all Gurobi license checkouts are handled by a single license server located in Ontario; it is therefore important for you to make sure that your use of Gurobi limits as much as possible license checkout attempts. Rather than checking out a license for each invocation of Gurobi in a job--which may occur dozens or even hundreds of times--you should ensure that your program, whatever the language or computing environment used, only makes a single license checkout and then reuses this license token throughout the lifetime of the job. This will improve your job's performance since contacting a remote license server is very costly in time and will also improve the responsiveness of our license server for everyone who is using Gurobi. Failure to use Gurobi carefully in this regard may ultimately result in random intermittent license checkout failures for all users. If this happens, you will be contacted and asked to kill all your jobs until your program is fixed and tested to ensure the problem is gone. Some documentation on this subject for C++ programs may be found here, explaining how to create a single Gurobi environment which can then be used for all your models. Python users can consult this page, which discusses how to implement this same idea of using a single environment and thus a single license token with multiple models. Other programs when run in parallel that call Gurobi such as R can also easily trigger the problem especially when many simultaneous jobs are submitted and/or run in parallel.

Interactive Allocations

Gurobi Command-Line Tools

[gra-login2:~] salloc --time=1:00:0 --cpus-per-task=8 --mem=1G --account=def-xyz
[gra800:~] module load gurobi
[gra800:~] gurobi_cl Record=1 Threads=8 Method=2 ResultFile=p0033.sol LogFile=p0033.log $GUROBI_HOME/examples/data/p0033.mps
[gra800:~] gurobi_cl --help

Gurobi Interactive Shell

[gra-login2:~] salloc --time=1:00:0 --cpus-per-task=8 --mem=1G --account=def-xyz
[gra800:~] module load gurobi
[gra800:~] echo "Record 1" > gurobi.env    see *
[gra800:~] gurobi.sh
gurobi> m = read('/cvmfs/restricted.computecanada.ca/easybuild/software/2017/Core/gurobi/8.1.1/examples/data/glass4.mps') 
gurobi> m.Params.Threads = 8               see **
gurobi> m.Params.Method = 2
gurobi> m.Params.ResultFile = "glass4.sol"
gurobi> m.Params.LogFile = "glass4.log"
gurobi> m.optimize()
gurobi> m.write('glass4.lp')
gurobi> m.status                           see ***
gurobi> m.runtime                          see ****
gurobi> help()

where

   * https://www.gurobi.com/documentation/8.1/refman/recording_api_calls.html
  ** https://www.gurobi.com/documentation/8.1/refman/parameter_descriptions.html
 *** https://www.gurobi.com/documentation/8.1/refman/optimization_status_codes.html
**** https://www.gurobi.com/documentation/8.1/refman/attributes.html

Replaying API calls

You can record API calls and repeat them with

[gra800:~] gurobi_cl recording000.grbr

Reference: https://www.gurobi.com/documentation/8.1/refman/recording_api_calls.html

Cluster Batch Job Submission

Once a Slurm script has been prepared for a Gurobi problem, it can be submitted to the queue by running the sbatch script-name.sh command. The jobs status in the queue can then be checked by running the sq command. The following Slurm scripts demonstrate solving 2 problems provided in the examples directory of each Gurobi module.

Data Example

The following Slurm script utilizes the Gurobi command-line interface to solve a simple coin production model written in LP format. The last line demonstrates how parameters can be passed directly to the Gurobi command-line tool gurobi_cl using simple command line arguments. For help selecting which parameters are best used for a particular problem and for choosing optimal values, refer to both the Performance and Parameters and Algorithms and Search sections found in the Gurobi Knowledge Base as well as the extensive online Gurobi documentation.

File : script-lp_coins.sh

#!/bin/bash
#SBATCH --account=def-group   # some account
#SBATCH --time=0-00:30        # specify time limit (D-HH:MM)
#SBATCH --cpus-per-task=8     # specify number threads
#SBATCH --mem=4G              # specify total memory
#SBATCH --nodes=1             # do not change

#module load StdEnv/2016      # for versions < 9.0.3 
module load StdEnv/2020       # for versions > 9.0.2

module load gurobi/9.5.0

rm -f coins.sol
gurobi_cl Threads=$SLURM_CPUS_ON_NODE Method=2 ResultFile=coins.sol ${GUROBI_HOME}/examples/data/coins.lp

Python Example

This is an example Slurm script for solving a simple facility location model with Gurobi Python. The example shows how to set the threads parameter with a dynamically created gurobi.env file when using the Gurobi python interface. This must be done in the working directory for each submitted job, otherwise Gurobi will (by default) start as many execute threads as there are physical cores on the compute node, instead of using the number of physical cores allocated to the job by Slurm, potentially slowing down the job and negatively impacting other user jobs running on the same node.

File : script-py_facility.sh

#!/bin/bash
#SBATCH --account=def-group   # some account
#SBATCH --time=0-00:30        # specify time limit (D-HH:MM)
#SBATCH --cpus-per-task=4     # specify number threads
#SBATCH --mem=4G              # specify total memory
#SBATCH --nodes=1             # do not change

#module load StdEnv/2016      # for versions < 9.0.3 
module load StdEnv/2020       # for versions > 9.0.2

module load gurobi/9.5.0

echo "Threads ${SLURM_CPUS_ON_NODE:-1}" > gurobi.env

gurobi.sh ${GUROBI_HOME}/examples/python/facility.py

Using Gurobi in Python virtual environments

Gurobi brings it's own version of Python which does not contain any 3rd-party Python packages except Gurobi. In order to use Gurobi together with popular Python packages like NumPy, Matplotlib, Pandas and others, we need to create a virtual Python environment in which we can install both gurobipy and e.g. pandas.

Before we start, we need to decide which combination of versions for Gurobi and Python to use. For example compare the Python versions supported by the Gurobi 8.11 and 9.0.1 modules:

[name@server ~] $ module load StdEnv/2016  (not needed for gurobi>9.0.2)
[name@server ~] $ module load gurobi/8.1.1
[name@server ~] $ cd $EBROOTGUROBI/lib
[name@server ~] $ ls -d python*
  python2.7        python2.7_utf32    python2.7_utf16
  python3.5_utf32  python3.6_utf32    python3.7_utf32

[name@server ~] $ module load StdEnv/2016  (not needed for gurobi>9.0.2)
[name@server ~] $ module load gurobi/9.0.1
[name@server ~] $ cd $EBROOTGUROBI/lib
[name@server ~] $ ls -d python*
  python3.7        python2.7_utf16  python2.7_utf32
  python3.5_utf32  python3.8_utf32  python3.6_utf32  python3.7_utf32

We see that gurobi/8.1.1 brings it's own installation of python2.7 and Python packages for Python 2.7, 3.5, 3.6 and 3.7 (pythonX.Y_utf32) while gurobi/9.0.1 by default uses python3.7 and brings Python packages for Python 2.7, 3.5, 3.6, 3.7 and 3.8 (pythonX.Y_utf32).

Creating the virtual environment

In this example we want to create a Python environment based on python/3.7 in which we want to use gurobi/9.0.1 and install the Pandas package. The following steps need to be done only once per system. First load the modules to create the virtual environment and activate it:

[name@server ~] $ module load gurobi/9.0.1 python/3.7
[name@server ~] $ virtualenv --no-download  ~/env_gurobi
  Using base prefix '/cvmfs/soft.computecanada.ca/easybuild/software/2017/Core/python/3.7.4'
  New python executable in /home/name/env_gurobi/bin/python
  Installing setuptools, pip, wheel...
  done.
[name@server ~] $ source ~/env_gurobi/bin/activate

Now that the environment has been activated, install the Python packages we want to use (in this case pandas) for example:

(env_gurobi) [name@server ~] $ pip install --no-index  pandas
  Ignoring pip: markers 'python_version < "3"' don't match your environment
  Looking in links:
  /cvmfs/soft.computecanada.ca/custom/python/wheelhouse/nix/avx2
  /cvmfs/soft.computecanada.ca/custom/python/wheelhouse/nix/generic
  /cvmfs/soft.computecanada.ca/custom/python/wheelhouse/generic
  Collecting pandas,
  Collecting numpy>1.13.3 (from pandas)
  [...]
  Successfully installed numpy-1.18.4 pandas-1.0.3 python-dateutil-2.8.1 pytz-2020.1 six-1.15.0

The third and final step is to install gurobipy into the environment:

(env_gurobi) [name@server ~] $ cd $EBROOTGUROBI
(env_gurobi) [name@server ~] $ python setup.py build --build-base /tmp/${USER} install
  running build
  running build_py
  creating /tmp/name
  creating /tmp/name/lib
  creating /tmp/name/lib/gurobipy
  copying lib/python3.7_utf32/gurobipy/__init__.py -> /tmp/name/lib/gurobipy
  copying lib/python3.7_utf32/gurobipy/gurobipy.so -> /tmp/name/lib/gurobipy
  running install
  running install_lib
  creating /home/name/env_gurobi/lib/python3.7/site-packages/gurobipy
  copying /tmp/name/lib/gurobipy/gurobipy.so -> /home/name/env_gurobi/lib/python3.7/site-packages/gurobipy
  copying /tmp/name/lib/gurobipy/__init__.py -> /home/name/env_gurobi/lib/python3.7/site-packages/gurobipy
  byte-compiling /home/name/env_gurobi/lib/python3.7/site-packages/gurobipy/__init__.py to __init__.cpython-37.pyc
  running install_egg_info
  Writing /home/name/env_gurobi/lib/python3.7/site-packages/gurobipy-9.0.1-py3.7.egg-info

Using the virtual environment

Python scripts can now import both Pandas and Gurobi:

[name@server ~] $ cat my_gurobi_script.py
  import pandas as pd
  import numpy as np
  import gurobipy as gurobi
  from gurobipy import *
  etc

Once created we can activate our Gurobi environment at any time with

[name@server ~] $ module load gurobi/9.0.1
  source ~/env_gurobi/bin/activate
  (env_gurobi) python my_gurobi_script.py

Note that we now use python instead of gurobi.sh! This is an example job script that we can use:

File : gurobi-py_example.sh

#!/bin/bash
#SBATCH --time=0-00:30        # time limit (D-HH:MM)
#SBATCH --cpus-per-task=1     # number of CPUs to use
#SBATCH --mem-per-cpu=1000M   # memory per CPU (in MB)

module load StdEnv/2016.4
module load gurobi/9.0.1

source ~/env_gurobi/bin/activate

echo "Threads ${SLURM_CPUS_ON_NODE:-1}" > gurobi.env   # set number of threads

python  my_gurobi_script.py

Using Gurobi with Java

To use Gurobi with Java, you will also need to load a Java module and add an option to your Java command in order to allow the Java virtual environment to find the Gurobi libraries. A sample job script is below:

File : gurobi-java.sh

#!/bin/bash
#SBATCH --time=0-00:30        # time limit (D-HH:MM)
#SBATCH --cpus-per-task=1     # number of CPUs (threads) to use
#SBATCH --mem=4096M           # memory per CPU (in MB)

module load java/14.0.2
module load gurobi/9.1.2

java -Djava.library.path=$EBROOTGUROBI/lib -Xmx4g -jar my_java_file.jar

Using Gurobi with Jupyter notebooks

Various topics can be found by visiting Resources, then clicking Code and Modeling Examples and finally Optimization with Python – Jupyter Notebook Modeling Examples. Alternatively visit support.gurobi.com and search on Jupyter Notebooks.

A demo case of using Gurobi with Jupyter notebooks on our systems can be found in this video recording, i.e. at time 38:28.

Cite Gurobi

Please see How do I cite Gurobi software for an academic publication?