Minutes of the QUEST-ESM NCAS FAMOUS meeting of Monday, 1st October 2007
Present: Ros, Jeff, Lois, Robin, Simon, Annette, Jonathan
A new version of FAMOUS is nearly ready. This is still MOSES I but includes the science changes Robin has been working on over the last year (sea-ice tuning, ozone, orography, iceberg-substitute flux adjustment) and various technical improvements (very long absolute time convention for filenames, orbital forcing mod). Robin is currently running a control on HPCx. When this is done, Annette will do one on quest, and we will compare statistically as useful. As this is a major release, testing an ensemble (which is statistically more rigorous) would be valuable. If there are no problems, the release will be made, on puma in the FAMOUS-enhanced version of the UMUI. The mods etc. will also be stored on puma, as well as being installed on HPCx and quest. It is appropriate to use puma for sharing UM-related files in this way. The very long absolute time convention requires an option in the UMUI and some of the small executables have to be replaced, because the namelist entries for filenames have to allow for longer strings. This is backward compatible.
Annette has run successfully on the quest cluster, which has now generally been stable for a month or so. There are unclear issues about scripts and commands to submit jobs, and bottlenecks for output. This will become clearer to us when we have more experience with using it, as Robin will soon do.
Simon has carried out some timing tests for FAMOUS on HECToR and quest. He reports that 1x8 PEs in the atmos is about twice as fast as 8x1 on both machines. Such a sensitivity is not usual and must somehow be related to FAMOUS low resolution, as it relates to commuications overhead. On HPCx we run on 16 PEs but perhaps there is a similar sensitivity. The fastest Simon obtained on HECToR was 124 years/day on 8 PEs and 138 on 16. He pointed out we are using code options for faster execution but loss of bit-comparability with different PE configurations, and will check the effect of not using these options, as bit-comparability is very desirable.