Version 30 (modified by willie, 3 years ago) (diff)

Performance of Standard Configurations

Global Atmosphere 7.0/Global Land 7.0

Suite id Description UM Version Date Run Platform Decomposition* OpenMP Threads Length of run Dump Frequency Wallclock Data output vol Comment
u-ab878 GA7.0 N96 AMIP 10.3 10.02.17 ARCHER 16 x 12 2 30 days 10 days 01:07:53
u-ag689/n216 GA7.0 N216 AMIP 10.5 22.06.17 ARCHER 42 x 36 1 10 days 10 days 00:29:27
u-ak155 GC3.1 N216 ORCA025 10.6.1 07.06.17 ARCHER A:42x36 O:24x26 X:12 (6 nodes) 1 30 days 10 days 01:24:29
u-ai862 GC3.0 N96 ORCA025 10.6 20.07.17 ARCHER A:16x18 O:18x26 X:8 (4 nodes) 1 10 days 10 days 00:27:03
u-an561 UKESM0.6-CN PI, new oxidants file, lana, StratTrop 10.6 06.29.17 — 07.17.17 ARCHER A:48x24 O:9x12 1 1yr (12 x 1 mnth cycles) 10 days 12 x 52 mins 111GB (archived (no dumps, pp+compressed netcdf))
u-q606 UKESM0.8-CN N216 ORCA025 (some improvements) 10.7 04.10.17 ARCHER A:36x24 O:30x30 X:4 nodes (8/2) 1 1 mnth 10 days 2:00:00 working copy only
u-aq711 piControl N216-ORCA025 10.7 17.10.17 ARCHER A:36x24x2 O:30x30 X:8 nodes (16/2) 2 (atm) 1 mnth 10 days 1:45:00 25% variation in run times - expensive because multi-threaded (necessary?) working copy only
u-as037 (r59528) GC3.1 N96 ORCA1 PI Control CMIP6 10.7 11.11.17 ARCHER A:12x8 O:9x8 X:6 (3 nodes) 2 1m 10 days 00:56:59 162Gb (History_Data - 2month run)

*Decomposition: A = Atmos (EW x NS), O = Ocean (EW x NS), X = XIOS processes (nodes)

A more detailed view of u-as037 (r 61036) running on ARCHER

This looks at the performance of HadGEM3-GC3.1 N96ORCA1 PI Control for CMIP6 over a wide range of atmosphere decompositions with a fixed number of IO servers (6). The ocean decomposition was 9x8 and the XIOS used 6 processes. The atmosphere was run with two threads, except for one run at 36x28 where three threads were used (3,186 processes in the graph below).

The following graphs show the model rate (in model years per day) against cost and total number of processes. Each point in the graph is the result of running the coupled suite for a model month. The cost axis was obtained by rescaling the processors axis in the lower graph. The cost of computing the graph was 440 kAU.

Model Rate vs Cost

Model rate vs total processes

Attachments (7)

Download all attachments as: .zip