Improve GW performance efficiency with different tasks per node and layout #4141
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -271,7 +271,7 @@ case "${fv3_res}" in | |
| export WRTTASK_PER_GROUP_PER_THREAD_PER_TILE=15 #Note this should be 10 for WCOSS2 | ||
| elif [[ "${RUN}" = "gfs" || "${RUN}" = "gcafs" ]]; then | ||
| export layout_x=12 | ||
| export layout_y=15 | ||
|
There was a problem hiding this comment. Is this in combination with the There was a problem hiding this comment. I agree that some justification would be needed to have this change (on top of thorough testing since this seems to have only been tested on Ursa and Gaea). There was a problem hiding this comment. 768/15 = 51.2 so I also think this means we can't be guaranteed reproducibility. There was a problem hiding this comment. @DavidHuber-NOAA Yes, that is the thought, so we hope that more communications can be within node. There was a problem hiding this comment. @JessicaMeixner-NOAA From my understanding 768/15 = 51.2, which is not an integer, which should impact load balance, but not reproducibility. But with C768 is a pretty high-resolution, there are should be plenty of tasks to do, load-balance is probably not as important as at low-res. I guess reduce cross-node communication is more important. There was a problem hiding this comment. @dpsarmie can you weigh in here. It’s my understanding the layout chosen here will not reproduce with other layouts since it’s not divisible. Maybe things have changed? |
||
| export WRITE_GROUP=4 | ||
| export WRTTASK_PER_GROUP_PER_THREAD_PER_TILE=20 #Note this should be 10 for WCOSS2 | ||
| fi | ||
|
|
||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@GeorgeVandenberghe-NOAA had noted that the UFS did not run properly on C6 when allowed to use more than 144 tasks at high resolution. Is this issue fixed?