paolino@cola_iges_org
New Member
Since I didn't receive any guidance here, I went ahead and ported my Bluevista CCSM3.0 runs over to Blueice. A recipe detailing my procedure is below. Please send your comments and suggestions.
Please note that a model run at Bluevista will not be reproducible at Blueice, at least with the procedure I've used.
After unpacking your CCSM3.0 tar archive (I'm using beta12):
1. in directory: $CCSMROOT/scripts/ccsm_utils/Tools
in file check_machine: change "bluesky" to "blueice"
2. in directory: $CCSMROOT/scripts/ccsm_utils/Machines
- Rename four files:
- batch.ibm.bluevista ==> batch.ibm.blueice
- env.ibm.bluevista ==> env.ibm.blueice
- l_archive.ibm.bluevista ==> l_archive.ibm.blueice
- run.ibm.bluevista ==> run.ibm.blueice
- edit: batch.ibm.blueice
set mach = blueice
- edit: l_archive.ibm.blueice
set mach = blueice
3. edit file: $CCSMROOT/models/bld/Macros.AIX
- add -lessl -lmass -lpessl -lmassv to ULIBS
- I retained all libraries included on Bluevista runs;
esmf (at least) needs to be retained.
4. run create_newcase with -mach blueice
5. MAYBE: in env_mach.blueice change MAX_TASKS_PER_NODE from 8 to 16. Also check to see that your run architecture is defined in 16 processor chunks. Check your GAU charges! On my first attempt I only used half the processors I had assigned to myself, and so was effectively charged double.
6. Verification
I compared Bluevista and Blueice outputs for indentical runs. In my case, CAM, CLM and POP start from observed ICs. I looked at daily output from CAM and CLM. I used the NCL routine vinth2p to interpolate the hybrid levels to pressure levels. I haven't checked POP or CSIM output.
In a comparison of day 0 output, data are identical, verifying that I've used identical ICs. Minute differences are seen in vegetation fields QVEGE and H2OCAN -- I'm assuming that's due to some esoterica in the CLM code initializing vegetation fields.
Minute, "round-off" errors are evident at 24 hours, but the atmospheric fields really don't diverge until after about two weeks. For my purposes, this indicates that the climatologies of the Bluevista and Blueice model builds are identical.
7. Timings: Don't expect anything close to optimized code. Runtimes increased by 27% compared to the same suite running at Bluevista
Please note that a model run at Bluevista will not be reproducible at Blueice, at least with the procedure I've used.
After unpacking your CCSM3.0 tar archive (I'm using beta12):
1. in directory: $CCSMROOT/scripts/ccsm_utils/Tools
in file check_machine: change "bluesky" to "blueice"
2. in directory: $CCSMROOT/scripts/ccsm_utils/Machines
- Rename four files:
- batch.ibm.bluevista ==> batch.ibm.blueice
- env.ibm.bluevista ==> env.ibm.blueice
- l_archive.ibm.bluevista ==> l_archive.ibm.blueice
- run.ibm.bluevista ==> run.ibm.blueice
- edit: batch.ibm.blueice
set mach = blueice
- edit: l_archive.ibm.blueice
set mach = blueice
3. edit file: $CCSMROOT/models/bld/Macros.AIX
- add -lessl -lmass -lpessl -lmassv to ULIBS
- I retained all libraries included on Bluevista runs;
esmf (at least) needs to be retained.
4. run create_newcase with -mach blueice
5. MAYBE: in env_mach.blueice change MAX_TASKS_PER_NODE from 8 to 16. Also check to see that your run architecture is defined in 16 processor chunks. Check your GAU charges! On my first attempt I only used half the processors I had assigned to myself, and so was effectively charged double.
6. Verification
I compared Bluevista and Blueice outputs for indentical runs. In my case, CAM, CLM and POP start from observed ICs. I looked at daily output from CAM and CLM. I used the NCL routine vinth2p to interpolate the hybrid levels to pressure levels. I haven't checked POP or CSIM output.
In a comparison of day 0 output, data are identical, verifying that I've used identical ICs. Minute differences are seen in vegetation fields QVEGE and H2OCAN -- I'm assuming that's due to some esoterica in the CLM code initializing vegetation fields.
Minute, "round-off" errors are evident at 24 hours, but the atmospheric fields really don't diverge until after about two weeks. For my purposes, this indicates that the climatologies of the Bluevista and Blueice model builds are identical.
7. Timings: Don't expect anything close to optimized code. Runtimes increased by 27% compared to the same suite running at Bluevista