aro ald informalsession

Slide1: 

AROME workshop Prague, 11-12 April 2003

Slide2: 

Agenda of the workshop Friday morning : parallel sessions Scientific meeting Info about history of the (CSSI+LSMG) AROME/ALADIN link Friday afternoon : plenary meeting       Presentations, questions Saturday morning: plenary meeting        Conclusions, draft of implementation plan

Slide3: 

• Basics on what AROME is (or will be) • History of the AROME-ALADIN link • About computing resources More precisely this morning

Slide4: 

What the acronym means… Application de la Recherche à l'Opérationnel à Méso-Echelle Application of Research to Operations at MEsoscale Basics on what AROME is (or will be)

Slide5: 

A Limited Area Model at 2.5 - 3 km resolution a scale where intense convection becomes explicit in other words, where the atmosphere is really treated as 3D (of course non-hydrostatic) Still a lot of physical parametrizations … some of them quite expensive (micro-physics, 3D-turbulence) Basics on what AROME is (or will be)

Slide6: 

A system with fine-scale oriented data assimilation Currently ALADIN has: - either no assimilation (ex: Météo-France) - or assimilation of low density observations AROME will use satellite, radar… and classical obs. One of the key issues: the water cycle Basics on what AROME is (or will be)

Slide7: 

A system with fine-scale oriented data assimilation satellite:    not only ATOVS or IASI (10-40 km, 4-6h)    also MSG info (3-5km, 15mn) - water vapour - clouds even if more difficult to use Basics on what AROME is (or will be)

Slide8: 

A system with fine-scale oriented data assimilation radar:    the best space/time resolution    the richest info on water also not easy to use Basics on what AROME is (or will be)

Slide9: 

A system meant to become operational Objective for Météo-France : 2008-2010 AROME will be the core of our short range NWP system Among the goals: - general improvement of the forecast ex: low clouds or clear sky ? - severe weather conditions - better response to accidents Basics on what AROME is (or will be)

Slide10: 

A system meant to become operational A case study for illustration: Aude floods November 1999 1980 km 1780 km 320 km 360 km 2 versions of Meso-NH research model (10km, 2.5km) Basics on what AROME is (or will be)

Slide11: 

A system meant to become operational A case study for illustration: Aude floods November 1999 Cumulated rainfall 12 Nov. 1999 12 UTC to 13 Nov. 1999 06 UTC Basics on what AROME is (or will be)

Slide12: 

A type of objectives shared by other NWP groups US: the WRF project (http://www.wrf-model.org) A coordinated work NOAA-NCAR-a few others Objectives: almost the same (1-10 km + assim) UK: since the "new dynamics" (2002) they have the tools but they didn't choose the path yet HIRLAM: also share the objective a strategy for assimilation, try to find one for model Not forgetting DWD, the first having dreamed of it… Basics on what AROME is (or will be)

Slide13: 

Discussion

Slide14: 

The early stage: an internal M-F discussion 19th Jan 2001: discussion paper of F. Bouttier la définition d'un modèle dynamique de prévision atmosphérique régional non-hydrostatique à échelle fine. Une option envisagée est une fusion logicielle entre Aladin et méso-NH, mais ce n'est pas le seul choix possible. Ce débat intéresse au premier chef GMME et GMAP, ainsi qu'en second plan le CEPMMT, le LA, et la communauté Aladin. of course ALADIN (the partners, not only the model) from the beginning taken into consideration History of the ALADIN-AROME link

Slide15: 

The early stage: an internal M-F discussion March 2001: the "Meso" project La phase de développement du projet méso s'étalant sur plusieurs années, il n'y a pas lieu de s'inquiéter à court terme sur l'avenir des engagements pris: ni Méso-NH, ni Aladin ne sont en danger. A l'horizon 2005-2006, cependant, une redéfinition de certaines collaborations sera nécessaire, que l'on s'efforcera de rendre la plus positive possible en fonction de la direction prise par les développements à ce moment-là. We are not waiting until 2005-2006 ! History of the ALADIN-AROME link

Slide16: 

The early stage: an internal M-F discussion Until automn 2001: working groups assimilation, dynamics, convection, physics in general, coupling issues, chemistry, sea surface, land surface, validation, software management The one about dynamics set up a frame for the comparison of the possible solutions Meso-NH, ALADIN-NH, imported, new ? History of the ALADIN-AROME link

Slide17: 

The early stage: an internal M-F discussion From this early stage it was clear that (among others): - AROME assimilation would be derived from   ALADIN 3D-Var the development that M-F didn't want to do… not 4D-Var because of "too incremental" - AROME physics would be derived from Meso-NH - the choice for the dynamics was more difficult History of the ALADIN-AROME link

Slide18: 

The early stage: an internal M-F discussion From this early stage it was clear that (among others): - the total compatibility with IFS-ARPEGE   would be lost - fine-scale / global oriented physics - new code modularity - the operational cost would become an issue   for most of the partners History of the ALADIN-AROME link

Slide19: 

The first presentations to the partners December 2001 : LACE Council (Bratislava) still called "Meso" January 2002 : ALADIN Assembly of Partners (Casablanca) project now called "AROME" "It’s time to start thinking about the consequences of this evolution The Assembly of Partners could set up the appropriate framework Météo-France needs the awareness ot its Partners for this long-term challenging target" Some partners were surprised by the ambitious objective History of the ALADIN-AROME link

Slide20: 

The first presentations to the partners Behind the surprise: the jump over the 3-7 km range the "grey zone" where the atmosphere is treated: - no longer as quasi-2D - not exactly as really 3D M-F strategy: no effort put in this specific range because the 2.5-3 km range soon affordable (oper in 2008-2010) Other partners may adopt a different strategy History of the ALADIN-AROME link

Slide21: 

The first presentations to the partners The question for the partners was : When will the ALADIN->AROME transition take place for me ? 2010 ? 2015 ? 2020 ? And potentially : What will happen in the "interim phase" when Météo-France will run AROME and I'll still be running ALADIN ? Météo-France needing to keep 10-km versions (oversea, special assistances, coupling of the "real AROME") But this was not an urgent issue History of the ALADIN-AROME link

Slide22: 

The 2002 maturation Presentations : - April 2002 (LACE Council, Dubrovnik) - Oct. 2002 (Assembly of Partners, Bucharest) At Bucharest a longer discussion than anticipated : some partners fearing the situation where Météo-France would put resources on AROME and neglect the ALADIN evolution And this sentence (in both presentations) " Structured external collaborations will only make sense after the decision-making and feasibility-study phases." History of the ALADIN-AROME link

Slide23: 

The 2002 maturation Feasibility studies have been carried out esp. as regards the dynamics but also assimilation, surface etc. The decision-making : "CIPN" meeting in December 2002 History of the ALADIN-AROME link

Slide24: 

The "CIPN" meeting mid-December 2002 Choice for the AROME dynamics: ALADIN-NH scientifically at the top much more efficient than any other candidate a success for ALADIN Now the basic ingredients for AROME are Dynamics: ALADIN (-NH) Physics: Meso-NH Assimilation: ALADIN But the work is not just to copy/paste ! History of the ALADIN-AROME link

Slide25: 

The "CIPN" meeting mid-December 2002 Choice for the AROME dynamics: ALADIN-NH Remember: when the distributed memory computers appeared some people said that spectral models were: - inadapted to MPP architecture - inadapted to LAM - inadapted to NH They are still there (from ECMWF to AROME…) History of the ALADIN-AROME link

Slide26: 

The "CIPN" meeting mid-December 2002 Offer for the ALADIN partners: convergence instead of (delayed) transition Leads to the concept of a 10-km version of AROME the core of AROME + a convection parametrization + likely adaptations to the 10-km case based on Meso-NH experience to be operationally used also at Météo-France Nota: 9 or 9.325 or 10.7654 km will also be possible… History of the ALADIN-AROME link

Slide27: 

The "CIPN" meeting mid-December 2002 Offer for the ALADIN partners: convergence instead of (delayed) transition The convergence implies different constraints on the scientific work AROME-oriented ALADIN evolutions and vice versa for the 10-km version also very important, on the code evolution the deported versions have to follow History of the ALADIN-AROME link

Slide28: 

The "CIPN" meeting mid-December 2002 With this convergence strategy AROME becomes a toolbox adapted to different computing capabilities Every partner able to jump to the "real" AROME when technically ready (esp. computer) History of the ALADIN-AROME link

Slide29: 

Convergence option discussed during the workshop Objective to continue in a mutually beneficial situation Of course the convergence leads to much closer milestones than the sudden transition When ready, the 10-km version will be immediately usable by almost all the partners A need to have a mixed scientifical/technical/political discussion (hence elsewhere than at the annual Assembly) Now is a good time for that History of the ALADIN-AROME link

Slide30: 

Discussion

Slide31: 

Again an issue due to the cost of AROME ALADIN is relatively cheap the same model on supercomputers and workstations Because the resources have been put on large-scale 4D-Var With AROME becomes expensive again : - the 2.5 - 3 km resolution without compromise - the detailed physics (esp. 3D turb., microphysics) The reason for addressing this point… About computing resources

Slide32: 

Supercomputers at Météo-France (since begin. of ALADIN) Step by step: 1988: 25% of a Cray-2 1992: a full Cray-2 x 4 1993: Cray C90/4 x 3 1995: Cray C90/8 x 2 1998: Fujitsu VPP700/26 x 5 2000: Fujitsu VPP5000/31 x 4 2003: Fujitsu VPP5000/124 x 4 2006: ? ? ? ? ? ? ? x 3 About computing resources

Slide33: 

Supercomputers at Météo-France (since begin. of ALADIN) With respect to the beginning of the ALADIN story: 1988: 25% of a Cray-2 1 1992: a full Cray-2 4 1993: Cray C90/4 12 1995: Cray C90/8 24 1998: Fujitsu VPP700/26 120 2000: Fujitsu VPP5000/31 480 2003: Fujitsu VPP5000/124 1920 2006: ? ? ? ? ? ? ? 5760 About computing resources

Slide34: 

Supercomputers at Météo-France (since begin. of ALADIN) How does this compare with Moore's law ? (x 2 each 18 months) 1988-2006: 1 to almost 6000 x 2 each 17.3 months ! Moore's law would lead to 4096 About computing resources

Slide35: 

Supercomputers at Météo-France (since begin. of ALADIN) Evolution of the finest operational model resolution: 1985: 36 km (PERIDOT, grid point) 1993: 30 km (ARPEGE) 1995: 25 km (ARPEGE) 1998: 10 km (ALADIN/France) (+ALADIN 15km quasi-oper PECO 1994, oper LACE 1996) Same ratio PERIDOT/ALADIN and ALADIN/AROME PERIDOT and AROME at full CPU consumption not the case for ALADIN/France About computing resources

Slide36: 

Supercomputers at Météo-France (since begin. of ALADIN) Current trend of computing evolution: top supercomputers: slower than Moore's law less customers (but still the Met Services ! ) other computers: not really slowing down More an issue for M-F than for the partners About computing resources

Slide37: 

(Super)-computers for NWP at the partners In general a more recent history Situation in 2003: highly variable about 1/8 to 1/150 of M-F current Fujitsu 1/30 to 1/600 of the mid-2003 conf Delay / M-F: 7 to 13 years equivalent computer evolution About computing resources

Slide38: 

Computing requirements for AROME The mesh / time step ratio work on numerical schemes: from the beginning of semi-implicit (≈1980) until now: a factor of 14 gained almost 6-year equivalent computer evolution but the constraint will come more from the physics About computing resources

Slide39: 

Computing requirements for AROME Let's consider the following configuration 10km domain: as current 10km ALADIN ex. for France: 2500 km x 2500 km 2.5-3 km domain: smaller, just adapted to the country ex. for France: 1500 km x 1500 km About computing resources

Slide40: 

Computing requirements for AROME This is a typical starting configuration Orders of magnitude: - 10 km AROME: 3 x more / 10-km ALADIN - 2.5-3 km AROME: 50-100 x more / 10km AROME Converted in computer evolution accord. to Moore's law - 10 km AROME: 2.5 years / 10-km ALADIN - 2.5-3 km AROME: 8.5-10 years / 10km AROME About computing resources

Slide41: 

Computing requirements for AROME So all together the equivalent of about 12 years of evolution How can M-F imagine AROME being oper in 2008-2010 ? Because ALADIN-France doesn't saturate the Fujitsu run on 2 procs / 31 of the 2000 machine factor of 15 ≈ 6-year evolution the countdown starts in 1994 The same may apply (but to less extent) to other partners About computing resources

Slide42: 

Computing requirements for AROME Now looking at the other partners CPU evolution more difficult to assess depending on economical progress, european integration etc. still we'll try to give estimates taking into account the CPU ratio / France the "minimum AROME-area ratio" / France About computing resources

Slide43: 

Computing requirements for AROME Minimum AROME Area About computing resources

Slide44: 

Computing requirements for AROME Minimum AROME-area ratio / France MA : 1 - 2 ≈ 0     -1.5 year PL, RO, TU : 2 - 4 ≈ -1.5  -3 years PT, HR : 4 - 6 ≈ -3     -4 years AU, BG, HU, CZ : 6 - 8 ≈ -4     -4.5 years SK, MO, BE : 8 - 10 ≈ -4.5  -5 years SI : 10 - 15 ≈ -5     -6 years In other words, most of the partners:      - will wait for AROME 3-6 years less due to the size      - have an AROME requirement close to a VPP5000/31 About computing resources

Slide45: 

Computing requirements for AROME Coupling both aspects CPU evolution and domain size: Assuming the Moore's law continuation for the partners (hopefully too conservative for the currently less powerful) Assuming AROME at M-F in 2008 on the 2006 "x3-machine" (for the partners having invested) an AROME-delay of about: - 1-2 years for the best situations - 8-10 years for the worst situations (hence hoping that this could be reduced) - in-between for the others with 2-year uncertainty (computer replacement strategies) About computing resources

Slide46: 

Computing requirements for AROME We have to take this diversity into account Without forgetting the "first step" aspect of this "minimum domain" approach (which has some scientific limitations) About computing resources

Slide47: 

Computing requirements for AROME By the way, there was also a delay for ALADIN: - quasi-oper 1994 - most of real operations starting from 1996 to 2000 even if less directly CPU-linked About computing resources

Slide48: 

Computing requirements for AROME The same causes may lead to the same solutions more centralization at least in a first phase ? Not to be decided today ! About computing resources

Slide49: 

Discussion