MSMAccelerator

We believe running MD should be faster and easier. If you agree, MSMAccelerator is for you!

MSMAccelerator is a tool for abstraction, allowing the researcher to worry about what systems they want to run, and how to sample those systems most efficiently, rather than worry about resource or data management.

Overview

MSMAccelerator is an adaptive sampling engine for Markov state model (MSM) powered molecular dynamics simulations.

The basic idea is to interlace simulation and analysis together -- MSMAccelerator automates the process of building models, identifying the undersampled regions of phase space, shooting new simulations, and then reanalyzing the data.

The MD simulations can be done remotely with heterogeneous compute hardware. MSMAccelerator connects to the compute nodes using WorkQueue (http://nd.edu/~ccl/software/workqueue/), a scalable Master/Worker framework. The only requirement is that the compute nodes be able to communicate with the head node via an open port. MSM construction is done on the "head" node by the MSMAccelerator executable.

MSMAccelerator currently interfaces with Gromacs, Amber, OpenMM. Support for NAMD is in the works. Adding support for your favorite MD package is easy, and requires minimal modification of the code.

Quick Start

So you want running MD to be as painless as possible, eh? Here are the steps you need to take

• Install MSMBuilder, WorkQueue, MSMAccelerator (see below)
• Acquire a PDB file specifing what protein you want to run
• Set up a project.yaml file. See tutorial/project.yaml for an example. We've included some other examples in examples/. Likely one will fit your needs.
• Find computation resources. On each cluster you want to run, install cctools and your MD package of choice. Instructions and some helpful scripts are in cluster-tools.
• Start up a master instance: MSMAccelerator -p myproject.yaml
• Start up your workers on a remote cluster: pbs_submit_workers
• Kick back and relax. You'll be getting MSMs, build automatically from optimally run MD simulations shortly. All data will be automatically retrieved and organized on your master node.
• Publish. Yeah, we said it.

Architecture

MSMAccelerator is centered around two components. First, the WorkQueue master/worker framework is the bridge between the command and control processes and the worker nodes running simulation. MSMAccelerator submits "jobs" to a queue, WorkQueue handles sending them to remote workers, running the command(s) and then returning the MD trajectories. When trajectories are returned, MSMAccelerator is triggered and begins data analysis.

The second major component is a MySQL database that stores (paths to) trajectories, models, and metadata associated with the running set of simulations. The various portions of MSMAccelerator largely stay in sync by coordinating via the database. Using MySQL is a bit of a pain compared to SQLite, but (I think) required. The key reason is that MSMAccelerator is a threaded application in which both threads need to make write operations, and SQLite has no support for this type of concurrency.

There are three main modules. First is QMaster, which controls the WorkQueue instance, adding jobs to the queue and retrieving them from the queue. The second is Brain, which creates the jobs, and the third is Builder, which builds the Markov state models.

In addition to these components, there is a core Project class that loads the configuration file (project.yaml) and initiates the database connection. The sampling module contains the adaptive sampling logic.

MSMAccelerator abstracts out the difference between MD engines by using a set of python scripts called drivers. Each MD engine is associated with a driver that is responsible for taking a small number of command line options (PDB file, forcefield name, etc), and then setting up and running the simulation. Details like the temperature, type of thermostat, etc, should be set inside the driver file and are thus invisible to the rest of MSMAccelerator.

Dependencies

MSMAccelerator depends on

• MSMBuilder (http://www.github.com/simtk/msmbuilder)
• WorkQueue (http://nd.edu/~ccl/software/workqueue)
• specific python modules (see below)

MSMAccelerator is written in python2.7. We recommend the Enthought python distribution (EPD), which comes with pre-compiled binaries for most of the dependencies. EPD is free for academics (http://www.enthought.com/).

WorkQueue and its python bindings are required, and are not always trivial to install. WorkQueue is distributed as a part of "CCTools", and we use a forked version of the CCTools package, which is hosted at

https://github.com/rmcgibbo/cctools-3.4.2-fork

See the install instructions in that package for details on compiling it.

For the database component, we use MySQL, mysql-python and SQLAlchemy. Either a dedicated MySQL server or a running instance on your local machine is required.

mysql-python does not ship by default with EPD, but can be installed with

easy_install mysql-python

On a fresh Ubuntu 11.10 instance, I used the following commands to get mysql setup (with python bindings)

sudo apt-get install mysql
sudo apt-get install libmysqlclient-dev
easy_install mysql-python

If you're not using EPD python, the python packages that you'll need to install include, but may not be limited to

• numpy
• scipy
• tables
• MSMBuilder
• SQLAlchemy
• mysql-python