UM On Gadi
General details on how to port an existing UM job to Gadi can be found on the ACCESS wiki, this page just describes runs we use
Performance details of various runs are being recorded on a spreadsheet for future reference - follow the links next to each experiment name to go to a page with the details of that job
In general on Gadi the UM scales better - increasing the number of cpus gives results much faster in walltime than on Raijin, however SU cost is roughly similar so with the same allocation you won't be able to run much more model time.
UM7.3
PLEASE NOTE - we are cleaning up input files, so not all of the ACCESS inputs will be copied over, only what is needed. Please test your runs before Jan 6th to make sure the files you need on Gadi are in place
General information about converting an existing UM 7.3 job to Gadi is available on the ACCESS wiki, however as there are a wide variety of different configurations you may still see errors for your specific configuration. If you do run into trouble please let CMS know and we'll do our best to get you running.
We recommend starting with a processor decomposition of 16 x 12 for UM 7.3 jobs - see the spreadsheet graphs for more detail
Run ID | Description | Performance Details |
---|---|---|
vauya | ACCESS 1.3 AMIP | [1] |
vauyc | ACCESS 1.0 AMIP | [2] |
UM8.5
Follow the version 7.3 instructions, but use the hand edit file '~access/gadi/handedits/um8.5'
UM10.4
PLEASE NOTE - we are cleaning up input files, so not all of the ACCESS inputs will be copied over, only what is needed. Please test your runs before Jan 6th to make sure the files you need on Gadi are in place
UM11.4
At climate resolutions (n96) UM 11.4 scales well up to about 1000 cpus. We recommend starting with a decomposition of around 28 x 24. OpenMP threads should in general not be used at this resolution.
Run ID | Description | Performance Details |
---|---|---|
u-bo915 | GA 7 | [3] |
u-bm651 | HighResLAM | [4] |
|}