So, is the bottom line that when you run 7.4.1 on COS8 in parallel that you get (slightly) different results each time?
On 2/2/2024 11:20 AM, Horn, Mitchell Jacob wrote:
Hi FS Devs,
I’m experiencing unreproducible thickness results when running any 7+ version with parallelization enabled. Running recon-all without parallelization produces consistent thickness results. I’m running this in AlmaLinux8 (a library-equivalent downstream OS to CentOS8).
I’m attaching a table (table1) of 12 recons with bert:
- 3 parallelized with CentOS8-compiled 7.4.1
- 3 non-parallelized with CentOS8-compiled 7.4.1
- 3 parallelized with CentOS7-compiled 7.4.1
- 3 non-parallelized with CentOS7-compiled 7.4.1
I suspected the downstream CentOS8 libm was the culprit (because of testing I did this last 2023 summer). I ran 3 more recons parallelized with the CentOS7-compiled 7.4.1, but before running the recon-all command, set LD_PRELOAD to a copy of the CentOS7 libm libraries. The thickness results were then consistent, see the second table below (table2). I could not run this experiment on the CentOS8-compiled version, as that one is obviously not backward compatible with CentOS7 libm.
As a quick test of the OS-dependency, I submitted 3 parallel recons on MLSC with 7.3.3. Each reported different thickness. See table 3 (table3).
I’m asking if you can please confirm whether running any 7.+ version with parallelization is generating reproducible results for you in CentOS8 (or equivalent)?
P.S. - I tested 6.0 (CentOS6-compilation) with parallelization, and the results were consistent.
Best,
Mitch
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer