You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
## DESCRIPTION OF CHANGES:
Add the ability for forecast jobs on Jet and Hera to use resources more efficiently, and to allow for Jet jobs to run on more partitions, which will be helpful during the HFIP Allocation Season. This involved additional Slurm tags in the rocoto xml, and changing pre-defined grids' layouts and blocksize to be consistent with
Other changes:
- Converted nthreads to a configurable option in model_configure
- Updated layout and blocksize for GSD_HRRR25km pre-defined grid.
- Changed dependency for post jobs to also allow for completion of run_fcst to trigger jobs.
- Reduce the number of tries for all tasks to 1. Anything higher is wasteful of resources and time, especially when there are forecast tasks known to fail in the test suite.
The concept adopted here is based on tests performed by Sam Trahan on the AVID real-time setup of HRRR. He found that a more efficient way to run the forecast model is by using the following settings:
<partition>kjet,xjet,sjet,vjet</partition>
<cores>240</cores>
<native>--cpus-per-task 4</native>
<native>--exclusive</native>
along with `OMP_NUM_THREADS=4`
To adopt these settings in a more general way, he also provided a two rules to follow in order to avoid crashes when choosing layouts and blocksize:
blocksize <= grid_size_y / layout_y / 2
blocksize >= grid_size_x / layout_x / 2 \
## TESTS CONDUCTED:
I ran the test suite (except for nco tests) on both Jet and Hera. Applying the above techniques, I fixed a couple of the tests that were already failing for develop -- regional_003, regional_004, and new_JPgrid. (After merging with develop, I realize that new_JPgrid may have already been fixed)
## CONTRIBUTORS:
@SamuelTrahanNOAA, Dom Heinzeller, and several colleagues in GSL/ATD.
* Add Jenkinsfile that includes a build and test pipeline, which
leverages the unified build and test scripts
* Supported platforms are Cheyenne, Gaea, and Orion
* Supported compilers are GNU and Intel
Description
Update CMEPS to the latest emc/develop branch. This update contains a bugfix for ifrac mapping from ice->atm in both nems_orig and nems_frac mode.
Related to
Dependent on CMEPS PR #22
The text was updated successfully, but these errors were encountered: