250 likes | 381 Views
External Projects related to WG 6. COSMO Working Group 6: Reference Version and Implementation Working Group Coordinator: Ulrich Schättler. Interoperability Priority Project POMPA: Performance On Massively Parallel Architectures. Contents. Status of the Interoperability Programme.
E N D
External Projects related to WG 6 COSMO Working Group 6: Reference Version and Implementation Working Group Coordinator: Ulrich Schättler
Interoperability Priority Project POMPA: Performance On Massively Parallel Architectures Contents COSMO General Meeting 2010
Status of the Interoperability Programme Ulrich Schättler DWD Philipp Glatt MeteoSwiss
To support future collaborative effort in European Limited Area Modelling by: Defining a standard output format Providing software tools (adaptors) to convert LAM output to the standard format Enabling European LAMs to run from appropriate alternative model fields To display or use in NMHS post-processing system To provide operational backup Idea of Interoperability COSMO General Meeting 2010
Plans for the Adaptor For Parameter List 1: ps, psml, t2m, td2m, … GME UM ARPEGE IFS Fieldextra Read (Write) Grib 2 and transform to COSMO-Model grid: only 2D fields UM_local ALADIN HIRLAM COSMO COSMO General Meeting 2010
Plans for the Adaptor For Parameter List 2: u, v, w, t, p, qx, … to be done status unclear GME ARPEGE GFS? GSM? INT2LM Read (Write) Grib 2 and transform to COSMO-Model grid: 2D fields, full atmosphere, soil parameters ALADIN HIRLAM COSMO COSMO General Meeting 2010
lum2lm (Grib1): Temperature (lowest COSMO Level) from GME IFS UM Aemet UM DWD COSMO General Meeting 2010
lum2lm (Grib1): PP from reference pressure (lowest level) from GME IFS UM Aemet UM DWD COSMO General Meeting 2010
lum2lm (Grib1): Wind speed (lowest level) from GME IFS UM Aemet UM DWD COSMO General Meeting 2010
In December 2009, the “Interoperability Team” met in Reading for discussing last years work and do some planing for 2010. The first 2 deliverables were finally agreed: The Standard Format shall be the native grid and native parameters Documentation supplied with the standard format should be the minimum required Some Grib 2 Issues were discussed: There is a problem with coding the vertical coordinate / vertical grid of COSMO in Grib 2 (see tomorrows session on Postprocessing) Workshop in Reading COSMO General Meeting 2010
Vertical coordinate / vertical grid of the COSMO-Model A COSMO-internal discussion about coding the vertical grid in Grib2 has been initiated early 2010 For the Gal-Chen height based hybrid coordinate it is possible to code the grid with typeOfLevel=118, the corresponding orography (as a 2D field) and vertical coordinate parameters This is NOT possible for e.g. the SLEVE coordinate The COSMO discussion led to a proposal to the WMO Commission for Basic Systems (IPET-DRC: Inter-Programme Expert Team on Data Representation and Codes). A new vertical coordinate type is proposed, where no vertical coordinate parameters together with a 2D field are given, but a full 3D field with the heights of every grid point is specified. The proposal has been discussed also with Enrico Fucile (ECMWF) Work in 2010 COSMO General Meeting 2010
grib_api grib_api has been established at some COSMO centres to be the official Grib(2) library (DWD, MeteoSwiss, ARPA-SIM Bologna) Work is in progress to improve performance on NEC machines (together with NEC, MeteoFrance and ECMWF) It has been implemented in FieldExtra for reading and writing. FieldExtra therefore can now produce the parameters of the standard output format list directly Work is in progress, to implement it into INT2LM Work in 2010 COSMO General Meeting 2010
New Grib 2 test data set and documentation A new Grib2 test data set, which has been produced by FieldExtra, has been provided. This still has the old (Grib1-style) coding of the vertical coordinate No Grib2 test data has been provided with typeOfLevel=118, because this might not be the default case any more (if proposal is accepted) No surface test data (according to the guidelines from the ET on Surface Aspects) has been provided up to now The minimum required level documentation has not been updated according to the plans from last workshop describe native grids describe fields in the standard output list: what field is it, what does it represent in the generating model Work in 2010 COSMO General Meeting 2010
Priority Project POMPAPerformance on MassivelyParallel Architectures (To be confirmed by STC) Project Leader Oliver Fuhrer MeteoSwiss
Computer architectures are about to change. Multi-core, many-core systems are on their way. Many people talk about General Purpose GPUs. But first: Identify the problems of the COSMO-Model on existing machines! Background: Future Computer Architectures COSMO General Meeting 2010
Fraunhofer / SCAI investigated LM_RAPS_4.0 on IBM BlueGene The RAPS code is not as optimized on NEC SX-9 as the current operational model is Timings for COSMO-DE 421 461 50 grid points, dt=25 s Modern architectures surely are more similar to BlueGene than they are to SX-9 There are problems in the I/O and the overall scalability! Investigations from Fraunhofer / SCAI on BlueGenewith RAPS 4.1 COSMO General Meeting 2010
Scalability of COSMO-EU: 665 657 40, 24h COSMO General Meeting 2010
Scalability of COSMO-DE: 421 461 50, 21h COSMO General Meeting 2010
Scalability of COSMO-Europe: 1500 1500 50, 2.8 km, 3 h, no output COSMO General Meeting 2010
I/O: Accessing the disks and the global communication involved disturbes scalability heavily. Although the communications besides I/O are almost all local, the speedup degrades when using many processors. What cannot be seen on the pictures before: Although the speedup of the computations is not bad, the efficiency of the code is not satisfying: NEC SX-9: 13 % of peak IBM pwr6: about 5-6 % of peak Cray XT4: about 3-4 % of peak (?) The Problems of the COSMO-Code … COSMO General Meeting 2010
Swiss national initiative HP2C High Performance High Productivity Computing Plan: Enable software to run efficiently on new computer architectures (multi-core, many-core, GP-GPUs, etc.) ETH Zürich is taking part with the COSMO-CLM: HP2C-CCLM This project has 3 subtasks Cloud-resolving climate-change scenarios Preparing the CCLM and existing model extensions for future computers Refactoring CCLM components for emerging architectures HP2C-CCLM runs from Spring 2010 to December 2012 More Background COSMO General Meeting 2010
POMPA is the „missing link“ between the Swiss national initiative and our consortium POMPA gives additional resources in the optimization of the COSMO-Model for future computer architectures POMPA can ensure that the operational COSMO-Model really benefits from the outcomes of HP2C-CCLM Finally: PP POMPA COSMO General Meeting 2010
Thank you very much for your attention COSMO General Meeting 2010