HIRLAM AT KNMI

Toon Moene

Below you find an overview of the operational setup of HIRLAM at KNMI.

Operational compute power

SGI Origin 2000 with 16 180 Mhz R10000 processors, 2.5 Gbyte of main memory, 1 Mbyte of cache.

This machine is also used for development work. Operational HIRLAM gets 12 of the 16 processors using a software product called MISER.

Operational HIRLAM: 0.5 x 0.5 degrees

A 48 hour forecast is made every 6 hours on the area indicated below. The analysis cycle is 3 hours. The cut-off time for observations is 2 hours.

Hirlam 4.3.4 code base.
Eulerian advection.
Straco convection/condensation scheme.

Dynamics time step: 4 minutes.
Physics time step: 4 minutes.

Grid: 164 x 130 x 31. South pole at 15 W, 30 S.
Nested in 1.5 x 1.5 ECMWF boundaries, on a rotated grid oriented the same as Hirlam grid.
Resources needed:

Analysis: 150 seconds

Surface obs: 3000
Aireps: 200-2000
Satobs: 200
Temps: 130
Pilots: 10

Forecast: 2700 seconds

The large extent of this area to the east and south is due to a request from Dutch public television, which want to show the whole Mediterranean area with their weather presentations.

High resolution HIRLAM: 0.1 x 0.1 degrees

A 12 hour forecast is made every 3 hours on the area indicated below. The analysis cycle is 3 hours. The cut-off time for observations is 1 hour.

This run is semi-operational, in the sense that the output isn't guaranteed although during office hours assistance in case of problems is available. The distribution of the output is the same as for the "normal" HIRLAM run.

Hirlam 4.6.3 code base.
Semi-Lagrangian advection.
Straco convection/condensation scheme.
Holtslag vertical diffusion.

Dynmaics time step: 3 minutes.
Physics time step: 3 minutes.

Grid: 164 x 150 x 31. South pole at 0 W, 35 S.
Nested in 3-hourly 0.5 x 0.5 operational HIRLAM boundaries.
Resources needed:

Analysis: 450 seconds

Surface obs: 1000
Aireps: 20-600
Satobs: <10
Temps: 25
Pilots: <3

Forecast: 3300 seconds

After upgrading from 4.3.4 to 4.6.3, we didn't manage anymore to complete the run in the hour before the main HIRLAM run, so now we run it after the main job.

The following map shows what the extra resolution may buy: A realistic wind field over the IJsselmeer, and katabatic winds down the Ardennes.

Future Plans

This year we'll acquire a new computer system for operations. The performance goal of the machine will be:

For the 48 hour forecast:

0.2x0.2 degree resolution over the full area.
Semi-Lagrangian advection with a 6 minute time step.
(If possible) 40 layers - double the resolution in the PBL.

For the 12 hour forecast:

0.05x0.05 degree resolution over the small area.
Semi-Lagrangian advection with a 90 seconds time step.
(If possible) 40 layers - double the resolution in the PBL.

The necessary CPU power will be a factor of 10 higher than what we use now.

We expect to be offered distributed memory systems; if we cannot use 3D-Var data assimilation yet at the beginning of next year, we'll use the OI analysis on our current machine.