evaluating the spread of climate model ensembles based on
play

Evaluating the Spread of Climate Model Ensembles Based on Computing - PowerPoint PPT Presentation

Evaluating the Spread of Climate Model Ensembles Based on Computing Environment Selection Tom Robinson Multicore Workshop 2019 Outline Motivation Ensemble method Ensemble description Ensemble spreads and comparison


  1. Evaluating the Spread of Climate Model Ensembles Based on Computing Environment Selection Tom Robinson Multicore Workshop 2019

  2. Outline • Motivation • Ensemble method • Ensemble description • Ensemble spreads and comparison • Conclusions

  3. Motivation • Reproducibility is important • Floating point and rounding differences between runs prevents bit-for-bit reproducibility • “Climate answers” are dependent on the selection of platform/compiler (options) • What is the “model spread” due to rounding error? • Is the model spread platform dependent?

  4. Ensemble Method • GFDL AM4 (github.com/NOAA-GFDL/AM4) • Simulate rounding error – Single random point – Initial mid-level T 10 -13 K – Different point for each ensemble member • Model run for one year

  5. Ensembles Ensemble Name Compiler Platform Processor # of ensembles Base Production intel 16 Gaea B/H 300 AVX intel 16 Gaea B/H 100 Intel 18 intel 18 Gaea B/H 100 Cray cray Gaea B/H 95 Theta intel 16 theta KNL 118 Hera intel 19 Hera Skylake 47

  6. Average standard deviation • Find the point-by-point standard deviation – Take a global average • Plot and compare – Point by point mean • Are the means similar? – Point by point standard deviation – Compare across ensembles • Is spread platform dependent?

  7. Global Spread Surface Pressure 7 Base 7 Base Average spread 30 members Average spread 50 members BH/Cray BH/Cray 6 Skylake/Intel19 6 Skylake/Intel19 KNL/Intel16 KNL/Intel16 5 5 BH/Intel18 BH/Intel18 BH/intel16avx 4 4 3 3 2 2 1 1 0 0 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 7 Base Average spread 100 members BH/Cray 6 BH/Intel18 KNL/Intel16 5 BH/Intel18 4 3 2 1 0 1 2 3 4 5 6 7 8 9 10 11 12

  8. Global Spread U wind 4.5 4.5 Average spread 30 members Average spread 50 members 4.25 4 4 3.75 3.5 3.5 3.25 Base 3 3 BH/Cray Base Skylake/Intel19 2.75 BH/Cray KNL/Intel16 2.5 2.5 Skylake/Intel19 BH/Intel18 BH/Intel16avx2 KNL/Intel16 2.25 BH/Intel18 2 2 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC 4.5 1.75 u Average spread 100 members u Standard Deviation of Standard Deviation 50 members 1.65 4 1.55 1.45 3.5 1.35 1.25 Base 3 1.15 BH/Cray BH/Cray Skylake/Intel19 47Skylake/Intel19 1.05 KNL/Intel16 KNL/Intel16 2.5 0.95 BH/Intel18 BH/Intel18 BH/Intel16avx2 0.85 Base 2 0.75 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC 1 2 3 4 5 6 7 8 9 10 11 12

  9. Mean ps (base)

  10. ps Standard Deviation

  11. KNL-Base Mean Difference *All values within 1 standard deviation

  12. Standard Deviation Diff (theta-base)

  13. Standard Deviation Diff (cray-base)

  14. Standard Deviation %Diff (KNL-base)

  15. Standard Deviation %Diff (cray-base)

  16. Standard Deviation %Diff (KNL-base)

  17. Standard Deviation %Diff (skylake-base)

  18. Base30-Base %diff

  19. Conclusions • Ensemble means are not platform dependent • Ensemble spreads over a local region are platform/compiler dependent • You should use a large ensemble to report the error due to rounding on your computing platform. – Global Average for summary – Map of values for patterns/weaker areas

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend