Development of High Resolution Global NWP Model
Ken-ichi Kuma
Numerical Prediction Division
Japan Meteorological Agency
1) Issues to be solved
The current JMA global model has the resolution of T213 which corresponds to about 60km
mesh, while the regional model has the resolution of 20km mesh. In order to provide the
global scale to regional scale forecast with a unified model, we need to develop the very high
resolution global model( about 20km mesh). To realize such high resolution global model,
there are several issues to be solved as follows;
1 Semi-Lagrangian time integration scheme to improve computational efficiency.
2 Assessment of computational requirement in high resolution global spectral model.
3 High resolution global data such as topography.
4 Physical processes which can work properly with 20km mesh.
elapse time*number of nodes
2) Current development
At Japan Meteorological Agency, the development of Semi-Lagrangian global model is under
way. We have completed the shallow water version of the Semi-Lagrangian model and now
expanding to 3D model. It is expected that
Semi-Lagrangian
model
will
be
60000
implemented into the operation NWP model
52320
in early 2003. In 2006, it is expected to
50000
th
introduce the 8 generation super computer
40000
by which we can run TL999 (Truncation at
total wave number 999 with linear grid)
30000
which corresponds to 20km (or less) mesh
20600
20000
resolution.
Our Semi-Lagrangian model will still follow
10000
the spectral formulation. Thus, the cost for
5696
Legendre
transformation
and
data
0
communication among computational nodes
100
200
300
400
500
for wave-grid transformation must be
R esolution (truncation w ave num ber)
assessed under the real computational
environment. Since we cannot run TL999 on
Fig.1 The relation between horizontal
the current JMA computing system, we are
resolution (wave truncation) and total elapse
to use T426 resolution with 40 nodes and
time(second)
compare the performance with the
operational T213 resolution.
The experiment is conducted on Hitachi's SR8000E, which has 80 nodes with distributed
memory (8 Gbytes each). Each node has a peak performance of 9.6GFLOPS with 8 RISCbased CPUs. Fig.1 shows how computational cost increases as the horizontal resolution
increases. It should be noted that T213 is computed with 16 computational nodes, while
T319 and T426 are computed with 40 nodes. T426 model takes 9 times larger total
computational time than T213 model. Considering that the doubled horizontal resolution
requires 2X2X2=8 times computation, the overhead due to Legendre and communication is
satisfactory small at T426. It is also noted that the actual computational speed is about 16 %
of the theoretical peak performance in T213 model even if we include the time for I/O (about
2-3% of total time). This computational performance is notably higher than that for the other
RISC based machine.
Fig.2 shows ratio for execution time in major components of the model. As the horizontal
resolution increases, the ratio of physical process (PHYSCS) is reduced while that of
dynamical
process
Execution time ratio for major subroutines
(TNDNCY) increases.
It should be noted
that the moist process
T426
(GMOIST)
includes
both moist physical
PHYSCS
processes and waveTNDNCY
grid transformation.
T319
TINTGS
TINTGS includes the
GMOIST
process related to
semi-implicit
time
T213
integration which is
computed on wave
space.
0%
20%
40%
60%
80%
100%
High
resolution
global data includes
Fig.2 The execution time ratio for major subroutines with different
those used for the
horizontal resolution.
initial conditions and
those used for the boundary conditions. As to the former category, we need to collect the
satellite data with the original resolution. For instance, ATOVS (Advanced TIROS
Operational Vertical Sounder) data has 17.4km mesh resolution which needs to be used for
20km mesh model. One of the most important boundary condition data is the global
topography. For the future high resolution model, we have replaced Navy's 10 minites data
by GTOPO's 30 seconds data.
Fig.3 shows the topography and precipitation in 24 hour for T213 and T426 models. The
difference of precipitation over the ocean is very small, suggesting the robustness of cumulus
Fig.3 Topography (contour, 500m interval) and 24hour precipitation (shaded, mm) in
T213 (left) and T426 (right) model. The initial time is 12 UTC on July 4th 2000.
vertical level
parameterization
over
the
different
resolution. On the other hand, over the
The levelw hich determ ines the C FL.
southern slope of Tibetan plateau, the
40
precipitation pattern differs significantly. It
37
is of interest to investigate whether this
34
difference gives the different general
circulation through a feedback mechanism.
31
We need to run the high resolution model for
28
long time range, which is feasible in the
25
Earth Simulator.
22
The performance of the physical processes in
the higher resolution is not verified yet.
19
However, our experience in the regional
16
model (about 20km mesh resolution) implies
0%
10%
20%
30%
40%
50%
60%
the possible problems. Our regional NWP
model often exaggerates the intensification
Fig.4 Histogram of occurrences for each
for the extra-tropical cyclone. It is not clear
vertical level to determine the CFL condition.
that either the high resolution model or
The statistics for 11 month integration with
physical processes in the regional model
T106L40 model.
causes the problem. In order to separate two
issues, we are planning to implement the physical processes of the global model into the
regional model.
We have also increased the
number of model vertical layers
from 30 to 40 and lifted the model
top from 10hPa to 0.4hPa. One of
the problems arises from this
change is the strong wind at the
top level which reduces the time
interval of integration, thus
increases the computation. Fig.4
shows
the
histogram
of
occurrences for each vertical level
to determine the CFL condition.
More than half of the cases, the
highest level determines the CFL
Fig.5 Simulated temperature at 1hPa in July. Contour
condition. It is partly explained
interval in 10K
by the fact that the wind speed in
the upper stratosphere is high.
However, in the winter for the southern hemisphere, wind speed exceeds more than 250m/s,
which seems to be unrealistic.
Fig.5 shows the monthly mean temperature at 1hPa in July obtained from 11 month
integration of T106L40 model. The temperature in the polar night is unrealistically low
(190K) and close to the radiative equilibrium temperature. This implies the lack of some
dynamical processes in the model in the polar night stratosphere.
3) Summary
JMA is planning to develop the very high resolution global model which corresponds to
20km mesh model. The feasibility of the global model with such high resolution must be
examined in various aspects. We have already developed the high performance global model.
This model needs to be run intensively under the high resolution environment.