Transcript PPT - Reading e-Science Centre

Global Hydrology Modelling and
Uncertainty: Running Multiple
Ensembles with the University of
Reading Campus Grid
Simon Gosling1, Dan Bretherton2, Nigel Arnell1 & Keith Haines2
1
Walker Institute for Climate System Research, University of Reading
2 Environmental Systems Science Centre (ESSC), University of
Reading
Outline
•
•
•
•
•
Uncertainty in climate change impact assessment
The NERC QUEST-GSI project & requirement for HTC
Modification to the CC impact model & Campus Grid
Results: impact on global river runoff & water resources
Conclusions & future developments
Uncertainty in Climate Change
Impact Assessment
Uncertainty in climate change impact
assessment
• Global climate models (GCMs) use different but
plausible parameterisations to represent the climate
system.
• Sometimes due to sub-grid scale processes (<250km)
or limited understanding.
Uncertainty in climate change impact
assessment
• Therefore climate projections differ by institution:
2°C
The NERC QUEST-GSI Project
and the Requirement for HTC
The NERC QUEST-GSI project
• Overall aim: To examine and assess the implications of
different rates and degrees of climate change for a wide
range of ecosystem services across the globe
• Our specific aims for global hydrology & water
resources:
A) To assess the global-scale consequences of different
degrees of climate change on river runoff and water
resources
B) To characterise the uncertainty in the impacts associated
with a given degree of climate change
The NERC QUEST-GSI project
• A) achieved by investigating impacts associated with the
following 9 degrees of global warming relative to present:
0.5 1.0 1.5 2.0 2.5 3.0 4.0 5.0 6.0ºC
• B) achieved by exploring impacts with the climate change
patterns associated with 21 different GCMs (climate
model structural uncertainty)
• Assessed impacts by applying above climate change
scenarios to the global hydrological model (GHM) MacPDM.09
– A global water balance model operating on a 0.5°x0.5° grid
– Reads climate data on precipitation, temperature, humidity,
windspeed & cloud cover for input
The challenge
GCM used to provide climate data
Prescribed Temperature
0.5
1.0
1.5
2.0
2.5
3.0
4.0
5.0
6.0
UKMO HadCM3
1
1
1
1
1
1
1
1
1
CCCMA CGCM31
3
3
3
2
3
3
3
3
3
IPSL CM4
3
3
3
2
3
3
3
3
3
MPI ECHAM5
3
3
3
2
3
3
3
3
3
NCAR CCSM30
3
3
3
2
3
3
3
3
3
UKMO HadGEM1
3
3
3
2
3
3
3
3
3
CSIRO MK30
3
3
3
2
3
3
3
3
3
CCSR MIROC32HI
4
4
4
4
4
4
4
4
4
CCSR MIROC32MED
4
4
4
4
4
4
4
4
4
CNRM CM3
4
4
4
4
4
4
4
4
4
GFDL CM21
4
4
4
4
4
4
4
4
4
GISS MODELEH
4
4
4
4
4
4
4
4
4
GISS MODELER
4
4
4
4
4
4
4
4
4
INM CM30
4
4
4
4
4
4
4
4
4
MRI CGCM232A
4
4
4
4
4
4
4
4
4
GFDL CM20
4
4
4
4
4
4
4
4
4
NCAR PCM1
4
4
4
4
4
4
4
4
4
BCCR BCM20
4
4
4
4
4
4
4
4
4
CCCMA CGCM31T63
4
4
4
4
4
4
4
4
4
GISS AOM
4
4
4
4
4
4
4
4
4
CSIRO MK5
4
4
4
4
4
4
4
4
4
Running on Linux Desktop:
• 1 run = 4 hours
• 1st Priority runs
9 runs = 36 hours
• 2nd & 3rd Priority runs
63 runs = 252 hours (~11 days)
• 4th Priority runs
189 runs = 756 hours (~32 days)
Running on Campus Grid:
189 runs = 9 hours
Modifications to Mac-PDM.09 and
the Campus Grid
Modifications to MacPDM.09
• Climate change scenarios previously downloaded
from Climatic Research Unite (CRU) at UEA and reformatted to be compatible with Mac-PDM.09
– Around 800Mb of climate forcing data needed for 1 MacPDM.09 simulation
– Therefore ~160GB needed for 189 simulations
– Integrated ClimGen code within Mac-PDM.09 as a
subroutine to avoid downloading
– Ensured all FORTRAN code was compatible with the
GNU FORTRAN compiler
• But the large data requirements meant the Campus
Grid storage was not adequate…
Campus Grid data management
• Total Grid storage only 600GB, shared by
all users; 160GB not always available.
• Solution chosen was SSH File System
(SSHFS http://fuse.sourceforge.net/sshfs.html)
• Scientist’s own file system was mounted on
Grid server via SSH.
– Data transferred on demand to/from compute
nodes via Condor’s remote I/O mechanism.
Campus Grid data management (2)
Using SSHFS to run models on Grid with I/O to remote file system
Campus Grid
Grid
server
...
Remote FS
mounted
using
SSHFS
Grid storage,
not needed
Data
transfer via
SSH
Large
file
system
Scientist’s data
server in
Reading
Campus Grid data management (3)
SSHFS advantages:
• Model remained unmodified, accessing
data via file system interface.
• It is easy to mount remote data with
SSHFS, using a single Linux command.
Campus Grid data management (4)
Limitations of SSHFS approach
• Maximum simultaneous model runs was 60
for our models, implemented using a
Condor Group Quota
– Can submit all jobs, but only 60 allowed to run
simultaneously.
– Limited by Grid and data server CPU load
(Condor load and SSH load)
• Software requires sys.admin. to install.
• Linux is the only platform
Campus Grid data management (5)
Other approaches tried and failed
• Lighter SSH encryption (Blowfish)
– No noticeable difference in performance
• Models work on local copies of files
– Files transferred to compute nodes before runs
– Resulted in even more I/O for Condor
– Jobs actually failed
• Mount data on each compute node
separately
– Jobs failed because data server load too high
Results
Global Average Annual Runoff
Multiple ensembles for various prescribed
temperature changes
18
81
9 model runs
The ensemble mean
Global Average Annual Runoff Change from
Present (%)
But what degree of uncertainty is there?
Uncertainty in simulations
Number of models in agreement
of an increase in runoff
Results
Catchment-scale Seasonal Runoff
The Liard
The Okavango
The Yangtze
Seasonal Runoff
Agreement of increased snowmelt induced runoff
Less certainty
regarding
wet-season
changes
Agreement of dryseason becoming
drier
Large uncertainty throughout
the year
Results
Global Water Resources Stresses
Calculating stresses
• A region is stressed if water availability is
less than 1000m3/capita/year
• Therefore stress will vary according to population
growth and water extraction:
– Stress calculated for 3 populations scenarios in the 2080s
• SRES A1B
• SRES A2
• SRES B2
• Calculated for different prescribed warming (0.56.0ºC)
Global water resources stresses
Global Increase in Water Stress with 2080s A1B Population
The range of uncertainty
Global Increase in Water Stress with 2080s A1B Population
Conclusions
• HTC on the Campus Grid has reduced total simulation
time from 32 days to 9 hours
– This allowed for a comprehensive investigation of climate
change impacts uncertainty
– Previous assessments have only partly addressed climate
modelling uncertainty
• e.g. 7 GCMs for global runoff
• e.g. 21 GCMs for a single catchment (we looked at 65,000)
• Results demonstrate:
– GCM structure is a major source of uncertainty
– Sign and magnitude of runoff changes varies across GCMs
– For water resources stresses, population change uncertainty
is relatively minor
Further developments
• Several other simulations have just been completed
on the Campus Grid & are now being analysed:
– NERC QUEST-GSI project:
 204-member simulation
 3 future time periods, 4 emissions scenarios, 17 GCMs (3x4x17=204)
 816 hours on Linux Desktop - 10 hours on Campus Grid
– AVOID research programme (www.avoid.uk.net)
 Uses climate change scenarios included in the Committee on Climate
Change report
 420-member simulation
 4 future time periods, 5 emissions scenarios, 21 GCMs (4x5x21=420)
 70 days on Linux Desktop – 24 hours on Campus Grid
– 1,000-member simulation planned to explore GHM
uncertainty
Further developments
Forcing repositories at other institutes
• Forcing = hydrological model input
• Avoid making local copies in Reading
• Additional technical challenges:
– Larger volume of data (GCMs not run locally)
– Slower network connections (for some repos.)
– Sharing storage infrastructure with more users
– No direct SSH access to data
Further developments
• Possible solutions
– Mount repos. on compute nodes with Parrot
(http://www.cse.nd.edu/~ccl/software/parrot)
• This technique is used by CamGrid
• Parrot talks to FTP, GridFTP, HTTP, Chirp + others
• No SSH encryption overheads
– May need to stage-in subset of forcing data
before runs
• Options include Stork (http://www.storkproject.org/)
Acknowledgements
The authors would like to thank David Spence and the Reading
Campus Grid development team at the University of Reading for
their support of this project.
Thank you for your time
Visit www.walker-institute.ac.uk