Earth System Modeling Infrastructure Cecelia DeLuca/ESMF-NCAR March 31-April 1, 2009 CHyMP Meeting www.esmf.ucar.edu Outline • • • Elements of interoperability platforms Integrating across elements Summary www.esmf.ucar.edu.
Download ReportTranscript Earth System Modeling Infrastructure Cecelia DeLuca/ESMF-NCAR March 31-April 1, 2009 CHyMP Meeting www.esmf.ucar.edu Outline • • • Elements of interoperability platforms Integrating across elements Summary www.esmf.ucar.edu.
Earth System Modeling Infrastructure Cecelia DeLuca/ESMF-NCAR March 31-April 1, 2009 CHyMP Meeting www.esmf.ucar.edu Outline • • • Elements of interoperability platforms Integrating across elements Summary www.esmf.ucar.edu Elements of interoperability platforms 1. 2. 3. 4. 5. Tight coupling tools and interfaces - hierarchical and peer component relationships - frequent, high volume transfers on high performance computers Loose coupling tools and interfaces - generally peer-peer component relationships - lower volume and infrequent transfers on desktop and distributed systems Science gateways - browse, search, and distribution of model components, models, and datasets - visualization and analysis services - workspaces and management tools for collaboration Metadata conventions and ontologies - ideally, with automated production of metadata from models Governance - coordinated and controlled evolution of systems www.esmf.ucar.edu Tight coupling tools and interfaces Examples: • Earth System Modeling Framework (ESMF) - NASA, NOAA, Department of Defense, community weather and climate models, U.S. operational numerical weather prediction centers (HPC focus) http://www.esmf.ucar.edu • Flexible Modeling System (FMS) – NOAA precursor to ESMF, still used at the Geophysical Fluid Dynamics Laboratory for climate modeling http://www.gfdl.noaa.gov/fms/ • Space Weather Modeling Framework (SWMF) – NASA-funded, used at the University of Michigan for space weather prediction www.esmf.ucar.edu How coupling tools work: • Users wrap their native data in framework data structures • Users adopt standard calling interfaces for a set of methods that enable data exchange between components • Development toolkits help users with routine functions (regridding, time management, etc.) www.esmf.ucar.edu ESMF: Standard interfaces • Three ESMF component methods: Initialize, Run, and Finalize (I/R/F) • Each can have multiple phases • Users register their native I/R/F methods with an ESMF Component • Small set of arguments: AppDriver (“Main”) Call Initialize Call Run Initialize Call Finalize Run Finalize Parent GridComp “Hurricane Model” Call Initialize Call Run Initialize Call Finalize Run Finalize Child GridComp “Atmosphere” Initialize Run Finalize Child GridComp “Ocean” Initialize Run Finalize Child CplComp “Atm-Ocean Coupler” call ESMF_GridCompRun (myComp, importState, exportState, clock, phase, blockingFlag, rc) www.esmf.ucar.edu ESMF: Distributed data representation 1. Representation in index space (Arrays) • • • Simple, flexible multi-dimensional array structure Regridding via sparse matrix multiply with user-supplied interpolation weights Scalable to 10K+ processors - no global information held locally Supported Array distributions 2. Representation in physical space (Fields) • Built on Arrays + some form of Grid • Grids are: logically rectangular, unstructured mesh, or observational data streams • Regridding via parallel on-line interpolation weight generation, bilinear or higher order options • Intrinsically holds significant amounts of metadata dynamic, usable for multiple purposes, limited annotation required www.esmf.ucar.edu ESMF: Coupling options • • • Generally single executable for simpler deployment Push mode of data communication is very efficient Coupling communications can be set up and called in a coupler, or called directly from within components (for I/O, data assimilation) • Hierarchical components for organization into subprocesses Recursive components for nesting higher resolution regions Coupling across C/C++ and Fortran Ensemble management • • • www.esmf.ucar.edu ESMF-based hierarchical structure of GEOS-5 atmospheric GCM ESMF: Performance portability • • • • ESMF is highly performance portable, low (<5%) overhead 3000+ regression tests run on 30+ platform/compiler combinations nightly See http://www.esmf.ucar.edu/download/platforms Newer ports include native Windows, Solaris Using TeraGrid Build and Test Service to simplify regression testing Performance at the petascale… ASMM Run Time Comparison on XT4 (w/o -N1) ASMM Run-Time Comparison 1000 3D Array 3D Array (-N1) 100 Bundle (-N1) msec 10 0 Number of Processors www.esmf.ucar.edu 16 38 4 81 92 40 96 20 48 10 24 51 2 25 6 12 8 64 32 16 8 1 4 Plot from Peggy Li, NASA/JPL Tested on ORNL XT4, -N1 means 1 core per node. Bundle Time (msecs) Scaling of the ESMF sparse matrix multiply, used in regridding transformations, out to 16K processors. (ESMF v3.1.0rp2) ESMF: Higher order interpolation techniques in CCSM Interpolation noise in the derivative of the zonal wind stress Interp. noise grid index in latitudinal direction Black = bilinear Red = higher-order ESMF v3.1.1 Green = higher order ESMF v4.0.0 www.esmf.ucar.edu • ESMF higher order interpolation weights were used to map from a 2-degree Community Atmospheric Model (CAM) grid to a POP ocean grid (384x320, irregularly spaced) • 33% reduction in noise globally in quantity critical for ocean circulation compared to previous bilinear interpolation approach • ESMF weights are now the CCSM default HAF ESMF: Model map SWMF Legend GAIM Ovals show ESMF components and models that are at the working prototype level or beyond. CCSM4 Dead atm GFS Data atm GFS Atm Phys POP Ocean Dead ocean Data ocean Stub ocean CICE ice Dead ice Data ice Stub ice CLM Dead land Data land Stub land GFS Atm Dynamics NOAA Department of Defense University NASA Department of Energy National Science Foundation GFS I/O FIM NEMS NMM-B Atm Dynamics NMM-B Atm Phys Ice sheet ESMF coupling complete ESMF coupling in progress NMM History Strat Chem Component (thin lines) Model (thick lines) GEOS-5 Param Chem GEOS-5 Atm Dynamics GOCART GEOS-5 GWD GEOS-5 FV Dycore FV Cub Sph Dycore Tracer Advection Land Info System GSI GEOS-5 Atm Physics GEOS-5 Hiistory GEOS-5 Atm Chem GEOS-5 Radiation GEOS-5 Aeros Chem GEOS-5 LW Rad GEOS-5 Solar Rad GEOS-5 Surface GEOS-5 Land GEOS-5 Topology GEOS-5 Veg Dyn GEOS-5 Lake GEOS-5 Catchment GEOS-5 Moist Proc GEOS-5 Turbulence GEOS-5 Land Ice GEOS-5 OGCM Poseidon GEOS-5 Data Ocean GEOS-5 Salt Water GEOS-5 Ocean Biogeo WRF POP ROMS MOM4 UCLA AGCM COAMPS MITgcm MITgcm Atm MITgcm Ocean pWASH123 NCOM www.esmf.ucar.edu SWAN CICE HYCOM ADCIRC Loose coupling tools and interfaces Examples: • OpenMI http://www.openmi.org • Web service approaches Coupling options: • Generally multiple executable • Pull mode of data communication simple but not efficient (ask for a data point based on coordinates) • Generally peer-peer component relationships • Coupling across multiple computer languages (Python, Java, C++, etc.) www.esmf.ucar.edu Science gateways – access centers Examples: • Earth System Grid (ESG) – DOE, NCAR, NOAA support, used to distribute Intergovernmental Panel on Climate Change data and for climate research http://www.earthsystemgrid.org • Hydrologic Information System (HIS) - NSF funded, used to enhance access to data for hydrologic analysis http://his.cuahsi.org • Object Modeling System (OMS) - USDA effort, used for agricultural modeling and analysis http://javaforge.com/project/1781 www.esmf.ucar.edu Metadata conventions and ontologies Examples: • Climate and Forecast (CF) conventions - spatial and temporal properties of fields used in weather and climate http://cf-pcmdi.llnl.gov • METAFOR Common Information Model (CIM) – large EU-funded project, climate model component structure and properties (including technical and scientific properties) http://metaforclimate.eu • WaterML – Schema for hydrologic data developed by the Consortium of Universities for the Advancement of Hydrologic Science (CUAHSI) http://his.cuahsi.org/wofws.html www.esmf.ucar.edu Governance Pervasive issue in community modeling Divergent effects of • Multiple institutions • Geographic dispersion • Multiple domains of interest (working groups) Must be balanced by strong integration body - strategies: • Meets frequently enough to affect routine development (quarterly) • Meets virtually to get sufficient representation • Includes user and other stakeholder representatives • Authorized to prioritize and set development schedule • Supported by web-based management tools www.esmf.ucar.edu Integrating across interoperability elements Examples from the Curator project (NSF and NASA) • Automated output of CF and CIM XML schema from ESMF (tight coupling + ontology) • Ingest of ESMF-generated schema into ESG, propagation into tools for search, browse, inter-comparison and distribution of model components and models (tight coupling + ontology + science gateway) • Implementation of dataset “trackback” in ESG that connects datasets with detailed information about the models used to create the data (tight coupling + ontology + science gateway) • Implementation of personal and group workspaces in ESG (science gateway + governance) www.esmf.ucar.edu Integrating across interoperability elements (cont.) • Translation of ESMF interfaces into web services to enable invocation of ESMF applications from a science gateway, and enable data and metadata from the run to be stored back to the gateway (tight coupling + loose coupling + science gateway + ontology, new TeraGrid funding) Web service interface ESMF interface Tightly coupled HPC components www.esmf.ucar.edu Loosely coupled components Issue of switch from push to pull data interactions… Screenshot: Component trackback www.esmf.ucar.edu Screenshot: Faceted search www.esmf.ucar.edu Summary • Cross-domain interoperability platforms have multiple elements • Many of these elements already exist • Integration activities (such as Earth System Curator) are the next focus www.esmf.ucar.edu Image courtesy of Rocky Dunlap, Georgia Institute of technology