SLIDE 1
Outline
- Motivation
- Ensemble method
- Ensemble description
- Ensemble spreads and comparison
- Conclusions
Evaluating the Spread of Climate Model Ensembles Based on Computing - - PowerPoint PPT Presentation
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
SLIDE 2
SLIDE 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?
SLIDE 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
SLIDE 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
SLIDE 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?
SLIDE 7
Global Spread Surface Pressure
1 2 3 4 5 6 7 1 2 3 4 5 6 7 8 9 10 11 12
Average spread 30 members
Base BH/Cray Skylake/Intel19 KNL/Intel16 BH/Intel18 1 2 3 4 5 6 7 1 2 3 4 5 6 7 8 9 10 11 12
Average spread 50 members
Base BH/Cray Skylake/Intel19 KNL/Intel16 BH/Intel18 BH/intel16avx 1 2 3 4 5 6 7 1 2 3 4 5 6 7 8 9 10 11 12
Average spread 100 members
Base BH/Cray BH/Intel18 KNL/Intel16 BH/Intel18
SLIDE 8
Global Spread U wind
2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
Average spread 30 members
Base BH/Cray Skylake/Intel19 KNL/Intel16 BH/Intel18 2 2.5 3 3.5 4 4.5 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
Average spread 50 members
Base BH/Cray Skylake/Intel19 KNL/Intel16 BH/Intel18 BH/Intel16avx2 2 2.5 3 3.5 4 4.5 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
u Average spread 100 members
Base BH/Cray 47Skylake/Intel19 KNL/Intel16 BH/Intel18 BH/Intel16avx2 0.75 0.85 0.95 1.05 1.15 1.25 1.35 1.45 1.55 1.65 1.75 1 2 3 4 5 6 7 8 9 10 11 12
u Standard Deviation of Standard Deviation 50 members
BH/Cray Skylake/Intel19 KNL/Intel16 BH/Intel18 Base
SLIDE 9
Mean ps (base)
SLIDE 10
ps Standard Deviation
SLIDE 11
KNL-Base Mean Difference
*All values within 1 standard deviation
SLIDE 12
Standard Deviation Diff (theta-base)
SLIDE 13
Standard Deviation Diff (cray-base)
SLIDE 14
Standard Deviation %Diff (KNL-base)
SLIDE 15
Standard Deviation %Diff (cray-base)
SLIDE 16
Standard Deviation %Diff (KNL-base)
SLIDE 17
Standard Deviation %Diff (skylake-base)
SLIDE 18
Base30-Base %diff
SLIDE 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