Climate simulations - T
Transcription
Climate simulations - T
DLR.de • Chart 1 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 A diagnostic interface for ICON – Coping with the challenges of high-resolution climate simulations Bastian Kern Deutsches Zentrum für Luft- und Raumfahrt Institut für Physik der Atmosphäre Oberpfaffenhofen DLR.de • Chart 2 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Weather – Climate? Screenshot, 26.04.2016: http://www.wetter.de/klima/europa/deutschland/ muenchen-s99000036/april.html DLR.de • Chart 3 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Weather – Climate? Screenshot, 26.04.2016: http://www.wetter.de/klima/europa/deutschland/ muenchen-s99000036/april.html DLR.de • Chart 4 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Weather – Climate? Screenshot, 26.04.2016: http://www.wetter.de/klima/europa/deutschland/ muenchen-s99000036/april.html DLR.de • Chart 5 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Overview Numerical models numerical weather prediction climate simulation Challenges (and possible answers) Interface for on-line diagnostics implementation, performance Outlook DLR.de • Chart 6 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast 1904: “Das Problem der Wettervorhersage, betrachtet vom Standpunkte der Mechanik und der Physik”, Met. Z. Necessary and sufficient conditions for solving the forecast problem: • Know with sufficient accuracy the (initial) state of the atmosphere • Know with sufficient accuracy the (physical) laws determining the evolution of the atmosphere One diagnostic and one prognostic step Suggests a graphical calculation method. Gaining higher accuracy using smaller fixed time intervals, while higher calculation costs. http://www.nb.no/ (blds_01075) DLR.de • Chart 7 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast 1922: “Weather Prediction by Numerical Process” Calculation of initial pressure tendency from full partial differential equations. Took two years by hand to calculate for one point (in the centre of the calculation region). “Forecast”: p = 145 hPa in 6h Lewis Fry Richardson by Walter Stoneman, NPG x164373 © National Portrait Gallery, London, CC BY-NC-ND 3.0 DLR.de • Chart 8 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast 1922: “Weather Prediction by Numerical Process” Calculation of initial pressure tendency from full partial differential equations. Took two years by hand to calculate for one point (in the centre of the calculation region). “Forecast”: p = 145 hPa in 6h Level LFR Model filtered 1 48.3 48.5 -0.2 2 77.0 76.7 -2.6 3 103.2 102.1 -3.0 4 126.5 124.5 -3.1 Surface 145.1 145.4 -0.9 P. Lynch, Journal of Computational Physics, 2008, Table 1 DLR.de • Chart 9 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast 1922: “Weather Prediction by Numerical Process” Calculation of initial pressure tendency from full partial differential equations. Took two years by hand to calculate for one point (in the centre of the calculation region). “Forecast”: p = 145 hPa in 6h “Forecast Factory”: 64000 human computers calculating weather forecasts in a big “theatre” first massive parallel computer system Lewis Fry Richardson by Walter Stoneman, NPG x164373 © National Portrait Gallery, London, CC BY-NC-ND 3.0 DLR.de • Chart 10 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast 1946 proposal: “Weather prediction was, par excellence, a scientific problem suitable for solution using a large computer.” John von Neumann LANL (http://www.lanl.gov/) DLR.de • Chart 11 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast Courant, Friedrichs, Lewy, Math. Annalen, 1928 John von Neumann LANL (http://www.lanl.gov/) DLR.de • Chart 12 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast 1946 proposal: “Weather prediction was, par excellence, a scientific problem suitable for solution using a large computer.” Charney leader of meteorological group • unfiltered primitive equations Richardson • baroclinic quasi-geostrophic system computationally too demanding • barotropic vorticity equation satisfactory initial results 1950: first weather forecast on ENIAC. 5 scientist plus 3 operators 5 weeks 2 satisfactory “forecasts” one 24h “forecast” took 24h John von Neumann LANL (http://www.lanl.gov/) DLR.de • Chart 13 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast ENIAC (U.S. Army Photo) John von Neumann LANL (http://www.lanl.gov/) DLR.de • Chart 14 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast MIT Museum DLR.de • Chart 15 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 The first weather forecast MIT Museum DLR.de • Chart 16 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 http://www.nco.ncep.noaa.gov/sib/verification DLR.de • Chart 17 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Evolution in weather prediction http://www.ecmwf.int/en/forecasts/charts/medium/ anomaly-correlation-ecmwf-500hpa-height-forecasts DLR.de • Chart 18 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Evolution in computing power “PHONIAC” ENIAC integrations as Java application on a Nokia 6300 ENIAC: 24h PONIAC: <1s Lynch & Lynch, Weather, 2008 CRAY-1: ~250 MFLOPS (vector) Nokia 6300: ~237 MFLOPS ENIAC: 140 kW CRAY-1: 115 kW Nokia 6300: 1.5 W DLR.de • Chart 19 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Challenges in weather prediction Atmosphere as chaotic system Slight deviations in initial conditions can cause large differences in forecast https://commons.wikimedia.org/wiki/ User:Dschwen, CC BY 2.5 DLR.de • Chart 20 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Challenges in weather prediction Atmosphere as chaotic system Slight deviations in initial conditions can cause large differences in forecast Sophisticated methods for initial conditions Ensemble forecasts http://www.daserste.de/wetter (26.04.2016) DLR.de • Chart 21 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Weather forecast Climate projection? Weather forecast is an initial value problem Deterministic chaotic system boundary for meaningful forecast about 14 days Climate models are essentially numerical weather prediction models (some changes) DLR.de • Chart 22 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Weather forecast Climate projection? Weather forecast is an initial value problem Deterministic chaotic system boundary for meaningful forecast about 14 days Climate models are essentially numerical weather prediction models (some changes) How can we “forecast” climate in 2100? DLR.de • Chart 23 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Weather forecast Climate projection? Weather forecast is an initial value problem Deterministic chaotic system boundary for meaningful forecast about 14 days Climate models are essentially numerical weather prediction models (some changes) How can we “forecast” climate in 2100? We can do climate projections statistical analyses Climate projection is boundary value problem Model produces realistic weather systems from internal variability DLR.de • Chart 24 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Weather forecast Climate projection? Numerical Weather Prediction Climate projection http://www.daserste.de/wetter (26.04.2016) http://www.ipcc.ch/report/ar5/syr/ DLR.de • Chart 25 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations https://www.dkrz.de/about/media/galerie/Media-DKRZ/hlre-3 DLR.de • Chart 26 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations „Mistral - Phase 2“ (spring 2016) ~ 68.000 cores 3 PetaFLOPS memory: ~200 Terabyte disks: 50 Petabyte (Lustre) Top 500, November 2015 https://www.dkrz.de/about/media/galerie/Media-DKRZ/hlre-3 http://www.top500.org/list/2015/11/ DLR.de • Chart 27 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations – Challenges “With the HLRE-II system "Blizzard", about 8 Petabytes of new simulation results had to be archived per year. Starting with the availability of the new HLRE-3 "Mistral" in summer 2015, we expect the yearly data growth to increase to up to 75 Petabytes per year.” (www.dkrz.de) https://www.dkrz.de/about/media/galerie/Media-DKRZ/hlre-3 DLR.de • Chart 28 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations – Challenges DLR.de • Chart 29 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations – Challenges DLR.de • Chart 30 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations – Challenges 1) simulation 2) data output memory 3) postprocessing (subset) 3b) postprocessing (reduced) 3a) postprocessing disk (scratch/work) archive (tape) 4) archiving DLR.de • Chart 31 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations – Challenges 1) simulation 2) data output memory 5) postprocessing (subset) 5b) postprocessing (reduced) 5a) postprocessing 3) archiving disk (scratch/work) 4) get data (subset) archive (tape) 6) archiving (reduced) DLR.de • Chart 32 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations – Challenges 1) simulation 2) data output memory 5) postprocessing (subset) 3) archiving disk (scratch/work) 5b) postprocessing (reduced) 5a) postprocessing Bottleneck 1: ~400 GB/s 4) get data (subset) archive (tape) 6) archiving Bottleneck 2: ~20 GB/s DLR.de • Chart 33 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Climate simulations – Challenges 1) simulation 2) data output (reduced) memory 1a) onlinediagnostics 3) archiving (reduced) disk (scratch/work) archive (tape) DLR.de • Chart 34 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 ICON – ICOsahedral Non-hydrostatic modelling framework Joint development of Deutscher Wetterdienst (DWD) and Max Planck Institut für Meteorologie (MPI-M) for numerical weather prediction and climate applications Important development goals • Improved conservation properties compared to GME and ECHAM • Non-hydrostatic dynamical core for application on all scales • Variable grid resolution (two-way-nesting) • Application as global and regional model • Improved performance on parallel computers (compared to GME / ECHAM) DLR.de • Chart 35 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 ICON – Grid structure G. Zängl, DWD DLR.de • Chart 36 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 ICON – Grid structure and refinement G. Zängl, DWD DLR.de • Chart 37 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 HD(CP)² - High Definition Clouds and Precipitation for advancing Climate Prediction HD(CP)² project: • High requirements to hardware High spatial resolution (Limited Area at 100m ~ 56 M grid cells + outer nesting domains) Small integration time-step • Scalability on O(105) cores Communication and memory handling ICON: • Unstructured grid and parallel decomposition • Nesting of multiple domains Figure: https://code.zmaw.de/projects/icon DLR.de • Chart 38 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 HD(CP)² - Model Diagnostics (subproject) • HD(CP)² Phase I milestones: • Diagnostic Interface in ICON • Prototype implementation of advanced diagnostic tool • Implementation in ICON: • Modular Earth Submodel System (MESSy) • #IFDEF MESSY … DLR.de • Chart 39 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 MESSy – Modular Earth Submodel System P. Jöckel DLR.de • Chart 40 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 HD(CP)² - Model Diagnostics (subproject) • HD(CP)² Phase I milestones: • Diagnostic Interface in ICON • Prototype implementation of advanced diagnostic tool • Implementation in ICON: • Modular Earth Submodel System (MESSy) • #IFDEF MESSY … • Prototype advanced diagnostic tool GRAGG (Grid AGGregation) • “Aggregation on user defined regular grids” examples: spatial mean, joint-PDFs DLR.de • Chart 41 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 GRid AGGregation (GRAGG) • Init-Phase: Grid cell indices CPU1: 10,11,12,17,19 CPU2: CPU3: 17, 23 CPU4: 1,2,3,4,12,13,14 CPU5: • Integration time-step: Partial results per CPU • Output time-step: (MPI-) All-Reduce • Problems: • Min/max in integration time-step Inter-CPU communication • “Hybrid” parallelisation (MPI + OpenMP) 17 23 11 12 19 1 4 3 2 12 13 14 10 17 20 5 15 6 16 DLR.de • Chart 42 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 GRid AGGregation (GRAGG) GRAGG ICON 3D Spatial mean on user-defined grid Diff. DLR.de • Chart 43 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 GRid AGGregation (GRAGG) GRAGG ICON 3D jointPDF in highlighted grid cell Diff. DLR.de • Chart 44 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 HD(CP)² - Model Diagnostics (subproject) • HD(CP)² Phase I milestones: • Diagnostic Interface in ICON • Prototype implementation of advanced diagnostic tool • Implementation in ICON: • Modular Earth Submodel System (MESSy) • #IFDEF MESSY … • Prototype advanced diagnostic tool GRAGG (Grid AGGregation) • “Aggregation on user defined regular grids” examples: spatial mean, joint-PDFs • VIsual Satellite OPerator (VISOP), developed at LMU example DLR.de • Chart 45 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 VIsual Satellite OPerator (VISOP) • Leonhard Scheck et al., Jour. Quant. Spect. and Rad. Transfer, 2016 800nm MODIS simulation, “inner nest” @ 312m, 26.04.2013, 12 UTC approx. 3-D calculated from VISOP data (2D fields) 3-D calculated from ICON output (3D fields) DLR.de • Chart 46 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Performance test on “Mistral” ICON, 3-domains, std. ICON output, 15 min. diagnostic submodel output std. optimised • Diagnostic Interface: no overhead • VISOP (optimised): max. +2.72% (column based) • GRAGG (optimised): max. +60.79% (high MPI communication) Kern & Jöckel, GMDD, in prep. DLR.de • Chart 47 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Outlook • Further optimisation of prototype diagnostic tool • Memory footprint • MPI communication • Parallel output of diagnostic interface DLR.de • Chart 48 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Outlook • Further optimisation of prototype diagnostic tool • Memory footprint • MPI communication • Parallel output of diagnostic interface • HD(CP)² Phase II (1 April 2016 – 30 March 2019) • Advanced on-line diagnostics • On-line feature identification, feature tracking, and combination with trajectories DLR.de • Chart 49 > Challenges of high-resolution climate simulations > Bastian Kern • HPCN Workshop, Göttingen > 10.05.2016 Summary • First weather forecast • Numerical weather prediction and climate simulations • Challenges of high resolution simulations bottlenecks • ICON as new tool for NWP and climate applications • Diagnostic interface based on MESSy • Examples of diagnostic tools • Performance • Outlook