openfe/docs/guide/execution/quickrun_execution.rst

.. _userguide_quickrun:

Execution with Quickrun
=======================

The planning and preparation of a campaign of alchemical simulations using ``openfe`` is intended to be achievable on a local workstation in a matter of minutes.
The *execution* of these simulations however requires a large amount of computational power, and beyond running single calculations locally, is intended to be distributed across a HPC environment.
Doing this requires storing and sending the details of the simulation from the local workstation to a HPC environment, which can be done via the :func:`.Transformation.to_json` function which :ref:`creates a saved JSON version of the data<dumping_transformations>`.
These serialized JSON files are the currency of executing a campaign of simulations and contain all the information required to execute a single simulation.

To read the ``Transformation`` information and execute the simulation, the command line interface provides the ``openfe quickrun`` command, the full details of which are given in :ref:`the CLI reference section<cli_quickrun>`.
Briefly, this command takes in the ``Transformation`` information represented as JSON, then executes a simulation according to those specifications.
For example, the following command executes a simulation defined by ``transformation.json`` and produces a results file named ``results.json``.

::

  openfe quickrun transformation.json -o results.json


Executing within a job submission script
----------------------------------------

You may need to submit computational jobs to a queueing engine, such as Slurm.
The ``openfe quickrun`` command can be used within a submission script as follows:

::

  #!/bin/bash

  #SBATCH --job-name="openfe job"
  #SBATCH --mem-per-cpu=2G

  # activate an appropriate conda environment, or any "module load" commands required to
  conda activate openfe_env

  openfe quickrun transformation.json -o results.json


Parallel execution of repeats with Quickrun
===========================================

Serial execution of multiple repeats of a transformation can be inefficient when simulation times are long.
Higher throughput can be achieved with parallel execution by running one repeat per HPC job.
Most protocols are set up to run three repeats in serial by default, but this can be changed by either:

 1. Defining the protocol setting ``protocol_repeats`` - see the :ref:`protocol configuration guide <cookbook/choose_protocol.nblink>` for more details.
 2. Using the ``openfe plan-rhfe-network`` (or ``plan-rbfe-network``) command line flag ``--n-protocol-repeats``.

Each transformation can then be executed multiple times via the ``openfe quickrun`` command to produce a set of repeats.
However, **you must use unique results files for each repeat to ensure they don't overwrite each other**.
We recommend using folders named ``results_x`` where x is 0-2 to store the repeated calculations as our :ref:`openfe gather <cli_gather>` command also supports this file structure.

Below is an example of a simple script that will create and submit a separate job script (``\*.job`` named file) for every alchemical transformation (for the simplest SLURM use case) in a network running each repeat in parallel and writing the results to a unique folder:

.. code-block:: bash

   for file in network_setup/transformations/*.json; do
     relpath="${file:30}"  # strip off "network_setup/"
     dirpath=${relpath%.*}  # strip off final ".json"
     jobpath="network_setup/transformations/${dirpath}.job"
     if [ -f "${jobpath}" ]; then
       echo "${jobpath} already exists"
       exit 1
     fi
     for repeat in {0..2}; do
       cmd="openfe quickrun ${file} -o results_${repeat}/${relpath} -d results_${repeat}/${dirpath} --n-protocol-repeats 1"
       echo -e "#!/usr/bin/env bash\n${cmd}" > "${jobpath}"
       sbatch "${jobpath}"
     done
   done

This should result in the following file structure after execution:

::

    results_parallel/
    ├── results_0
    │   ├── rbfe_lig_ejm_31_complex_lig_ejm_42_complex
    │   │   └── shared_RelativeHybridTopologyProtocolUnit-79c279f04ec84218b7935bc0447539a9_attempt_0
    │   │       ├── checkpoint.nc
    │   │       ├── simulation.nc
    │   ├── rbfe_lig_ejm_31_complex_lig_ejm_42_complex.json
    ├── results_1
    │   ├── rbfe_lig_ejm_31_complex_lig_ejm_42_complex
    │   │   └── shared_RelativeHybridTopologyProtocolUnit-a3cef34132aa4e9cbb824fcbcd043b0e_attempt_0
    │   │       ├── checkpoint.nc
    │   │       ├── simulation.nc
    │   ├── rbfe_lig_ejm_31_complex_lig_ejm_42_complex.json
    └── results_2
        ├── rbfe_lig_ejm_31_complex_lig_ejm_42_complex
        │   └── shared_RelativeHybridTopologyProtocolUnit-abb2b104151c45fc8b0993fa0a7ee0af_attempt_0
        │       ├── checkpoint.nc
        │       ├── simulation.nc
        └── rbfe_lig_ejm_31_complex_lig_ejm_42_complex.json

The results of which can be gathered from the CLI using the ``openfe gather`` command, in this case you should direct
it to the root directory which includes the repeat results and it will automatically collate the information

::

 openfe gather results_parallel

Optimizing GPU performance with NVIDIA MPS
==========================================

You can further optimize execution of ``openfe quickrun`` using NVIDIA's Multi-Process Service (MPS).
See NVIDIA's documentation on `MPS for OpenFE free energy calculations <https://developer.nvidia.com/blog/maximizing-openmm-molecular-dynamics-throughput-with-nvidia-multi-process-service/?ref=blog.omsf.io#mps_for_openfe_free_energy_calculations>`_ for details.

See Also
--------

For details on inspecting these results, refer to :ref:`userguide_results`.