Optimizing GCHP benchmarks on AWS #511
Description
Your name
Bob Yantosca
Your affiliation
Harvard + GCST
Please provide a clear and concise description of your question or discussion topic.
We have recently updated the AWS cloud benchmark instance from c5.12xlarge to c7i.16xlarge. This has resulted in a 2x speedup for the GEOS-Chem Classic benchmarks. However, the GCHP benchmarks do not show a speedup of the same magnitude:
Using c7i.16xlarge
gchp-c24-1Mon.14.7.0-alpha.6
Start Time 2025-09-02T15:36:26+0000
End Time 2025-09-02T22:09:59+0000
Elapsed Time 06:33:33
Using c5.12xlarge
gchp-c24-1Mon-14.7.0-alpha.1
Start Time 2025-06-23T20:04:29+0000
End Time 2025-06-24T04:19:10+0000
Elapsed Time 08:14:41 The run scripts for the GCHP cloud benchmarks are located in our https://github.com/geoschem/gc-cloud-infrastructure repository:
- https://github.com/geoschem/gc-cloud-infrastructure/blob/main/benchmarks/scripts/stages/runGCHP.sh
- https://github.com/geoschem/gc-cloud-infrastructure/blob/main/benchmarks/scripts/stages/modules/gchp-modules.sh
The GCHP benchmarks run on a single node using 60 cores, and so do not use ParallelCluster. I'm wondering if there are some tweaks we can add to the run scripts to optimize performance. @yuanjianz, I know you recently added some run scripts for AWS on EFA using ParallelCluster... if you have any ideas I can try please let me know.
Tagging @yuanjianz @laestrada @lizziel @msulprizio @yuyao-cyber