Conversation
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
@manodeep, there's three prereleases here that hopefully will be useful for testing performance:
Would you be able to run your usual tests? |
|
Piggy-backing off this PR to test the performance impact of the other extreme of adding |
|
🚀 Attempted to deploy 🖥️
|
|
The performance numbers for the previous builds are
Printing summary table of top 4 results (all times are in mins, sorted by 'Simulated Years per wallclock-day'):
edit All the runs were bitwise identical. [~/perf-esm1p6/test-mavx2-vs-avx512 @gadi-login-04]$ ~/codes/utils/check_for_repro_multi_dirs.sh *
Checking that all arguments are directories ...done
Checking that all arguments are unique directories ...done
Comparing Atmosphere in dirs = (20260226-PI-config-base,20260226-PI-config-base-pr192-3): Status - PASSED
Comparing Ocean in dirs = (20260226-PI-config-base,20260226-PI-config-base-pr192-3): Status - PASSED
Comparing Atmosphere in dirs = (20260226-PI-config-base,20260226-PI-config-base-pr192-4): Status - PASSED
Comparing Ocean in dirs = (20260226-PI-config-base,20260226-PI-config-base-pr192-4): Status - PASSED
Comparing Atmosphere in dirs = (20260226-PI-config-base,20260226-PI-config-base-pr192-5): Status - PASSED
Comparing Ocean in dirs = (20260226-PI-config-base,20260226-PI-config-base-pr192-5): Status - PASSED
Comparing Atmosphere in dirs = (20260226-PI-config-base-pr192-3,20260226-PI-config-base-pr192-4): Status - PASSED
Comparing Ocean in dirs = (20260226-PI-config-base-pr192-3,20260226-PI-config-base-pr192-4): Status - PASSED
Comparing Atmosphere in dirs = (20260226-PI-config-base-pr192-3,20260226-PI-config-base-pr192-5): Status - PASSED
Comparing Ocean in dirs = (20260226-PI-config-base-pr192-3,20260226-PI-config-base-pr192-5): Status - PASSED
Comparing Atmosphere in dirs = (20260226-PI-config-base-pr192-4,20260226-PI-config-base-pr192-5): Status - PASSED
Comparing Ocean in dirs = (20260226-PI-config-base-pr192-4,20260226-PI-config-base-pr192-5): Status - PASSED |
|
!redeploy |
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
Sadly the compile job got killed (exceeding memory/time on the login node) |
|
!redeploy |
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
Printing runtime and imbalance results (all times are in mins, sorted by 'Simulated Years per wallclock-day'):
Printing summary table of top 7 results (all times are in mins, sorted by 'Simulated Years per wallclock-day'):
Deploy 10 and 11 are not bitwise identical with the rest, but bitwise identical with each other. All the other builds are bitwise identical (as noted in a previous comment) Performance slow-downs or improvements are in the few % level - nothing major as the 12% drop seen in OM2. However, UM7 seems to be waiting longer for MOM5 - so adding |
|
One thing was bothering me that even with -O2 for gcom4 (deployment 11, also includes
Results are bitwise identical - this is in contrast to PI-config, where the results were different. Not sure if that is expected, or why that would be the case. [~/perf-esm1p6/test-amip-with-gcom4-O2 @gadi-login-01]$ ~/codes/utils/check_for_repro_multi_dirs.sh *
Checking that all arguments are directories ...done
Checking that all arguments are unique directories ...done
Comparing Atmosphere in dirs = (20260303-dev-amip,20260303-dev-amip-pr192-11): Status - PASSEDIn case the new build scales better, I am now running two tests with 400 cores - one with the base config (which I am reasonably confident does not scale well to 4-nodes) and another for the deployment 11. Let's see what the numbers say. |
|
Updated summary table: Printing summary table of top 6 results (all times are in mins, sorted by 'Simulated Years per wallclock-day'):
From these runs, there are no bitwise identical pairs, except for the 2-node configs that I ran last night. [~/perf-esm1p6/test-amip-with-gcom4-O2 @gadi-login-04]$ ~/codes/utils/check_for_repro_multi_dirs.sh *
Checking that all arguments are directories ...done
Checking that all arguments are unique directories ...done
Comparing Atmosphere in dirs = (20260303-dev-amip,20260303-dev-amip-4nodes-400cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip,20260303-dev-amip-pr192-11): Status - PASSED
Comparing Atmosphere in dirs = (20260303-dev-amip,20260303-dev-amip-pr192-11-1node-100cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip,20260303-dev-amip-pr192-11-3node-306cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip,20260303-dev-amip-pr192-11-4nodes-400cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-4nodes-400cores,20260303-dev-amip-pr192-11): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-4nodes-400cores,20260303-dev-amip-pr192-11-1node-100cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-4nodes-400cores,20260303-dev-amip-pr192-11-3node-306cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-4nodes-400cores,20260303-dev-amip-pr192-11-4nodes-400cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-pr192-11,20260303-dev-amip-pr192-11-1node-100cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-pr192-11,20260303-dev-amip-pr192-11-3node-306cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-pr192-11,20260303-dev-amip-pr192-11-4nodes-400cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-pr192-11-1node-100cores,20260303-dev-amip-pr192-11-3node-306cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-pr192-11-1node-100cores,20260303-dev-amip-pr192-11-4nodes-400cores): Status - FAILED
Comparing Atmosphere in dirs = (20260303-dev-amip-pr192-11-3node-306cores,20260303-dev-amip-pr192-11-4nodes-400cores): Status - FAILED |
This PR is to test replacing
-xCORE*/-xavx/-axCORE-*with-mavx2across the ACCESS-ESM1.6 stack🚀 The latest prerelease
access-esm1p6/pr192-11at d863eb5 is here: #192 (comment) 🚀