This repository has been archived by the owner on Oct 23, 2020. It is now read-only.
Ocean irreproducible on Cori-KNL/Intel #1519
Comments
|
Further diagnosis by @greenrongreen pointed to a data alignment issue. Am testing and will submit a PR soon with a fix to compiler options. |
|
Added -fp-model consistent flag fixes this and is in PR #1520 |
|
Replaced by |
|
Sorry, meant to close related PR not issue |
mark-petersen
added a commit
to MPAS-Dev/MPAS-Model
that referenced
this issue
May 14, 2018
added -align array64byte flag to optimized Intel builds to fix reproducibility issues on some machines (KNL) also replaced deprecated -FR flag with -free to denote free-form source B4B Tested on Cori, Edison at NERSC Replaces previous PR in MPAS-Dev/MPAS MPAS-Dev/MPAS#1520 though see discussion therein. Fixes previously reported bug in MPAS-Dev/MPAS MPAS-Dev/MPAS#1519
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
During a QU240 test case on Cori-KNL, MPAS ocean model was found to be irreproducible in a pure MPI configuration when built using Intel at -O2 or higher. Namely, two identical 8-core jobs submitted one after the other generated non-b4b differences in the output file.
srun -n 8 ./ocean_model
mv output.nc orig.nc
srun -n 8 .ocean_model
nccompare output.nc orig.nc finds significant differences but should be b4b
The text was updated successfully, but these errors were encountered: