External Email - Use Caution        

Hi Ajay

The memory footprint of each pipeline is something I wish there were better records of!  You're right that 4GB is outdated.  

It really depends on the pipeline.  For cross-sectional recon-all, 20Gb should do it.  For longitudinal pipelines I suspect (but haven't confirmed) that the RAM footprint is related to the number of timepoints a subject has.  The infant stream needs around 32Gb.  Those are the ones I'm familiar with.

Each command should have an fs_time output with a measure of the peak memory used, though it is not super accurate.

-Paul

On Mon, Jul 12, 2021 at 4:39 PM Ajay Kurani <dr.ajay.kurani@gmail.com> wrote:

        External Email - Use Caution        

Hi Freesurfer Experts,
   I am designing some new hardware for a mini-cluster running Freesurfer and other neuroimaging toolboxes and wanted to better understand requirements from Freesurfer.

1) Is there a minimum amount of ram per core that is needed/recommended or is the amount of ram only defined per subject.  Most of the threads I found say you need a minimum of 4 GB ram but these were from a long time ago and may not apply to the OPENMP code.  Our compute node will have between 52-64 cores so I wanted to know what is the ratio for Ram to CPU (multithreading turned off) that is recommended as opposed to Ram per subject. 

2) AMD vs Intel
We will run jobs on an existing cluster that has Dual Intel Xeon 6230R and for our new minicluster we are debating between AMD and Intel.  While AMD offers higher core-count per CPU, I am worried that results could potentially differ based on the CPU/underlying libraries.  We plan on using docker images so from your perspective do you have a recommendation ?

Best,
Ajay
_______________________________________________
Freesurfer mailing list
Freesurfer@nmr.mgh.harvard.edu
MailScanner has detected a possible fraud attempt from "secure-web.cisco.com" claiming to be https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer