Note: this repository is based on an old version of fsl_sub. Users should consider the recent fsl_sub release.
FSL is a popular tool for brain imaging. If you set it up on your laptop or desktop, many of the commands will take a long time to complete as they are run on a single CPU at a time. Big data centers typically set up a Grid Engine to run FSL tasks in parallel, but this is complicated and not possible on all operating systems (e.g. macOS). By simply installing this tiny file, FSL will run on parallel on any desktop or laptop. Specifically, any task that would have used SGE if it was available will be spawned to all your available CPUs. This solution is described in more detail on my optimizing FSL web page. Note that only some stages of FSL are designed to run in parallel (FEAT, MELODIC, TBSS, BEDPOSTX, FSLVBM, POSSUM). Therefore, this script may not always be the optimal way to accelerate FSL. Specifically, if you are using FEAT for fMRI analyses, you should be aware that only FLAME will be accelerated by the fsl_sub script provided here. An alternative to accelerating the processing of each individual would be to run several individuals in parallel. Further, while tools like bedpostx, eddy, and probtrackx may be able to take advantage of multiple CPUs, they will be dramatically faster using a CUDA-enabled graphics card.
28-March-2020
- Updated to correspond with FSL 6.0.3.
29-Jan-2017
- Example dataset included.
7-May-2017
- Explicitly use BASH shell, which is not default for Debian. Note: if you use FSL it is probably a good idea to set BASH as your default shell, as several FSL scripts will fail with other shells (e.g. DASH).
30-August-2016
- do not run "GPU" code in parallel
15-May-2015
- Not all versions of sh support "declare"
3-March-2015
- Initial version
Replace your previous version of fsl_sub with this code. Tasks will automatically run in parallel. SGE will take precedence over this code, so if you ever upgrade to a SGE cluster you do not have to modify this function
cd $FSLDIR/bin
cp fsl_sub fsl_sub_orig
sudo cp ~/Downloads/fsl_sub/fsl_sub fsl_sub
sudo chmod o+rx fsl_sub
Replace the copy of fsl_sub that came with FSL with the version included here. To find the location of the original fsl_sub type which fsl_sub from the command line. You will want to make sure that the version of fsl_sub is executable (sudo chmod +x fsl_sub).
After installation the code will automatically accelerate parallel processes. However, you can also control its behavior:
- If you want to test the benefit, you can temporarily disable the function by using the command “FSLPARALLEL=0; export FSLPARALLEL”
- If you want to test the benefit, you can temporarily force it to use precisely 8 cores with the command “FSLPARALLEL=8; export FSLPARALLEL”
- If you want to test the benefit, you can temporarily force it to automatically detect the number of cores (the default behavior) with the command “FSLPARALLEL=1; export FSLPARALLEL”
- To make permanent changes, add the desired FSLPARALLEL setting to your profile.
If you want to test the effectiveness of parallel processing on your computer, un-compress the folder "bedpostTest.zip" and run the shell script "runme_quick.sh" - this will FSL's Bedpost twice: once with parallel processing off and once with it on. Notice that FSLPARALLEL will accelerate any FSL program that is designed to use SunGridEngine to go run CPU tasks in parallel. Note that in the specific case of bedpost, serious users will want want to use the GPU-based bedpostx_gpu instead of the CPU-based bedpostx (e.g. the included script runme_gpu.sh). Also, the sample dataset only includes an MRI scan with six low-resolution slices (with little benefit from more than six threads), while real world datasets include dozens of high-resolution slices. The purpose of this test dataset is simply to test the installation of the parallel fsl_sub.
./runme_quick.sh
fsl_sub running with parallel processing disabled
...
real 19m59.990s
...
fsl_sub running with parallel processing enabled
...
real 8m54.404s
This software uses the FMRIB Software Library license, which is embedded in the source code.