Major MMF2/PAM updates #6354
Major MMF2/PAM updates #6354
Conversation
…to whannah/mmf/pam-updates
|
|
The plot below shows various 1-year averages from a recent simulation on OLCF Summit. This run had an average throughput of 1.57 sypd using 128 nodes. Note that the liquid and ice water paths fields show a substantial imbalance of too little liquid and too much ice. This needs to be addressed with tuning parameters in the future. The main parameters we want to target are not available to be modified from the MMF/PAM code until the EAMXX fork is pulled into the E3SM repo. So tuning changes are expected to be the focus of a future PR.
|
|
The latest commit should address issue #6363 |
| //------------------------------------------------------------------------------------------------ | ||
| //------------------------------------------------------------------------------------------------ | ||
| //------------------------------------------------------------------------------------------------ | ||
|
|
||
| // set number of CRM steps | ||
| int nstop = int(gcm_dt/crm_dt); | ||
| int nstop = int(gcm_dt/crm_dt_fixed); |
There was a problem hiding this comment.
This has an implied floor operation. Is that what you intend if they aren't evenly divisible? Another way to handle non-evenly divisible could be:
int nstop = std::ceil(gcm_dt / crm_dt_fixed);
crm_dt_fixed = gcm_dt / nstopThere was a problem hiding this comment.
I'm actually not sure which is better or if it even matters.... so I'm just gonna leave the implied floor in there.
|
@whannah1 anything else to do here? |
whannah1
force-pushed
the
whannah/mmf/pam-updates
branch
from
d6bfec1
to
2590f22
Compare
|
@rljacob looks like there's a new problem that's been created by other updates - which I think will all be fixed when the eamxx code in this repo is updated. Not sure how to navigate this, so I might just leave this in a holding pattern for a bit. |
|
Waiting for SCREAM update #6489 and a few more changes. |
This brings in several updates and fixes for the PAM CRM used in the MMF2 compsets to improve stability and throughput. Some of these required updates to the PAM submodule, but the most important change was the addition of adaptive sub-cycling in the PAM driver. This change allowed the CRM time step to be increased to 10 seconds, which dramatically improved the throughput while also making the model much more resilient to numerically unstable conditions due to unrealistically strong velocities (still not sure why these occur in the first place).
List of notable changes for PAM and/or the driver:
Fixes #6363
[non-BFB] only for MMF2