IBM demonstrates Deep Thunder
at the Supercomputing 1999 conference
The capabilities developed and utilized over the last three years in various
venues were employed at the Supercomputing
'99 Conference (SC99: November 13 - 19, 1999 in Portland, OR).
Deep Thunder was replicated in the IBM booth as part of the conference's
technical exhibition. The system was adapted to the Portland area.
Significant new capabilities for Deep Thunder were developed
for SC99. For the first time, "nested" forecasts at 16, 4 and 1 km
resolution (areas of 976x976, 244x244 and 61x61 km in size, respectively)
centered over Portland tied to multi-resolution visualizations were implemented
to make live predictions during the conference. Previously, only
a single resolution had been supported. This new capability is critical
for a number of commercial Deep Thunder applications, where the domain
of the forecast is tailored to the geographic region of interest enabling
one to "zoom in" on predictive forecasts. Maps of each of these nests
are shown below. For these and any of the subsequent images,
you can view a higher-resolution version by simply clicking on it.
You can also interact with this map via a scene in PanoramIX,
simplified VRML, or view a flyover
animation (or at higher resolution).
You can also interact with this map via a scene in PanoramIX,
simplified VRML.
You can also interact with this map via a scene in PanoramIX
or simplified VRML. To illustrate these capabilities,
consider a simple animation which
shows predicted clouds as an isosurface of total cloud water density registrated
with the map of the outer (16 km resolution) nest. This forecast
was produced on November 15.
The performance and resolution enhancements were mapped to three distinct
visualization and analysis applications. In addition to general improvements,
the visualization tools were adapted for multi-resolution operations via
a new method to encapsulate access to nested data. A 3d facility
for interactively browsing model results and tracking the simulation was
used, which includes flyover and time-based animation of various weather
variables, and "snapshots" for web access as PanoramIX scenes, simplified
VRML geometry and images. It works with data generated every 10 minutes
of forecast time. Two applications for the analysis of the post-processed
model output at hourly resolution provided complementary facilities.
One was focused on the entire 3d model domain while the other emphasized
surface and upper atmospheric layers. Details about these visualization
applications are available in a paper for
you to read, and examples are available further down this page.
These nests are shown in the images below produced by the browser application.
Each of the images illustrates cloud properties and surface temperature,
pressure and winds at 1 pm on November 16, 1999. The images show
a terrain map, pseudo-colored by contours of predicted surface temperature
overlaid with coastline, national boundary, state, county and river maps.
Some major cities in the area are also marked with predicted temperatures.
Forecast winds are illustrated by streamlines with directional arrows,
colored by speed. The streamlines are integrated from starting positions
identified as being critical points in the predicted wind flow. Clouds
are visualized as a white, translucent isosurface of cloud water density.
Inside the cloud surfaces are cyan surfaces of predicted reflectivity.
Points of high and low pressure, respectively are marked on the maps.
16 km nest
4 km nest
1 km nest
A similar representation for a forecast produced on November 17 is shown
via animation, but without the pressure
markers. It combines the 4 km and 1 km nests, which can be seen in
the area of higher resolution around Portland. It can also
be viewed at lower resolution. The
approach to wind visualization for the multi-resolution meshes captures
some of the orographically induced flows, such as down the Columbia River
Gorge.
Running in parallel on ten 200 MHz 2-way POWER3 nodes for compute and
an additional one for I/O, a 48-hour forecast was completed in about 8
hours for all three geographic nests. Comparing this performance
to previous experiments is a bit difficult because these forecasts are
far more computationally challenging. In addition to operating over
three nests, POWER3 cpus were used for the first time, additional vertical
layers were computed (31 vs. 28), a slower cloud microphysics package was
employed, the internal time step was much smaller on the inner meshs (48,
12 and 3 seconds, respectively vs. 30 seconds), and the forecasts were
run for two days rather than one. Crudely, one can note that the
total number of grid points for all time steps was about 5.5 larger other
recent efforts. Given that updates for the input data were available
at the conference every 12 hours, this level of performance was good as
well as practical for the size of the machine made available. This
has very strong implications for affordable applications.
Three workstations (two IBM RS/6000 43P-260s and an IBM Intellistation
M-Pro) and three laptops (IBM Thinkpads 770Z and 600E and RS/6000 860)
were available to interact with the model and analyze results. A
lecture on Deep Thunder was given on Tuesday, November 16 in the
IBM booth.
Due to a number of logistical difficulties, raw observations and access
to data assimilation via LAPS for the pre-processor step were unavailable
at SC99. Therefore, RAMS was initialized with the results from the
ETA synoptic scale model from NCEP, which are computed at 32 km resolution,
but sampled at 40 km for public availability. These same data were
also used for boundary conditions for the model. Since this model
is run only twice per day, two Deep Thunder runs were completed
each day of the conference.
These capabilities attracted conference attendees, the competition and
the press. For example, interactive results from Deep Thunder
were broadcast by all three local television stations (ABC, CBS and NBC
affiliates) during the conference.
After each RAMS execution, all of the results are collected and reorganized
into a form that can be used by standard meteorological analysis tools
as provided by NWS (e.g., AWIPS). These post-processed data were
made available for two interactive applications. This includes all
of computed variables from the model, but at hourly resolution unlike the
browser application that worked with a subset of variables but at six times
the temporal resolution. Here is a sample image and animation
created with one of the applications for the previously discussed forecast
run initiated on November at 17Z UTC (9 AM PST). This application
is a "slicer", which provides two- and 2-1/2-dimensional interaction with
surface and upper layers of the model data. Additional details about
the application is available in a paper
that discusses the visualization portion of Deep Thunder.
Four different surface variables have been selected in a combined visualization
with the 4 and 1 km nests integrated. Temperature is shown as pseudo-color.
Wind velocity is illustrated as streamlines with directional arrows arrows,
colored by speed. Colored line contours of relative humidity in increments
of 10% are shown. These planar representations are deformed vertically
by maximum reflectivity to create a shaded surface. A coastline map
(black) and state boundaries (white) are draped on the surface. Any
of the surface or upper air fields available from the model can be visualized
with any of these methods.
The image below is from the same forecast but produced with a 3d viewer,
which combines upper air field visualized volumetrically with surface data.
An 48-hour animation for the same period
can be viewed.
A surface variable (total precipitation) has been selected for display
as pseudo-colored filled contour bands, which are overlaid on a topographic
map. Any of the surface variables produced by the model may be presented
in this fashion. Coastlines (black), state boundaries (white) and
rivers (blue) are draped on the surface. An upper air variable (relative
humidity) has been selected for display via surface extraction. The
surface at 90% is requested in translucent white, which corresponds roughly
to a cloud boundary. Another field (reflectivity) has been selected
to show as a vertical slice, which is pseudo-color contoured. Gaps
in the slice correspond to regions where reflectivity is zero. Any
of the three-dimensional fields available from the model can be visualized
with either of these methods. The upper air wind data can be seen
along two vertical profiles, which are specified interactively, and via
streamribbons. The direction of the model wind field along these
"virtual sounding" is shown via vector arrows. Both the arrows and
ribbons are pseudo-colored by horizontal wind speed. The length of
the arrows also corresponds to the horizontal speed. Points along
the profile are used as seeds for the streamribbon integration. Each
profile is realized as a pseudo-colored tube, which is contoured by the
variable selected for isosurface realization (i.e., relative humidity).
The streamlines are marked for interactive comparison with a profile plot.
Profile #2 is in the 16 km nest and is not visible in this image.
To evaluate these model results, it is useful to compare
them to actual observations as well as
other model results.
lloydt@watson.ibm.com
Last updated January 4, 2000
