Conversation
…from benchmark_ALE
|
The model version in the
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
!redeploy |
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
!redeploy |
|
🚀 Attempted to deploy 🖥️
|
|
!redeploy |
|
🚀 Attempted to deploy 🖥️
|
…o all major components
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
!redeploy |
|
🚀 Attempted to deploy 🖥️
|
|
The compiler flag for the memory layout transformations did not make any significant performance impact
|
…om Helen for a ROMS 3-month run
|
🚀 Attempted to deploy 🖥️
|
|
🚀 Attempted to deploy 🖥️
|
|
!update-configs |
|
🔧 Opening/Updating Model Configuration PRs in ❌ One or more errors occurred with the workflow Configurations RequestedConfigurations requested from profile
Pull Requests Opened/UpdatedThe following PRs were opened/updated: More details can be found in the workflow run: https://github.com/ACCESS-NRI/ACCESS-OM3/actions/runs/34836506818 |
|
Ughhh I keep forgetting that the MCW branch name needs to be changed for the auto-bot-config-update |
benchmark_ALE test case had a fairly big performance impact from using a 3-stage compilation process - i) compile with
-fprofile-generate& run benchmark ii) combine generatedprofrawbinary files into a single profile file (usingllvm-profdata merge) and iii) recompile with-fprofile-use=/path/to/single/profile/fileIdeally, each config is compiled with it's own profile info, but it might be a reasonable trade-off to compile with a single profiling info that is generated from raw profiling data from all the configs and test cases.
This PR is to come up with some performance numbers to help us make that choice.
Performance summary table (25k IAF config)
From a single 30-day run of the experiments, so the numbers have some (unknown) error-bars on them. The profiling info in the first 6 build-cd deployments (i.e.,
pr258-1throughpr258-6, inclusive) are all from benchmark_ALE case.pr258-11andpr258-14using profiling info from a 1-year run with OM3 25k IAF.47-perf-rho-remap-diags, 1-year profile generation run)47-perf-rho-remap-diags, no profiling info)47-perf-rho-remap-diags, and47-perf-rho-remap-diagsprofiling info)47-perf-rho-remap-diags, OM3 25k IAF profiling info for 1 year)47-perf-rho-remap-diags,dev/gfdlprofiling info)dev/gfdlprofiling info)47-perf-rho-remap-diagsprofiling info)Performance summary table (25k RYF config)
From a set of 3 runs, with
repeat=Truefor a 30-day experiment. The profiling info in the first 6 build-cd deployments (i.e.,pr258-1throughpr258-6, inclusive) are all frombenchmark_ALEcase.pr258-11andpr258-14using profiling info from a 1-year run with OM3 25k IAF.Performance summary table (100k IAF config)
From a set of 3 runs, with
repeat=Truefor a 1-year experiment. The profiling info in the first 6 build-cd deployments (i.e.,pr258-1throughpr258-6, inclusive) are all frombenchmark_ALEcase.pr258-11andpr258-14using profiling info from a 1-year run with OM3 25k IAF. Since the base config runs on thenormal(i.e.,cascadelakequeue) - this uncovered a hidden variable of the hardware as well -- bothpr258-11andpr258-14crashed with illegal instruction error.🚀 The latest prerelease
access-om3/pr258-18at b1cced7 is here: #258 (comment) 🚀🚀 The latest prerelease
access-om3/pr258-20at 377d53b is here: #258 (comment) 🚀