Development of a subcycling option for the dynamics in the CAM spectral atmospheric dycore System Model
Run are posted immediately to this site for analysis and are not necessary correct or validated.
Talk to Kate (info below) if you have questions about the runs below.
Mark Taylor, SNL, Kate Evans, Pat Worley, and Jim Hack, ORNL, Dave Williamson, John Truesdale and Julie Caron, NCAR
Experimental setup use T85 F compset as baseline for comparisons for testing
1. run T85 with dtime=600, nsplit=1 for sanity check. Model runs are bit-for-bit and climates are identical. The run with nsplit is about the same cost as the version without.
Timings: dtime=600,nsplit=1: CCSM SYPD = 13.6
2. run T85 with dtime=600, nsplit=3 2 years T85 with and without subcycling, dtime=600 and nsplit = 3 in the 'odd' case
Timings: dtime=600,nsplit=3: CCSM SYPD = 6.3
3. run T85 with dtime=600, nsplit=4 2 years T85 with and without subcycling, dtime=600 and nsplit = 4 in the 'even' case
Timings: dtime=600,nsplit=4: CCSM SYPD = 6.7* (continuation run, so slightly higher)
4. run T85 with dtime=1200, nsplit=2 (so physics timestep matches the default run with no splitting) 2 years T85 with and without subcycling, dtime=1200 and nsplit = 2 in the 'even' case
5. run T85 with dtime=1800, nsplit=3 (so dyn timestep=600 as before) 2 years T85 with and without subcycling, dtime=1800 and nsplit = 3 in the 'odd' case
Timings: dtime=1800,nsplit=3: CCSM SYPD = 19.6
6. run T85 with dtime=1800, nsplit=4 (so dyn timestep is smaller than before) 2 years T85 with and without subcycling, dtime=1800 and nsplit = 4 in the 'even' case
Timings: dtime=1800,nsplit=4: CCSM SYPD = 17.9
Compare HOMME 1 degree with and without subcycling to evaluate climate differences
dtime=1800,nsplit=6, year HOMME ~1 degree preindustrial run versus corresponding run with no subcycling, dtime=300, nsplit=1
Apply to T341 F compset as a comparison
1. run T341 with dtime=150s, nsplit=1 to match original model
Timings: dtime=1800,nsplit=1: 62 days, CCSM SYPD=1.07
Quick ncview of the run shows 'ringing' effects near strong convection events as before.
2. run T341 with dtime=1800s, nsplit=15 (with npslit=12 to match original d_dyn=150s, model 'blew up'), 62 days, CCSM SYPD=2.27
Quick ncview of the run does not show 'ringing' effects near strong convection
events as before, athough we still see wavelike behavior near strong topography gradients.
For Additional Information:
Computational Earth Sciences Group
1 Bethel Valley Road
P.O. Box 2008, MS 6016
Bldg. 5600, C315
Oak Ridge, Tennessee 37831-6015
Phone: (865) 576-6517
Fax: (865) 576-5491
Email: evanskj at ornl dot gov