Pointers for the Future - Environmental Modelling: Finding Simplicity in Complexity - page 444

spatial and temporal discretizations. These problems
can be addressed by developing expensive high per-
formance computing clusters and parallel processing
techniques such as the UK's Met Office and NERC
joint supercomputer system (MONSooN) 29 or NASA's
Earth Exchange (NeX 30 ) - which brings together super-
computing capacity with remote sensing data feeds and
a framework for model application. Such systems are
clearly expensive to acquire and maintain and out
of reach for many modellers and applications. Net-
worked grid computing is an alternative that brings
together existing computer infrastructure (an organi-
zation's desktop and laboratory computers) and puts
them to work on complex modelling tasks when they
are not being used for other activities. The computers
used can be distributed within an institution or across
the web (such as the climateprediction.net volunteer
computing service). One of the most established vol-
unteer grid networks is the so-called BOINC (Berkeley
Open Infrastructure for Network Computing) network,
which is said to have more than 451 000 active com-
puters and is thus of greater computing capacity than
any single supercomputer. Within institutions a variety
of so-called 'middleware' programs such as Condor 31
can be used to divide computational tasks into chunks,
which are then distributed to nodes in the grid where
they are processed remotely, returned and re-assembled
into the required solution. Grid computing dramati-
cally cuts the cost of supercomputing but cannot easily
handle problems in which calculations must concur-
rently interact with each other, where data-transfer
requirements are huge or where the problem cannot
easily be broken down into chunks. Such problems
usually need significant rethinking for this type of grid
based parallelization.

and software product. Only the service used is paid for.
The system is stored as an image of the operating system
and software and can be applied to as many instances
(computers of a given specification) as are required
and these instances can connect to databases as neces-
sary. The modeller never sees the hardware and there is
no clear relationship between a software instance and a
hardware server since the instances run in a 'virtualized'
software environment where many operating systems
and configurations can operate on a single server. A
number of universities are developing their own cloud
systems and a number of commercial services exist
such as Amazon EC2, 32 Microsoft Windows Azure 33
and Google App Engine. 34 Such services hold promise
to separate modelling from computing needs, so that
modellers can focus on the modelling not the admin-
istration of computing systems. The only cloud system
to be focused on environmental applications so far is
the so-called Google Earth Engine project. 35 This is a
project of Google.org and is still in closed beta but essen-
tially provides a range of earth observation datasets and
a so-called API (Application Programming Interface)
that provides a limited computing environment for cal-
culations using these data and visualization of results of
those calculations using Google Earth. The calculations
are performed on Google's computing infrastructure
and are thus highly scaleable to even huge datasets
(held by Google). Neveretheless the system is some-
what limited for most modelling purposes whilst it
remains in closed beta form.

We expect developments in hardware computing
capacity (especially disk storage and speed, memory
and processing power as well as network speed) to
continue to facilitate ever more sophisticated modelling.
The increasing availability and ease of use of remote
sensing data products from the likes of NASA's MODIS
instrument, the Space Shuttle Topography Mission and
others will facilitate their use in spatial modelling. A
number of geodata portals such as the KCL geodata
portal 36 and the Consortium for Spatial Information 37
now provide such datasets in very easy-to-use formats.

Increasingly, modellers are not necessarily computer
whizzes. Maintaining your own set of networked mod-
elling computers whether running Linux, Mac or Win-
dows can be a bind, especially if they are only used
occasionally. Cloud computing offers access to fully
controllable computing resources that can be config-
ured as the modeller wishes and which are scalable as
computing needs change. This computing resource is
delivered as a software service rather than a hardware

•

32 See http://aws.amazon.com/ec2/ (accessed 6 April 2012).
33 See www.windowsazure.com/en-us/ (accessed 6 April 2012).
34 See http://code.google.com/appengine/ (accessed 6 April 2012).
35 See http://earthengine.google.org/#state 29 See www.metoffice.gov.uk/research/collaboration/jwcrp/

monsoon-hpc (accessed 6 April 2012).
30 See https://c3.nasa.gov/nex/about/ (accessed 6 April 2012).
31 See http://research.cs.wisc.edu/condor/ (accessed 6 April 2012).

= intro (accessed 6 April

2012).
36 See www.kcl.ac.uk/geodata (accessed 6 April 2012).
37 See SSwww.cgiar-csi.org/ (accessed 6 April 2012).