Norm Phillips work in Data Assimilation A.Hollingsworth Seattle Jan 2004

1
Norm Phillips work in Data Assimilation
A.Hollingsworth Seattle Jan 2004
The Interplay of Computer Power, Computer
Architecture and Numerical Algorithms in the
progress of Numerical Weather Prediction

• A.Hollingsworth, A.Simmons, W. Zwieflhoefer,
• M.Dragosavac, S.Uppala, J.Woollen, D.Marbouty,
• J-N Thepaut, R Engelen, A Dethof,
• ECMWF
• NCEP

2
Scope of Talk

• NWP requirements for resolution
• Semi-Lagrangian time-schemes
• Implementing efficient schemes on parallel
  machines
• Operational and scientific implications of such
  economies
• A look back to June 1944
• A look to the challenges of the future

3
Significance of model and analysis resolution

• Model resolution can control the success of
  forecasts for major rain systems.
• The next few pictures illustrate the importance
  of resolution (40km v 65km) in one of a series of
  episodes of heavy rains in the Mediterranean
• In the 40km model, the forecast successfully
  stretched, and then rolled up, the streamer of
  Potential Vorticity, while the 65 km model was a
  bust on these critical features.

4
9 September 2000 - 12 UTC D6 forecast
AN
TL511 D6
TL319 D6
AN TL511 D6
TL319 D6
5
Paying for Model Resolution

• The change in 2000 at ECMWF from a 65km (T319) to
  a 40 km (T511) model had a big positive impact,
  as was expected from hundreds of days of
  pre-operational trials.
• Dritschel et al. (1999) imply that 15km
  resolution is needed for a good 5-day forecast of
  the PV field realisable about 2010.
• Resolution is very costly (computer cost
  increases as cube of resolution).
• Several approaches can meet the need
• Increase the money stream (v.difficult)
• Keep the money-stream constant and rely on
  Amdahls Law.
• Keep the money-stream constant, rely on Amdahls
  Law, and use efficient time schemes
  (semi-implicit, semi-Lagrangian)

6

• The semi-Lagrangian scheme, integrates the
  equations of motion by
• Calculating, for every point on the grid, the
  trajectory from the departure point (at t-1)
• Interpolating the (t-1) values to the departure
  points
• Advecting the values forward along the
  trajectories

7

• Parallellism Communication on Distributed
  Memory Machines
• Efficient time schemes require the solution of
  global equations.
• One cannot get all the data into one processor,
  soA continuous complex shuffling / re-shuffling
  of the data is required to do a global
  calculation in many small steps
• The figure illustrates the many data
  transpositions needed to pass from physical to
  Fourier to legendre space and back again
• The implication is that the speed capacity of
  the inter-processor communication is as crucial
  as the processor power.

8
x 4
x 12
x 72
x 150
9
Implications of numerical efficiency for
productivity vendor competition

• Current semi-Lagrangian schemes offer very
  substantial gains in efficiency
• Without those gains ECMWF could not afford
• Deterministic model at 40 km (T511)
• 51 member Ensemble system at 80km (T255)
• Advanced 4D-Var assimilation system
• Assimilation of millions of pieces of data from
  several dozen satellite instruments
• Coupled seasonal forecast systems at 2 deg. (T95)
• At best ECMWF could afford a 2-day forecast at
  T511
• We certainly could not plan a 25 km system
  (T799/L90) in 2005

10
Sustained forecast improvements in both
hemispheres Convergence of skill between the
hemispheres
11
A look to the pastLet us now praise great men
-Day 6 June 1944
12
Observations for 12UTC 3 June 1944
423 pilot balloons Observations supplied by Jack
Woollen, NOAA/NWS/NCEP
13
Observations for 12UTC 3 June 1944
676 SYNOPS 112 SHIPS
14
10m wind and low cloud 00UTC 6 June
1944
T159 3D-Var analysis
15
10m wind and low cloud
06UTC 6 June 1944
T159 3D-Var analysis
16
10m wind and low cloud
12UTC 6 June 1944
T159 3D-Var analysis
17
10m wind and low cloud
18UTC 6 June 1944
T159 3D-Var analysis
18
The early morning of June 6 1944 Low cloud and
westerly winds off the Normandy beaches
Pictures from US Naval Historical Center
19
June 6 1944 Clear skies over the channel and
later over the Normandy beaches
20
Comparison with contemporary charts
Reproduced from www.meteo.fr
21
10m wind and cloud T799 from 00UTC 3
June 1944
22
10m wind cloud with 25km (T799) from 12UTC 3
June
23
http//www.ecmwf.int/newsevents/releases/030604.ht
ml
24
Historical perspective

• 1994 (ten years ago)
• Cray C90-16 installed in 1992
• 16 processors
• Sustained performance 6 gigaflops
• 2004 (one decade later)
• IBM Phase 3
• 4000 processors (250-fold increase)
• Sustained performance 2 teraflops (120-fold
  increase per decade)
• Distributed memory systems and a competitive HPC
  market
• 2013 (almost another decade)
• Preparation of 10 km system in the Strategy
  Review
• Unknown number of processors
• Sustained performance 200 teraflops (100-fold
  increase)

25
Climate Monitoring and the Chemical Weather
Forecast System

• Long-range transport of air pollutants is a well
  established fact- 40 of the material in
  Europe arrived from Asia, via N.America
• Satellite data provide a vast amount of data on
  atmospheric composition reactive gases, smog,
  aerosol, greenhouse gases.
• Extraction of the information on composition
  requires as the pre-requisite the sophisticated
  NWP assimilation systems to provide temperature,
  humidity, cloud,
• The GEMS consortium (10 Science institutes, 10
  regional air-quality labs, ECMWF) will provide an
  operational global composition monitoring
  capability and a global /regional air-quality
  operational forecast capability, built around
  ECMWF system (funding by EU).
• A fully interactive chemistry module in the
  global ECMWF model will pose substantial
  computational challenges.

26
Structure of an Earth-system model - all
processes and interactions must be well
represented
27
Monitor Greenhouse Gases- CO2, N2O, CH4, CO
CO2 Stratosphere May 2003
First analysis of stratospheric CO2 shows
Brewer-Dobson type of circulation. Variability is
also much smaller than in troposphere.
28
CO2 Troposphere May 2003
CO2 tropospheric columns are assimilated from
AIRS infrared observations. Monthly mean
distribution for May 2003
29
Monitor Forecast AEROSOL ( Fires)

• Model and assimilate global aerosol information
• Heritage -
• Instruments MERIS, MODIS x 2, MISR, SEAWIFS,
  POLDER
• Data Mgt tbd
• R/T
• Modelling
• Sources/ Sinks
• Data Assim.
• Validation

30
Monitor Forecast Reactive Gases
Ozone Hole 1 Oct 2003 Southern Hemisphere 1 Map
of total column O3 2 Cross-section 3
Validation v. Neumayer ozonesonde
31
Closing words

• Magna Laudatio to JNWP / NMC / NCEP
• for 50 years of pioneering leadership
• Good wishes from ECMWF for your continued
  success.
• The future will be at least as challenging as the
  past,
• on the science
• on the political science
• on the operational systems to meet customer needs

32
End

• Thank you for your attention!

33
7 September 2000 - 12 UTC D4 forecast
ANALYSIS TL511 D4
TL319 D4
34
8 September 2000 - 12 UTC D5 forecast
AN
ANALYSIS TL511 D5
TL319 D5
TL511 D5
TL319 D5
35
(No Transcript)
36
Group Captain J.M. Stagg
Meteorological advisor to Eisenhower Responsible
for reconciling the forecasts of three teams
The Met Office
The Royal Navy The US Air Force
Photograph from www.metoffice.com