Skip to content

Latest commit

 

History

History
161 lines (92 loc) · 5.08 KB

build-openmm.md

File metadata and controls

161 lines (92 loc) · 5.08 KB

OpenMM Installation Instructions

Installation using conda

The easiest and recommended way to install OpenMM is using the conda package manager. For that, install Miniconda in your user account.

wget https://repo.anaconda.com/miniconda/Miniconda3-*.sh
bash Miniconda3-*.sh

Append channel conda-forge to the list of channels:

conda config --append channels conda-forge

Separate environments are recommended for large packages that have many dependencies:

conda create --name omm
conda activate omm

Then install OpenMM

conda install openmm

Test the installation

python -m openmm.testInstallation

Check that cudatoolkit has been installed too if you have access to NVIDIA GPUs.

Accessory packages and tools

You may want to install accessory packages, for example mdtraj, openmm-plumed.

It may be necessary to request specific versions python or of other dependencies for compatibility. You can find that out:

conda search <package> --info

In that case reinstall python and then install the packages.

conda install python=3.9

Installing from source

This is a more complex procedure but can give more control on versions of libraries or other dependencies. Not justified unless it is impossible to install OpenMM out-of-the-box via conda.

Finding the version of the CUDA toolkit

GPU computing is a quickly evolving field and navigating among all the versions of cards, their drivers and compilers is essential. Locate the root of the CUDA installation and the CUDA compiler nvcc:

which nvcc
/usr/local/cuda/bin/nvcc --version

The command

nvidia-smi

shows important information on NVIDIA GPU hardware present. In computing centers this doesn't work on frontal or submission machines because they don't have GPUs for calculations. You need to find this information otherwise or, if possible, ssh into the computing nodes with GPUs to check.

Compilation of OpenMM on the PSMN computing center

  1. Install miniconda3, with numpy and not much else, since the PSMN machines already have most of the tools needed by OpenMM (SWIG, Doxygen, etc.) This will provide a local python installation for OpenMM. Activate conda if not by default:

     conda activate

    The version of cmake on the PSMN is too old for OpenMM 7.7, so install cmake also under conda.

  2. Download the OpenMM source code from https://github.com/openmm/openmm/releases/tag/7.7.0 and expand into a src dir; create a build dir.

     mkdir ~/src
 cd ~/src
 tar xzvf openmm-7.7.0.tar.gz
 mkdir build_openmm
 cd build_openmm

  3. Configure the build

    In the build_openmm dir do:

     ccmake ../openmm-7.7.0

    press 'c' to configure. Set the installation dir (replacing user) and CUDA compiler, also check that your local python was found (if not go back to 1.):

     CMAKE_INSTALL_PREFIX=/home/<user>/openmm
 PYTHON_EXECUTABLE=/home/<user>/miniconda3/bin/python

    then press 'g' to generate the build files.

  4. (This is outdated since end of 2022) Patch the source code to use sm_70. Edit the file openmm-7.6.0/platforms/cuda/src/CudaContext.cpp and add the two lines major = 7;, minor = 0; around line 227. Don't forget the end-of-line ';' this is C++.

     if (cudaDriverVersion < 8000) {
     // This is a workaround to support Pascal with CUDA 7.5.  It reports
     // its compute capability as 6.x, but the compiler doesn't support
     // anything beyond 5.3.
     if (major == 6) {
         major = 5;
         minor = 3;
     }
 }
 major = 7;
 minor = 0;
 gpuArchitecture = 10*major+minor;
 computeCapability = major+0.1*minor;

    Alternatively, you can apply the patch I created,

     cd ~/src
 patch -p0 < openmm-sm_70.patch

  5. Compile and install,

     make -j 12
 make install
 make PythonInstall

  6. Test the installation (it won't find a CUDA device on the frontal machine),

     python -m openmm.testInstallation

  7. In your submission script (to qsub) you may need to add

     export LD_LIBRARY_PATH=$HOME/openmm/lib

    so that OpenMM libraries are found.

That's it. You can test the installation as indicated above. These build instructions should also work with OpenMM 7.4.2.

TGNH Drude thermostat (OpenMM 7.4.2)

The temperature-grouped Nosé-Hoover thermostat (http://doi.org/10.1021/acs.jpclett.9b02983) is a plugin to OpenMM 7.4.2. The authors are in the process of upgrading it to the latest OpenMM but gave no estimated date.

  1. Download or git clone the plugin from https://github.com/scychon/openmm_drudeNose

  2. Configure

     mkdir build_drudeNose
 cd build_drudeNose
 ccmake ../openmm_drudeNose

    Set the following (change user):

     CMAKE_INSTALL_PREFIX=/home/user/openmm
 DRUDENOSE_BUILD_OPENCL_LIB=OFF
 OPENMM_DIR=/home/user/openmm
 PYTHON_EXECUTABLE=/home/user/miniconda3/bin/python

    Press 'c' and 'g'.

  3. Build and install

     make
 make install
 make PythonInstall

Done.