 |
 |
Hi - it seems unlikely that it makes sense to be working at 1mm standard space - switching to 2mm will save a lot of RAM? Cheers > On 7 Oct 2015, at 17:26, Noah Mercer <[log in to unmask]> wrote: > > > Ah, right - sorry I didn't include that info earlier: > > timepoints/volumes = 210 > voxels = 182 x 218 x 182 > > Does this level of memory usage seem reasonable given those numbers? > > > We've actually already done the guesstimation and set things up so that new jobs won't run if the RAM needs are too high, which leads to a very small number of jobs running in parallel and longer runtimes - hence the desire to decrease RAM utilization if possible. > > And speaking of long runtimes, we've noticed that in stage D the number of EVs makes an enormous difference in runtimes. If we have 5000 permutations and just 2 EVs stage D completes in about an hour, but if we go up to 4 EVs it will run for (so far) more than 20 hours without completing. That surprises me - can anyone say whether that's to be expected and if not suggest potential problems to look for? > > Thanks, > Noah > > > From: FSL - FMRIB's Software Library <[log in to unmask]> on behalf of Stephen Smith <[log in to unmask]> > Sent: Tuesday, October 6, 2015 4:37 AM > To: [log in to unmask] > Subject: Re: [FSL] memory utilization in dual regression analysis > > Hi - you didn't say how many timepoints / voxels you had so I can't comment on whether this is normal. If really necessary you could downsample the data spatially if you have RAM problems - though a decent cluster node should in general allow for at least 16GB per core RAM ideally. > > Make sure that the data is sensibly masked, as that will help reduce memory needs (it probably is already masked if you used standard FSL preproc). > > You can hopefully get your sysad to make sure the queues are setup such that new jobs won't get run if the total RAM needs are too high - you will proabbly need to guesstimate (empirically) how much an individual job needs in order for that functionality to work. > > Cheers. > > > >> On 5 Oct 2015, at 20:33, Noah Mercer <[log in to unmask] <mailto:[log in to unmask]>> wrote: >> >> >> We're running a dual regression analysis with a design that has 2 groups (n=12 each), 20 components and 4 covariates (i.e., age, education, gender, ICV) . The actual command line looks like this: >> >> dual_regression melodic_IC.nii.gz 1 design.mat design.con 5000 output_directory 24subjects_preprocessed_data.nii.gz >> >> This is running in a Torque-based grid environment and we're seeing memory usage of ~12GB fsl_maths process in stage A and up to ~20GB per fsl_glm process in stage C. Given the details of our particular setup this means that we are memory bound and can only run 3 stage A processes simultaneously and only 2 stage C processes simultaneously. All of this leads to a few questions: >> >> 1. Does this memory utilization seem within the expected range? >> >> 2. Does anyone have any suggestions on how to reduce the memory utilization (which would allow us to improve parallelization)? >> >> 3. If not, does anyone have any suggestions on how to reduce runtime other than by improving parallelization? (New/more hardware is not an option at the moment, unfortunately.) >> >> Thanks, >> >> Noah >> >> > > > --------------------------------------------------------------------------- > Stephen M. Smith, Professor of Biomedical Engineering > Head of Analysis, Oxford University FMRIB Centre > > FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK > +44 (0) 1865 222726 (fax 222717) > [log in to unmask] <mailto:[log in to unmask]> http://www.fmrib.ox.ac.uk/~steve <http://www.fmrib.ox.ac.uk/~steve> > --------------------------------------------------------------------------- > > Stop the cultural destruction of Tibet <http://smithinks.net/> > > > > > --------------------------------------------------------------------------- Stephen M. Smith, Professor of Biomedical Engineering Head of Analysis, Oxford University FMRIB Centre FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK +44 (0) 1865 222726 (fax 222717) [log in to unmask] http://www.fmrib.ox.ac.uk/~steve <http://www.fmrib.ox.ac.uk/~steve> --------------------------------------------------------------------------- Stop the cultural destruction of Tibet <http://smithinks.net/>