Parallelizing FSL – without the pain

FSL is a very popular analysis suite for neuroimaging data. Unfortunately, right out of the box FSL’s routines do not make good use of today’s multi-core machines. Consequently, researchers using FSL on individual workstations have difficulties harnessing the computing power of their equipment.

However, it is not the case that FSL doesn’t support modern computing environments. It is actually one of very few packages that come with built-in support for running computations on a cluster back-end. Unfortunately, it relies on the GridEngine software (formerly Sun, now from Oracle) to be installed locally, and configured properly. This is a task that is known to be non-trivial and can cause serious headache in a neuroscientist. While FSL has been ported to use other batch systems, such as Torque, none of them is much easier to deploy.

What would be helpful is a batch system that can be easily installed without complicated configuration procedures. Preferably, such a system is powerful enough to scale across multiple machines, whenever the demand for computing power increases. At NeuroDebian we looked at a number of projects and found Condor to be the most versatile tools, yet still straightforward to use in a single-machine scenario.

Over the past two years we have been working on getting Condor packaged for Debian. We are now very proud to announce that, a few days ago, Condor has been accepted as an official Debian package (in Debian experimental for the time being). NeuroDebian, as usual, provides backports of Condor for numerous Debian and Ubuntu releases.

With the availability of Condor packages, parallelizing FSL on a local workstation is now possible by following a simple four-step procedure.

1. Configure your machine to use the NeuroDebian repository

[ if you are already a NeuroDebian user you can skip this step ]

Go to http://neuro.debian.net and follow the setup instructions given on the front-page. This means selecting your Debian or Ubuntu release, select a package repository mirror close to you, and run the little shell script snippet that is presented to you. After having updated you package lists you are ready to go.

If you are not running Debian or Ubuntu on your machine, you can also install the NeuroDebian virtual machine which is able to utilize multiple CPU-cores. The remaining setup steps apply to VM users as well.

2. Install FSL

[ if you already have the latest FSL package version, you can skip this too ]

Open a terminal and install FSL being running the command:

% sudo apt-get install fsl

Of course you can also use any graphical package manager front-end (e.g. software center). The terminal commands are simply used for clarity and simplicity.

Check what version of FSL got installed. Any version from 4.1.9-2 onwards is fine. If you got an older version, you either haven’t succeeded doing step 1, or you are running a version of Debian or Ubuntu that is too old, and you would need to upgrade it first.

3. Install Condor

Install Condor with:

% sudo apt-get install condor

Any Condor package version from 7.7.4-1 onwards comes with all necessary features. During installation a few question will be asked. You should enable the Debconf-based configuration and ask for a “Personal Condor” installation (see screenshots).

../../_images/condor_install1.png ../../_images/condor_install2.png

Now you have a fully functional Condor installation that advertises one compute slot per detected CPU-core on your machine. You can display the current status with the condor_status command. On a dual-core machine this could look like this:

% condor_status

Name               OpSys      Arch   State     Activity LoadAv Mem   ActvtyTime

slot1@meiner       LINUX      INTEL  Owner     Idle     0.320  1972  0+00:00:04
slot2@meiner       LINUX      INTEL  Owner     Idle     0.000  1972  0+00:00:05
                     Total Owner Claimed Unclaimed Matched Preempting Backfill

         INTEL/LINUX     2     2       0         0       0          0        0

               Total     2     2       0         0       0          0        0

By default, Condor will only run jobs when the machine is idle. If you want to run jobs immediately when they are submitted you need to change the configuration slightly. To do this run:

% sudo dpkg-reconfigure htcondor

Note

26/06/2015 updated for condor to htcondor package rename

This will again ask a number of questions. Leave everything on the respective defaults, but enable running jobs regardless of other machine activity (see screen shot below).

../../_images/condor_install3.png

4. Enjoy parallelized FSL

Now you are ready to make use of Condor with FSL. Source the configuration from /etc/fsl/fsl.sh as usual:

% . /etc/fsl/fsl.sh

Whenever you want FSL to parallelize jobs through Condor, simply set the FSLPARALLEL environment variable to condor and run FSL as usual:

% export FSLPARALLEL=condor

You can check whether the setup is working by running condor_q to see the list of submitted/running jobs. Running the FEAT part of the FSL-FEEDS test suite could look like this:

% condor_q

-- Submitter: meiner : <127.0.0.1:52379> : meiner
 ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD
   1.0   michael         3/9  14:47   0+00:00:07 R  0   0.7  zsh -c /usr/share/
   2.0   michael         3/9  14:47   0+00:00:00 H  0   0.7  zsh -c /usr/share/
   3.0   michael         3/9  14:47   0+00:00:01 R  0   0.0  cluster2_sentinel.
   4.0   michael         3/9  14:47   0+00:00:00 H  0   0.7  zsh -c /usr/share/
   5.0   michael         3/9  14:47   0+00:00:01 R  0   0.0  cluster4_sentinel.
   6.0   michael         3/9  14:47   0+00:00:00 H  0   0.7  zsh -c /usr/share/
   7.0   michael         3/9  14:47   0+00:00:01 R  0   0.0  cluster6_sentinel.
   8.0   michael         3/9  14:47   0+00:00:00 H  0   0.7  zsh -c /usr/share/
   9.0   michael         3/9  14:47   0+00:00:01 R  0   0.0  cluster8_sentinel.
  10.0   michael         3/9  14:47   0+00:00:00 H  0   0.7  zsh -c /usr/share/
  11.0   michael         3/9  14:47   0+00:00:01 R  0   0.0  cluster10_sentinel

11 jobs; 0 completed, 0 removed, 0 idle, 6 running, 5 held, 0 suspended

Done

There is obviously much more that could be said about other exciting features of Condor, but we will leave this for another blog post in the future. As a little teaser we could mention that the Debian Condor package comes with an emulator of SGE’s qsub. Through this adaptor many software that can make use of SGE can also be used with Condor, for example NiPyPE

If you are now keen on exploring the possibilities of Condor, take a look at the manual that comes in the condor-doc package.

Comments

blog comments powered by Disqus