- 1 - HIGH PERFORMANCE COMPUTING IN OPERATIONAL METEOROLOGY Geoff Love President of the WMO Commission for Basic Systems
Dec 31, 2015
- 1 -
HIGH PERFORMANCE COMPUTING IN OPERATIONAL METEOROLOGY
Geoff Love
President of the WMO
Commission for Basic Systems
- 2 -
OVERVIEW
• A couple of definitions
• Where we have come from
• Where we are now
• Where we might be going in the short- and longer-terms
- 3 -
DEFINITIONS
• High Performance Computing: Computing performed on
a system that, at the time of its commissioning, qualified
as one of the top 500 (publicly benchmarked) systems in
terms of ability to deliver sustained floating point
operations.
- 4 -
DEFINITIONS
• Operational Meteorology: “Operational” requires that
production systems are supported in a robust way (code
upgrades are easily facilitated, data management is
streamlined, visualisation tools are available, etc.) - to be
distinguished from, for example, the research environment.
“Meteorology” includes both climate and weather
- 5 -
WHERE WE HAVE COME FROM ?
YEAR MACHINE GFLOP
• 1968 IBM 360 0.00065
• 1982 FACOM M200 0.006
• 1988 ETA 10P 0.12
• 1990 CRAY X-MP 0.23
• 1992 CRAY Y-MP2E 0.7
• 1993 CRAY Y-MP3E 1
• 1995 CRAY Y-MP4E 1.3
• 1997 NEC SX-4 32
• 1998 2xNEC SX-4 64
• 1999 NEC SX-5 104
• 2000 NEC SX-5 128
• 2001 2xNEC SX-5 256
A
Increase of Computer Power with Time
0123456789
1960 1970 1980 1990 2000 2010
yearlo
g10
of c
ompu
ter
pow
er
(kflo
ps)
LOG COMPUTERPOWER
- 6 -
WHERE WE HAVE COME FROM ?
A
- 7 -
WHERE WE HAVE COME FROM ?
•
A
- 8 -
WHERE WE HAVE COME FROM ?
•
A
- 9 -
SYSTEM EVOLUTION
1968 Regional analysis, regional prediction
1984 Experimental hemispheric prediction, regional
nesting
1986 Hemispheric prediction, regional prediction
1990 Global prediction
1994 Regional assimilation, global assimilation
- 10 -
WHERE ARE WE NOW ?
• Global and regional 3-D variational scheme for data
assimilation.
• Global, regional and mesoscale atmospheric and ocean
forecast systems. Ensemble production.
• Air quality modelling, including a variety of chemistry
options.
• Dispersion, tropical cyclone and hydrologic modelling.
• Climate simulation, regional downscaling - eg., catchment
scale water balances.
- 11 -
SYSTEM AND BASEDATE/TIME
Number ofCPUs used
APPROXIMATE START TIME
SST Analysis (REGIONAL) 1 0115 UTC 15 min
LAPS_PT375 00UTC 4/8 0145 UTC 30 min
MESO_LAPS_PT125 00UTC 8 0200 UTC 120 min
* MESO_LAPS_PT050
(SYDNEY) 00UTC4
0210 UTC 30 min
* MESO_LAPS_PT050
(MELB) 00UTC4
0230 UTC 30 min
EER and atmospheric transportcalculations from LAPSsystems
10235 UTC 55 min
WAVES (REGIONAL andMesoscale) 00UTC
1 0315 UTC 10 min
TLAPS375 00UTC 8 0355 UTC 35 min
EER from TLAPS375 00UTC 1 0500 UTC 20 min
TC_LAPS 00UTC if required 8 0500 UTC 10 min
GASP 00UTC 4/8 0630 UTC 90 min
GASP ensemble – singularvector 1 0700 UTC 120 min
GASP – ensemble prediction 3 0900 UTC 240 min
EER from GASP 00UTC 1 0730 UTC 60 min
WAVES GLOBAL 00UTC 4 0730 UTC 20 min
SPECIAL charts 00UTC 1 0930 UTC 20 min
Multi-operationalsystem environment
- 12 -
Visualisation
- 13 -
WHAT IS NEEDED TO SUPPORT THIS EFFORT ?
• Improving hardware, but of relatively stable design.
• Robust hardware.
• Software which can evolve to take best advantage of the
hardware but is sufficiently stable so as to support older
code, robust data management and modern visualisation
(and the like).
• Use of industry standards.
• A mechanism to develop and maintain those standards
likely to be peculiar to meteorology.
- 14 -
FUTURE TRENDS
.
- 15 -
FUTURE TRENDS
.
- 16 -
FUTURE TRENDS
• Centres will specialise - no one will do it all.
• There will be greater, and more successful efforts to
integrate models from different disciplines.
• Systems will be improved incrementally (modular
architecture).
• End-to-end modelling, including data quality monitoring,
assimilation, analysis and prognosis, visualisation,
archival, product generation and dissemination will occur.
- 17 -
FUTURE TRENDS
• The ultimate goal is clearly earth-system simulation
• The ultimate architecture would appear to be clusters of
powerful computing and data storage environments (the
level of interaction between modules, and time-critical
nature of the various applications / modules will drive
processor power-proximity relationship).
• Data management in meteorology will accommodate
explosive increases in data volumes, and be synergistic with
other geophysical modelling efforts.
- 18 -
After:http//www.top500.org
- 19 -
After: http//www.top500.org
- 20 -
- 21 -
- 22 -
FUTURE TRENDS
• There will always be a role for operational meteorology -
and a need for operational high performance computing.
• Operational meteorology will also be a component of a
more integrated whole.
• There will need to be significantly greater collaboration
across the boundary between meteorology and the other
geophysical and biological scientists performing earth-
system simulation. This interaction will grow in time.
- 23 -
The operational meteorologists (short) wish list:
• Keep the hardware improving according to Moore’s
law;
• Maintain a software environment that protects our
existing investment in model code;
• Provide the capability to manage and visualise the
increasingly large datasets that models and remote
sensing are providing.
- 24 -
THANK YOU