Obtaining a Benchmark for Mantid Fitting

There are several scenarios in which obtaining a benchmark for the accuracy and runtime of Mantid minimizers might be useful:

  • From the perspective of a Scientist, who wants to know the best Mantid minimizer to use for fitting their model to their data.

  • From the perspective of a Software Developer, who wants to optimize their code to improve its accuracy and runtime.

This page will explain how to benchmark the accuracy and runtime of different fitting minimizers in Mantid for different fitting problems. To do this, we will use a package called FitBenchmarking which is a cross-platform open source tool for comparing different minimizers and fitting frameworks.

Before you continue, make sure you have a recent version of python installed, and access to git bash (which is the recommended terminal to use).

Benchmarking Mantid from Source

This section will show you how to benchmark Mantid using a source code directory. This is useful for software developers who want to make quick changes to the code to see how they affect the fit accuracy.

  1. Open a git bash terminal and cd to the desired location.

  2. Activate your mantid conda environment.

  3. Pip install the FitBenchmarking package by following the FitBenchmarking installation instructions.

  4. Follow the instructions in the Running a Benchmark section below. Note that you must specify the PYTHONPATH in the last command, where the <config> is only required on Windows:

PYTHONPATH=/path/to/build/bin/<config> fitbenchmarking -o fitting_options.ini -p examples/benchmark_problems/Muon

Benchmarking Mantid from Install

This section will show you how to benchmark Mantid using an installed version. This is useful when you want a more realistic benchmark for fit runtimes.

Initial Setup

  1. Download and install the desired version of MantidWorkbench from the releases page. Note its installation path, hereby denoted as [INSTALL_PATH].

  2. Open a git bash terminal and cd to the desired location.

  3. Pip install the virtual environment package as described in the ‘Installing virtualenv’ section.

Creating your Benchmarking Environment

  1. From the same git bash terminal, run the following command:

If using Windows:

py -m virtualenv -p=[INSTALL_PATH]\bin\python.exe benchmark-env

This will create a virtual environment based on the python executable provided with your Mantid installation. For Windows, the location of this executable is C:\MantidInstall\bin\python.exe.

If using Ubuntu:

virtualenv --python=python3 benchmark-env

In this example, benchmark-env is the name given to your virtual environment.

  1. Activate your environment. This is explained in the ‘Activating a virtual environment’ section.

  2. It might be necessary to upgrade and install a few packages before installing FitBenchmarking if working from a Linux system:

python3 -m pip install --upgrade pip
python3 -m pip install --upgrade pillow
sudo apt-get install libglu1-mesa

  1. Install the FitBenchmarking package by following the FitBenchmarking installation instructions. Note that installing FitBenchmarking from source using the editable flag -e proved to be the most stable installation prior to the release of FitBenchmarking v0.2.

Your environment should now be ready for performing a benchmark of Mantid minimizers.

Running a Benchmark

The process for how to run a benchmark is explained extensively in the FitBenchmarking documentation, and so I recommend you give it a read. This section will give a basic example of how to perform a simple benchmark of three Mantid minimizers.

  1. Create a file called fitting_options.ini with the following contents

[FITTING]

software: mantid

num_runs: 1

[MINIMIZERS]

mantid: Levenberg-Marquardt
        Levenberg-MarquardtMD
        Simplex

[PLOTTING]

make_plots: yes

[LOGGING]

external_output: log_only

  1. Download the examples folder from the FitBenchmarking github repo by git cloning the repository. Alternatively, you can define your own fitting problems.

  2. From your activated virtual environment, run the following command. This will run the Muon fitting problems assuming you have the same directory structure as seen on the Fitbenchmarking repo.

fitbenchmarking -o fitting_options.ini -p examples/benchmark_problems/Muon

If benchmarking from source, you must also specify the PYTHONPATH. The PYTHONPATH needs to point to your build/bin/<config> folder where the <config> is only required on Windows and should be replaced by the config of your build e.g. build/bin/Release.

PYTHONPATH=/path/to/build/bin/<config> fitbenchmarking -o fitting_options.ini -p examples/benchmark_problems/Muon

When the benchmark is complete, it should open a browser which contains the results. You should read the FitBenchmarking documentation if you need help with how to interpret these results. The results will also be stored in your current folder location.

Tips

  • Make sure your git bash terminal is open in the correct location and has the virtual environment activated when running your benchmark.

  • Each time your run the benchmark, the old results will be overwritten unless you change the directory you run the fitbenchmarking command from. In later versions of FitBenchmarking (>v1.5) there will be an option to specify the results directory on the command line or via the .ini file.

  • Be aware that an ‘Unexpected Exception’ can sometimes occur when running the fitbenchmarking command after installing it from source without the editable flag -e.