wiki:Archer/cce8.3.7

Version 23 (modified by grenville, 4 years ago) (diff)

Monday June 29th

  • The default ARCHER Cray compiler is now cce8.3.7. NCAS-CMS have modified the UM infrastructure so that builds pick up compatible libraries and modules. Previous ad-hoc instructions for building need no longer be followed an indeed may cause confusion.
  • Models built with cce8.2.1 should run OK under the upgraded environment.
  • Do not perform incremental builds mixing code built under cce8.2.1 with that built under cce8.3.7. This means that your first compilation job after the upgrade should be a complete rebuild; subsequent incremental builds will be OK.

Pre June 24th

As part the regular software upgrade cycle, ARCHER will remove the Cray cce8.2.1 compiler on June 24th. cce8.2.1 has been the UM compiler since ARCHER became operational and we have maintained its use to minimize possible disruption but it does appear to be a source of OOM errors in some long-running UM jobs. These OOM errors have not appeared in jobs running code compiled with later Cray compilers, so it is timely to upgrade to the current ARCHER default - cce8.3.7.

cce8.3.7 GCOMs (gcom3.8 gcom4.2 gcom4.5 gcom4.6 gcom4.7 gcom5.1) are in $UMDIR/gcom/cce8.3.7

Include MPICH_NO_BUFFER_ALIAS_CHECK=1

Compiler upgrade notes

UM Version Description cce8.3.7 jobid cce8.2.1 jobid Comments
8.2 N96 HadGEM3A xlnia xleha Needed to reduce optimization on convection to O1 - didn't look for specific routine - only checked with 1 start file. We haven't changed the config file; see what happens in the community. This may be data driven
8.2 N512 - the SWAMMA driving model xjjzv xjjzw OOM errors seen (see ticket #1535); worked as is
8.2 12km SWAMMA xlnib xlehb OOM errors in the cce8.2.1 job; worked as is
8.4 N96 GA? xjfvy xlchx OOM errors in the cce8.2.1 job; needed fix for bug in locate_hdps.F90 - had to cast array indices to INTS - change added to the ncas branch see ticket #1485
6.6.3 HadGEM2-AMIP xlnic xjgcc runs OK built with -g - blows up somewhere with 8.2.1 settings — large_scale_cloud::ls_arcld is the culprit; build this with -O0 and the model runs OK
6.6.3 HadGEM2-ES xlnid xjgcu worked as is
6.6.3 HadGEM2-CC xlnie xjgcf needs large_scale_cloud::ls_arcld compiled with -O0 and the model runs OK
7.3 HadGEM3-A r2.0 xlnif xjeya worked as is
8.2 N96 GA4.0 AMIP xlnnb xjeye No modifications required
8.4 HG3A GA4.0 xlnnc xjeyf No modifications required
8.6 HG3A GA6.0 xloba xjvpu No modifications required
8.6 UKCA xlnzb —- No modifications required
6.1 HiGEM1a xjazo xlnig No modifications required - model and reconfiguration ran — haven't yet found where the leave files went