Hi Mark,
it is funny that you should ask about that now as I only last week found
some Matlab commands that allow for at least some degree of
parallelization, with or without changing code. You may want to look
into what
doc matlabpool % and/or
doc spmd % and/or
doc parfor
have to offer, all of which, however, require the parallel processing
toolbox. On my machine with a recent multicore processor, I am getting
substantially faster processing with matlabpool(6) and when truly
parallelizing a loop, parfor is tremendously faster than a simple for
loop (as several loops are run simultaneously). spmd is a tad
inconvenient as the output of a function thus called is in a new
(composite) format but the other ones may help, if Matlab is aware of
several cores. You can also set up a distributed processing environment
using matlabpool, but I have not tried that. For the record, I would
avoid option 1.
Cheers,
Marko
Mark Drakesmith wrote:
> Hi
>
> I am looking to implement DATREL for warping a large number of subjects
> (250 for now, but possibly going up to 750!). I am working on a HPC
> cluster. the pre-processing steps (coreg, segmentation, etc) can be
> performed quite easily (processing on one node per subject). Calculating
> the templates and warp fields however I'm not sure about as you need to
> specify all subject scans in the command, so I can't really split the
> task across multiple processors.
>
> I was wondering which of the following would be my best option:
>
> 1 I could compute the templates for a subset of the data and then, warp
> subsequent subjects to the template. If i can get away with this I could
> apply the warping to subsequent subjects in parallel. IN which case, how
> many subjects would i need to get an accurate template?
>
> 2. If if the above is not feasible, Is there any way i can parallelise
> DARTEL so I can use multiple nodes. Even if there just some parts of the
> code i can run on multiple nodes that would be helpful. One way i
> thought of doing this would be to compute warps and templates for groups
> of about 20 subjects per node, and then (somehow) combine these together
> to create a warp for the entire cohort.
>
> Any advice or help you can offer in these issues would be very helpful
>
> Many thanks
>
> Mark
>
> --
>
> Mark Drakesmith
> Research Associate
>
> Caridff University Brain Research Imaging Centre (CUBRIC)
> School of Psychology
> Cardiff University
> Park Place
> Cardiff
> CF10 3AT
>
> Tel: +44 (0) 29 2087 0354
> Fax: +44 (0) 29 2087 0339
>
--
____________________________________________________
PD Dr. med. Marko Wilke
Facharzt für Kinder- und Jugendmedizin
Leiter, Experimentelle Pädiatrische Neurobildgebung
Universitäts-Kinderklinik
Abt. III (Neuropädiatrie)
Marko Wilke, MD, PhD
Pediatrician
Head, Experimental Pediatric Neuroimaging
University Children's Hospital
Dept. III (Pediatric Neurology)
Hoppe-Seyler-Str. 1
D - 72076 Tübingen, Germany
Tel. +49 7071 29-83416
Fax +49 7071 29-5473
[log in to unmask]
http://www.medizin.uni-tuebingen.de/kinder/epn
____________________________________________________
|