COSMO model at ARPAE-SIMC

Last updated: 10 Jun 2020

Basic model set-up

Arpae-SIMC operates the COSMO model in two main configurations, one with a 5 km grid spacing over the Mediterranean area (COSMO-5M) and one with a 2.2 km grid spacing over Italy (COSMO-2I).

The domains of COSMO-5M (left) and COSMO-2I (right) used at Arpae-SIMC

Operational configuration

The run over the Mediterranean area uses the CNMCA LETKF determistic analysis as initial condition and ECMWF-IFS as boundary conditions, it is run twice a day, 00 and 12 UTC up to 72 hours on the Cineca HPC computing systems. The model configuration and grid is analogous to the one operated by CNMCA (COSMO-ME).

A backup run with the same resolution but on a subset of the Mediterranean grid is run on the internal Arpae-SIMC HPC system, using continuous data assimilation with nudging and DWD-ICON lateral boundary conditions.

The higher resolution pre-operational suite at 2.2 km includes a data assimilation run and several forecast runs.

The data assimilation run is based on Kenda system and on an ensemble background of 36+1 members. The boundary conditions for the ensemble members come from CNMCA LETKF ensemble data assimilation cycle (7km). The deterministic member takes the boundary conditions from the run at 5 km on Mediterranean area. Latent heat nudging is also applied to ensemble members using the Italian national radar composite.

The forecast runs are divided into:

• deterministic run, twice a day, 00 and 12 UTC up to 48 hours
• RUC (Rapidly Updating Cycle) deterministic run, 8 times a day up to 18 hours
• ensemble run, once a day at 21 UTC up to 51 hours, 20 members.

The deterministic runs use lateral boundary conditions from the run at 5 km on Mediterranean, while the ensemble forecast uses boundary conditions provided by the CNMCA 7 km COSMO-ME EPS. Initial conditions are provided by Kenda ensemble and deterministic analysis.

HPC systems

The main computing system used in Cineca center is a cluster of about 400 Intel-based (Broadwell) computing nodes with an Infiniband interconnection. Up to 22 nodes (~700 cores) are used for a single model run, while the number of nodes devoted to the COSMO operational procedures grows to about 80 when the required computational load is high.

The backup HPC system at Arpae-SIMC has 30 dual socket diskless nodes, connected by an Infiniband QDR network.

Detailed configuration