Memory issues in running GLMSingle on volumetric data

[vassiki]

Dear GLMSingle community,

We have been using the python implementation of GLMSingle for about 3 years, and it has been great! But now we have a problem: whereas in the pase we only worked with functional data in surface space, we are now trying to conduct the GLM on volumetric data. However, it seems that Linux runs out of memory and kills the code before it finishes executing. We’ve tried increasing our virtual machine’s memory as much as possible.

We are currently using version 1.2. The OS for the machine we are running GLMSingle on is Ubuntu 20.04, and we have 64 GB of RAM available. Two screenshots of the error are attached.

For context, the dimensions of each run, for one subject, are [93 x 121 x 100 x 261]. The pre-processed file is 800 MB in size. We have 10 runs total per subject. Information about our processing environment is also included in a separate screenshot.

Any suggestions for how to overcome these issues will be much appreciated!

Crash message:

OOM killer logs:

Memory details:

[kendrickkay]

Thanks for the informative description. I don't use the Python version, but a few thoughts. The easiest, of course, is to buy more RAM (if your computer has slots), or to find a computer with more RAM. Subselecting voxels of course could help, but that's annoying and hard to decide. Analyzing fewer runs at a time is another option, but that's also sub optimal and annoying.

What about the data format of your data? Can you use 32-bit floats?

Aside from that, another solution would be to revisit the code implementation of GLMsingle and think about memory optimization, but that's also a lot of work and may not necessarily garner much benefit.

[vassiki]

Thanks for the prompt response! I just tried to resave files as 32-bit float numpy arrays, but those files are still 1.7 GB in size. I also set wantfileoutputs and wantmemory parameters to [0,0,0,1] to minimize what needs to be held in memory during execution. The code still crashes as it's trying to run the Type A model. However, it seems to progress when I try to analyze fewer runs, so that may have to be the approach to take.

[kendrickkay]

I see. See how it goes. I have not myself done a within-session split of runs given to GLMsingle. One thing to potentially look out for is that the HRF selected for a voxel can/will be different across the splits that you create. Hence, there may be some discrepancy in the interpretation of the amplitude estimates that you get out from GLMsingle.