Meteorology and High- Performance Cyberinfrastructure: Applications in the Extreme

Transcription

1 Meteorology and High- Performance Cyberinfrastructure: Applications in the Extreme Kelvin K. Droegemeier School of Meteorology Center for Analysis and Prediction of Storms University of Oklahoma OSCER Symposium Friday, 13 September 2002

2 The Oklahoma Weather Center

3 The Oklahoma Weather Center A unique confederation of 11 Federal, State, and University of Oklahoma entities Employs over 600 professionals largest group of operational, research and academic meteorologists in the nation Over $0.5 billion in infrastructure Yearly research expenditures of $65 M Injects over $120 M annually into OK economy Develops much of the technology used operationally within the NWS

4 School of Meteorology (SoM) Center for Analysis and Prediction of National Severe Storms Laboratory Storms (CAPS) (NSSL) Oklahoma Climatological Survey Storm Prediction Center (SPC) (OCS) National Weather Service Forecast Cooperate Institute for Mesoscale Office Meteorological Studies (CIMMS) NEXRAD Operational Support Environmental Verification and Facility Analysis Center (EVAC) Warning Decision Training Branch International Center for Natural Hazard and Disaster Research (ICNHDR)

5 Major Achievements Pioneered Doppler weather radar technology

6 Major Achievements Created the nation s first high-density State-wide surface observation network the Oklahoma Mesonet

7 Major Achievements Developed the world s first ground-based mobile Doppler radar (Doppler on Wheels); recorded highest wind speeds on Earth (318 mph) during Moore, Oklahoma F5 tornado on 3 May 1999

8 Major Achievements Created the world s first computer-based forecast system designed specifically for intense, local storms now used operationally by American Airlines, FAA, Kennedy Space Center, and others

9 In the Beginning ENIAC

10 ENIAC Versus Today Weighed 30 tons Had 18,000 vacuum tubes, 1,500 relays thousands of resistors, capacitors, inductors Peak speed of 5000 adds/second and 300 multiplies/sec A 1.2 GHz Pentium IV processor is 500,000 times faster than the ENIAC A desktop PC with 1 Gbyte of RAM can store 5 million times as much data as the ENIAC

11 Charney, Fjortoft,, and von Neumann (1950) Numerically integrated one equation at one level in the atmosphere 24 hour forecast

12 The 1950 Grid Prediction Started at the Large Scale 736 km grid spacing

13 And Grid Spacing Has Been Decreasing Ever Since

14 Trends in Computer Model Forecast Skill Increasing Skill

15 Numerical Weather Prediction Make Observations Collect and Process Data Run Forecast Model on Supercomputer Dissemination to End Users Create Products

16 What Do Operational Forecast Models Currently Predict?

17 What Causes the Real Problems?

18 Why the Lack of Detail? This Thunderstorm Falls Through the Cracks

19 Why the Lack of Detail?

20 The Foundational Question in explicitly predict this type of weather? Can computer forecast model technology...

21 Center for Analysis and Prediction of Storms

22 Present NWS Operations CONUS RUC and Eta Models (32 & 40 km) NCEP Central Operations

23 Small-Scale Scale Weather is LOCAL and heterogeneous! Rain and Snow Intense Turbulence Fog Snow and Freezing Rain and Snow Rain Severe Thunderstorms

24 A Dynamic Environment (time = t) 20 km CONUS Ensembles 10 km 3 km 1 km

25 A Dynamic Environment (time = t+6 hours) 20 km CONUS Ensembles 10 km 3 km 10 km 3 km 3 km 3 km

26 Distributed Computing: The Grid 20 km CONUS Ensembles 10 km 3 km 10 km 3 km 3 km 3 km

27 Three Key Ingredients for Storm-Scale Scale Prediction #1 -- A computer forecast system that accurately represents complex physical processes and operates efficiently on parallel computers

28 Advanced Regional Prediction System (ARPS) Lateral boundary conditions from large-scale models Gridded first guess Mobile Mesonet Rawinsondes ACARS CLASS SAO Satellite Profilers ASOS/AWOS Incoming data Oklahoma Mesonet WSR-88D Wideband Data Acquisition & Analysis ARPS Data Analysis System (ADAS) Ingest Quality control Objective analysis Archival ARPS Data Assimilation System (ARPSDAS) Parameter Retrieval and 4DDA Single-Doppler Velocity Retrieval (SDVR) 4-D Vari ational Data Assimilation Variational Vel - ocity Adjustment & Thermodynamic Retrieval Forecast Generation ARPS Numerical Model Multi-scale non-hydrostatic prediction model with comprehensive physics Product Generation and Data Support System ARPSP LT and ARPSV IEW Plots and images Animations Diagnostics and statistics Forecast evaluation

29 The ARPS Cartesian grid-point code (600,000 lines of Fortran-90) developed new beginning in 1992 Uses 2-D 2 D domain decomposition and MPI/Open-MP Emphasis given to portability, scalability, parallelism, adaptability Solves nonlinear PDEs; ; equations per forecast Has has been run on dozens of machines ranging from Laptops to the CM-200 to a 1024 node Cray T3E to Alpha and Linux Clusters

30 100 ARPS Benchmark Timings 19x19x43 3km grid/processor Itanium 733MHZ CAPS Origin 2000 Platinum 1proc/node Seconds Processors Platinum 2proc/node NCSA Origin 2000 PSC ES-45 PSC ES-40 IBM WHII Power3 IBM NHII Power3 SUN SF6800 SUN-Fire 280-R IBM Regatta Power4 P4-1.6Ghz w/p3 Compile SGI Origin SGI Origin Sun Sparc III 900

31 ARPS Benchmark Timings 19x19x43 3km grid/processor PSC ES-40 IBM WHII Power3 IBM NHII Power3 SUN SF6800 Seconds SUN-Fire 280-R IBM Regatta Power4 P4-1.6Ghz w/p3 Compile SGI Origin Processors SGI Origin Sun Sparc III 900

32 Three Key Ingredients for Storm-Scale Scale Prediction #2 -- Sufficient computing power to accommodate high spatial resolution and produce a forecast at least 10x the speed of the weather

33

34 The Need for Teraflops Must fit the prediction model to the observations (data assimilation/retrieval) About times as expensive as the forecast Must use high spatial resolution 1-33 km resolution in sufficiently large domains Must quantify forecast uncertainty (ensembles) May need forecasts to produce an ensemble each forecast cycle Requirements: TFLOPS sustained; ; 0.5 TB memory; 20 TB storage

35 Three Key Ingredients for Storm-Scale Scale Prediction #3 -- Observations of sufficiently high time/space resolution

36 The Challenge of Using Doppler Radar Data in Numerical Weather Prediction The radar observes... one (radial) wind component precipitation intensity real wind We need... 3 wind components temperature humidity pressure water substance (6-10 fields) observed component CAPS was created to explore solutions to this retrieval problem and apply them to storm-scale scale numerical prediction

37 The Impact of Regional/Local Prediction Radar CAPS 12-hour Forecast (32 km) Radar (Tornadoes in Arkansas)

38 The Impact of Regional/Local Prediction CAPS 6-hour Regional Forecast (9 km) Radar Radar (Tornadoes in Arkansas)

39 The Impact of Regional/Local Prediction Radar CAPS 6-hour Local Forecast (3 km)

40 Real Time Forecasts

41 ARPSINTRP 40 proc Data and Process Flow Timeline AMS-99 Southern Plains (9 km) grid 9 h forecast Boundary conditions Initial conditions ARPS 121 proc U.S. (32 km grid) ARPS forecast NIDS2ARPS 4 proc 28 radars NEXRAD (NIDS) T 0:30 T+0:00 T+0:30 T+1:00 T+1:30 Data at NCSA Data from OU NEXRAD Retrieval Satellite ADAS 120 proc SAT2ARPS 1 proc Surface Rawinsonde Profiler MDCRS ARPSPLT 4 proc 9 min/h Raw data Processed data/ Model output Products

42

43 3 May 1999 Oklahoma Tornado Outbreak Copyright 1999 The Daily Oklahoman

44 May 3 Tornado Tracks/Intensity Courtesy Oklahoma City Area National Weather Service Forecast Office

45 NWS 12-Hour Accumulated Precipitation Forecast

46 NWS 12-Hour Accumulated Precipitation Forecast

47 CAPS Numerical Forecasts of the May 3 Tornadic Storms

48 March 28, 2000 Fort Worth Tornado

49 Fort Worth Tornadic Storm on TV Radar

50 NWS 12-hr NWS Forecast (32 km resolution) Valid 6 pm CDT (shading indicates 12-hr accumulated precipitation) No hint of precipitation in north Texas

51 (6 pm) (7 pm) (8 pm) KFWS Radar Observations

52 2 hrs (6 pm) 3 hrs (7 pm) 4 hrs (8 pm) KFWS Radar Observations ARPS 3 km Forecast With Radar Data

53 2 hrs (6 pm) 3 hrs (7 pm) 4 hrs (8 pm) KFWS Radar Observations ARPS 3 km Forecast With Radar Data ARPS 3 km Forecast Without Radar Data

54 How Good are the Forecasts?? Forecast Actual Event D/FW Airport 30 miles A perfectly predicted storm having a position error of 30 miles may be a terrible forecast on the scale of a single airport

55 How Good Are the Forecasts?? This same forecast would, on a larger scale, be viewed as exceptionally accurate

56 Amarillo Pinning Your Forecast on a Single Model Run??? Fort Worth

57 Amarillo Pinning Your Forecast on a Single Model Run??? Fort Worth 3 km ARPS Forecast Valid 00Z Monday, 16 June 1997

58 Storm-Scale Scale Ensembles in Action Amarillo Fort Worth 5-Member, 3 km ARPS Ensemble

59 Storm-Scale Scale Ensembles in Action Amarillo Fort Worth 5-Member, 3 km ARPS Ensemble

60 The Private Sector Enters the NWP Game

61 Raw Data Collection & Dissemination (NWS) Traditional Methodology

62 Traditional Methodology Raw Data Collection Computer & Dissemination (NWS) Forecast Models (NWS)

63 Traditional Methodology Raw Data Collection Computer & Dissemination (NWS) Forecast Models (NWS) Data/Products (NWS)

64 Traditional Methodology Raw Data Collection Computer & Dissemination (NWS) Forecast Models (NWS) Data/Products (NWS) Analysis, Advice Based on NWS Information

65 Traditional Methodology Raw Data Collection Computer & Dissemination (NWS) Forecast Models (NWS) Data/Products (NWS) Decision Makers Analysis, Advice Based on NWS Information

66 Traditional Methodology US National Weather Service Raw Data Collection Computer & Dissemination (NWS) Forecast Models (NWS) Data/Products (NWS) Decision Makers Analysis, Advice Based on NWS Information

67 Traditional Methodology US National Weather Service Raw Data Collection Computer & Dissemination (NWS) Forecast Models (NWS) Data/Products (NWS) Private Sector Decision Makers Analysis, Advice Based on NWS Information

68 The New Landscape Private Sector Raw Data Collection & Dissemination (NWS) Customized Computer Forecast Models/Decision Support Tools Customized Data Decision Makers Input to Customized Risk Models and Tools

69 The New Landscape Private Sector Raw Data Collection & Dissemination (NWS) Customized Computer Forecast Models/Decision Support Tools Customized Data Decision Makers Input to Customized Risk Models and Tools

70 Other Information Technology Activities at

71 NEXRAD Radar Network

72 Ingesting NEXRAD Radar Data via the Internet

73 Mesocyclone Climatologies Medium ITR Grant Courtesy Thomas Jones and Kevin McGrath, University of Oklahoma

74 NEXRAD NEXRAD in-situ surface observations NEXRAD Gap

75 Center for Adaptive Sampling of the Atmosphere (CASA) NSF ERC Proposal Oklahoma Cell Network Concept: Prof. D. McLaughlin, U of Massachusetts

76 FCC Cellular Database -- 20,455 sites

77 Data Handling Requirements Per Pixel Data storage: 10 bytes per observation Assume sensor nodes update each minute. Total Data Volume Regional Network: 25 Gbytes per observation; 26 Tbytes per day Nation-wide Network: 900 Gbytes per observation; 1.3 Pbytes per day

78 Dealing with the Flood of Information Generating it Moving it Analyzing (mining) it Storing it Finding it Accessing it Combining it As scientists and engineers, we spend well over 50% of our time dealing with the logistics of information

79 Linked Environments for Atmospheric Discovery (LEAD) NSF Large ITR Proposal

80 Linked Environments for Atmospheric Discovery (LEAD) NSF Large ITR Proposal