Difference between pages "GA4.0 N216 Development" and "ARCHER porting"

Revision as of 17:12, 14 January 2014

Several of the UKCA Release Jobs have been ported from HECToR to ARCHER, and their jobids are detailed on that page.

Timings

Scaling and efficiency profiling for N48L60 UKCa CheS+ on ARCHER

The plot shows how UKCA scales when ported to ARCHER. The most efficient number of nodes to use is 3 (8x9), although the model will run fastest on 6 nodes (e.g. 16x9 or 12x12). Moving to 3 nodes from 6 will increase the run-time by 50%.

Code Changes

Cray cce Fortran Compiler

Corrections required for the Cray cce compiler on archer can be found on the bugfixes page.

@@ Line 1: / Line 1: @@
+Several of the [[UKCA Release Jobs]] have been ported from HECToR to ARCHER, and their [[UKCA_Release_Jobs#Job_List|jobids]] are detailed on that page.
-As part of the GA4.0 UKCA development, in Cambridge we are developing a N216L85 GA4.0 model using the CheT(+GLOMAP) [[Developing_releases#2013_TropIsop.2FCheT_.2B_GLOMAP_.28Co-ordinator:_Graham_Mann.29|configuration]]. However, there is no GA4.0 N216 base model existing at MetUM version 8.4.
+==Timings==
-==Create a GA4.0 N216 Base model==
+[[File:CCMI_timings_ARCHER.png|300px|thumb|right|Scaling and efficiency profiling for N48L60 UKCa CheS+ on ARCHER]]
-There are a number of steps that need to be made before a base GA4.0 N216 model is ready for UKCA to be included:
+The plot shows how UKCA scales when ported to ARCHER. The most efficient number of nodes to use is 3 (8x9), although the model will run fastest on 6 nodes (e.g. 16x9 or 12x12). Moving to 3 nodes from 6 will increase the run-time by 50%.
-# Assess the models available at the [http://collab.metoffice.gov.uk/trac/GA/wiki/GADocumentation/GA4.0#Jobs GA development page]
-#* This shows that the only existing N216 job available is [http://collab.metoffice.gov.uk/trac/GA/wiki/GAJobs/xhcea '''xhcea'''] on MONSooN
+==Code Changes==
-#** This job ran for 30 years, and the output is available on MOOSE
-#** The original UKMO version of this job is '''ajthw'''.
+===Cray cce Fortran Compiler===
-#* There are however GA4.0 N96 jobs available at vn8.0 ([http://collab.metoffice.gov.uk/trac/GA/wiki/GAJobs/aliur '''aliur''']) and vn8.4 ([http://collab.metoffice.gov.uk/trac/GA/wiki/GAJobs/amche '''amche'''])
-# It is therefore necessary to regrid '''amche''' to N216 for use by UKCA
+Corrections required for the Cray cce compiler on archer can be found on the [[Bugfixes#Cray_compiler_on_ARCHER|bugfixes page]].
-#* There is a [http://collab.metoffice.gov.uk/trac/GA/wiki/GASettings/GA4.0_HRes guide on how change the horizontal resolution] on the GA twiki.
-#** The ancillary files can be taken from '''xhcea''' which already exist on MONSooN.
-#** There may be other changes that have been made between '''aliur''' and '''xhcea''' which are not represented in the guide. I have therefore asked Jeremy Walton to port '''aliur''' as well as '''amche''' over to MONSooN to check the differences.
-#* The new job '''xjgva''' has been created which is a GA4.0 N216L85 AMIP candidate, based on '''amche''' and using the differences from the [http://collab.metoffice.gov.uk/trac/GA/wiki/GASettings/GA4.0_HRes guide] and the differences between '''aliur''' and '''xhcea'''.
-#** Differences between vn8.0 GA4.0 N216 ('''xhcea''') and N96 ('''aliur'''): [[diff.xjgvx.xjgvz]]
-#** Differences between vn8.4 GA4.0 N216 ('''xjgva''') and N96 ('''amche'''): [[diff.xjgva.xjgvy]]
-#* It should be noted that the settings were mainly taken from the guide rather than the differences between vn8.0 jobs. Notable differences between '''xjgva''' and '''xhcea''', or '''xjgva''' and '''amche''', are:
-#*# In '''xjgva''' soil moisture has been initialised to <code>$USER_ANCIL/N216/orca025/smc_snow/v1/ajthma.smc_scaled</code> whereas '''amche''' sets this from the dump.
-#*# '''xjgva''' uses the same start dump as '''xhcea''': <code>$USER_ANCIL/N216/start_dumps/v1/ajthma.dai1910</code>.
-#*# In the ''Large Scale Precipitation'' section, '''xjgva''' uses '''7 substeps over full column''', rather than 10 in '''xhcea'''.
-#*# '''xjgva''' uses the standard vn8.4 spectral files, rather than the vn7.9 spectral files used by '''xhcea'''. There should be no difference in the components of the file used however.
-# '''xjgva''' compiled successfully and has been sent off for an equivalent 30-year run to '''xhcea'''.
-#* '''xjgva''' runs at approximately 2400 +/- 101 seconds for each 10-day jobstep, giving a maximum throughput of 12 model months per day.
-#* However, the actual throughput is quite variable, having between 3 and 12 model months per day with an average turnover of 9 model months at present.