NSF NCAR | NASA GSFC | DOE LANL ANL | NOAA NCEP GFDL | MIT | U MICH Architecture of the Earth System Modeling Framework GMAO Seasonal

NSF NCAR | NASA GSFC | DOE LANL ANL | NOAA NCEP GFDL | MIT | U MICH

www.esmf.ucar.edu

Architecture of the Earth System Modeling Framework

GMAO Seasonal Forecast

NCAR/LANL CCSM

NCEP Forecast

GFDL FMS Suite

MITgcm

NASA GMAO Analysis

ClimateData Assimilation

WeatherCecelia DeLucaGEM Snowmass, CO


www.esmf.ucar.edu

Outline

• Background and Motivation• Applications• Architecture• Implementation• Status• Future Plans• Conclusions


www.esmf.ucar.edu

Motivation for ESMFIn climate research and NWP...

increased emphasis on detailed representation of individual physical processes; requires many teams of specialists to contribute components to an overall modeling system

In computing technology... increase in hardware and software complexity in high-performance computing, shift toward the use of scalable computing architectures

In software …development of frameworks, such as the GFDL Flexible Modeling System (FMS) and Goddard Earth Modeling System (GEMS) that encourage software reuse and interoperability

The ESMF is a focused community effort to tame the complexity of models and

the computing environment. It leverages, unifies and extends existing software

frameworks, creating new opportunities for scientific contribution and collaboration.


www.esmf.ucar.edu

ESMF Project DescriptionGOALS: To increase software reuse, interoperability, ease of use and performance

portability in climate, weather, and data assimilation applications

PRODUCTS: • Core framework: Software for coupling geophysical components and utilities for

building components • Applications: Deployment of the ESMF in 15 of the nation’s leading climate and

weather models, assembly of 8 new science-motivated applications

METRICS:

RESOURCES and TIMELINE: $9.8M over 3 years, starting February 2002

Reuse Interoperability Ease of Adoption Performance

15 applications use ESMF component coupling services and 3+ utilities

8 new applications comprised of never-before coupled components

2 codes adopt ESMF with < 2% lines of code changed, or within 120 FTE-hours

No more than 10% overhead in time to solution, no degradation in scaling


www.esmf.ucar.edu

Outline



www.esmf.ucar.edu

Modeling Applications

SOURCE APPLICATION

GFDL FMS B-grid atmosphere at N45L18

FMS spectral atmosphere at T63L18

FMS MOM4 ocean model at 2°x2°xL40

FMS HIM isopycnal C-language ocean model at 1/6°x1/6°L22

MIT MITgcm coupled atmosphere/ocean at 2.8°x2.8°, atmosphere L5, ocean L15

MITgcm regional and global ocean at 15kmL30

GMAO/NSIPP NSIPP atmospheric GCM at 2°x2.5°xL34 coupled with NSIPP ocean GCM at 2/3°x1.25°L20

NCAR/LANL CCSM2 including CAM with Eulerian spectral dynamics and CLM at T42L26 coupled with POP ocean and data ice model at 1°x1°L40


www.esmf.ucar.edu

Data Assimilation Applications

SOURCE APPLICATION

GMAO/DAO PSAS interface layer with 2O0K observations/day

CAM with finite volume dynamics at 2°x2.5°L55, including CLM

NCEP Global atmospheric spectral model at T170L42

SSI analysis system with 250K observations/day, 2 tracers

WRF regional atmospheric model at 22km resolution CONUS forecast 345x569L50

GMAO/NSIPP ODAS with OI analysis system at 1.25°x1.25°L20 resolution with ~10K observations/day

MIT MITgcm 2.8° century / millennium adjoint sensitivity


www.esmf.ucar.edu

ESMF Interoperability DemonstrationsCOUPLED CONFIGURATION NEW SCIENCE ENABLED

GFDL B-grid atm / MITgcm ocn Introduction of global biogeochemistry into seasonal forecasts.

GFDL MOM4 / NCEP forecast New seasonal forecasting system.

NSIPP ocean / LANL CICE Extension of seasonal prediction system to centennial timescales.

NSIPP atm / DAO analysis Assimilated initial state for seasonal prediction system.

DAO analysis / NCEP model Intercomparison of systems for NASA/NOAA joint center for satellite data assimilation.

NCAR fvCAM/ NCEP analysis Intercomparison of systems for NASA/NOAA joint center for satellite data assimilation.

NCAR CAM Eul / MITgcm ocn Component exchange for improved climate prediction capability.

NCEP WRF / GFDL MOM4 Development of hurricane prediction capability.


www.esmf.ucar.edu

Interoperability Experiments Completed

NCAR Community Atmospheric Model (CAM) coupled to NCEP Spectral Statistical Interpolation (SSI) System, both set up as ESMF components

Experiment utilizes same observational stream used operationally at NCEP

3

NCAR Community Atmospheric Model (CAM) coupled to MITgcm ocean

Atmosphere, ocean, and coupler are set up as ESMF components

Uses ESMF regridding tools

1

GFDL B-grid atmosphere coupled to MITgcm ocean

Atmosphere, ocean, and coupler are set up as ESMF components

Uses ESMF regridding tools

2Temperature SSI import

Temperature SSI export

Temperature difference


www.esmf.ucar.edu

Outline



www.esmf.ucar.edu

Characteristics of Weather and Climate Simulation• Mix of global transforms and local communications• Load balancing for diurnal cycle, event (e.g. storm) tracking• Applications typically require 10s of GFLOPS,

100s of PEs – but can go to 10s of TFLOPS, 1000s of PEs• Required Unix/Linux platforms span laptop to

Earth Simulator• Multi-component applications: component

hierarchies, ensembles, and exchanges• Data and grid transformations between

components • Applications may be MPMD/SPMD,

concurrent/sequential, combinations • Parallelization via MPI, OpenMP, shmem, combinations• Large applications (typically 100,000+ lines of source code)

Platforms

assim

sea iceocean

landatm

physics dycore

assim_atm

atmland

Seasonal Forecastcoupler


www.esmf.ucar.edu

ESMF Architecture

1. ESMF provides an environment for assembling geophysical components into applications, with support for ensembles and hierarchies.

2. ESMF provides a toolkit that components use to

i. increase interoperability

ii. improve performance portability

iii. abstract common services

ESMF InfrastructureData Classes: Bundle, Field, Grid, Array

Utility Classes: Clock, LogErr, DELayout, Machine

ESMF SuperstructureAppDriver

Component Classes: GridComp, CplComp, State

User Code


www.esmf.ucar.edu

Hierarchies and Ensembles

assim

sea iceocean

landatm

physics dycore

assim_atm

atmland

Seasonal Forecast

coupler

ESMF encourages applications to be assembled hierarchically and intuitively

Coupling interfaces are standard at each layer

Components can be used in different contexts

assim_atm

Ensemble Forecast

assim_atm assim_atm

ESMF supports ensembles with multiple instances of components running sequentially (and soon, concurrently)


www.esmf.ucar.edu

ESMF Class Structure

DELayoutCommunications

StateData imported or exported

BundleCollection of fields

GridCompLand, ocean, atm, … model

F90

Superstructure

Infrastructure

FieldPhysical field, e.g. pressure

GridLogRect, Unstruct, etc.

Data Communications

C++

RegridComputes interp weights

CplCompXfers between GridComps

UtilitiesMachine, TimeMgr, LogErr, I/O, Config, Base etc.

ArrayHybrid F90/C++ arrays Route

Stores comm paths

DistGridGrid decomposition

PhysGridMath description


www.esmf.ucar.edu

ESMF Data Classes

Model data is contained in a hierarchy of multi-use classes. The user can reference a Fortran array to an Array or Field, or retrieve a Fortran array out of an Array or Field.

• Array – holds a cross-language Fortran / C++ array• Field – holds an Array, an associated Grid, and metadata• Bundle – collection of Fields on the same Grid• State – contains States, Bundles, Fields, and/or Arrays• Component – associated with an Import and Export State


www.esmf.ucar.edu

ESMF DataMap Classes

These classes give the user a systematic way of expressing interleaving and memory layout, also hierarchically (partially implemented)

• ArrayDataMap – relation of array to decomposition and grid, row / column major order, complex type interleave

• FieldDataMap – interleave of vector components • BundleDataMap – interleave of Fields in a Bundle


www.esmf.ucar.edu

ESMF Standard MethodsESMF uses consistent names and behavior throughout the

framework, for example• Create / Destroy – create a new object, e.g. FieldCreate• Set / Get – set or get a value, e.g. ArrayGetDataPtr• Add / Get / Remove – add to, retrieve from, or remove from a

list, e.g. StateAddField• Print – to print debugging info, e.g. BundlePrint• And so on


www.esmf.ucar.edu

Superstructure

• ESMF is a standard component architecture, similar to CCA but designed for the Earth modeling domain and for ease of use with Fortran codes

• Components and States are superstructure classes• All couplers are the same derived type (ESMF_CplComp) and have a standard

set of methods with prescribed interfaces• All component models (atm, ocean, etc.) are the same derived type

(ESMF_GridComp) and have a standard set of methods with prescribed interfaces

• Data is transferred between components using States.• ESMF components can interoperate with CCA components – demonstrated at

SC03


www.esmf.ucar.edu

Becoming an ESMF GridComp• ESMF GridComps have 2 parts: one part user code, one part ESMF code• The ESMF part is a GridComp derived type with standard methods

including Initialize, Run, Finalize• User code must also be divided into Initialize, Run, and Finalize methods –

these can be multi-phase (e.g. Run phase 1, Run phase 2) • User code interfaces must follow a standard form – that means copying or

referencing data to ESMF State structures• Users write a public SetServices method that contains ESMF

SetEntryPoint calls - these associate a user method (“POPinit”) with a framework method (the Initialize call for a GridComp named “POP”)

• Now you’re an ESMF GridComp


www.esmf.ucar.edu

Infrastructure

• Data classes are Bundles, Fields, and Arrays• Tools for expressing interleaved data stuctures• Tools for resource allocation, decomposition, load

balancing• Toolkits for communications, time management,

logging, IO


www.esmf.ucar.edu

Virtual Machine (VM)

• VM handles resource allocation• Elements are Persistent Execution Threads or PETs• PETs reflect the physical computer, and are one-to-one with Posix

threads or MPI processes• Parent Components assign PETs to child Components• PETs will soon have option for computational and latency /

bandwidth weights• The VM communications layer does simpleMPI-like

communications between PETs (alternative communication mechanisms are layered underneath)


www.esmf.ucar.edu

DELayout

• Handles decomposition• Elements are Decomposition Elements, or DEs (decomposition that’s 2

pieces in x by 4 pieces in y is a 2 by 4 DELayout)• DELayout maps DEs to PETs, can have more than one DE per PET (for

cache blocking, user-managed OpenMP threading) • A DELayout can have a simple connectivity or more complex

connectivity, with weights between DEs - users specify dimensions where greater connection speed is needed

• DEs will also have computation weights• Array, Field, and Bundle methods perform inter-DE communications


www.esmf.ucar.edu

ESMF Communications

• Communication methods include Regrid, Redist, Halo, Gather, Scatter, etc.

• Communications methods are implemented at multiple levels, e.g. FieldHalo, ArrayHalo

• Communications hide underlying ability to switch between shared and distributed memory parallelism


www.esmf.ucar.edu

Load Balancing

Three levels of graphs:• Virtual Machine – machine-level PETs will have computational and

connectivity weights• DELayout – DE chunks have connectivity weights, will have

computational weights• Grid – grid cells will have computational and connectivity weights

Intended to support standard load balancing packages (e.g. Parmetis) and user-developed load balancing schemes


www.esmf.ucar.edu

Outline



www.esmf.ucar.edu

Open Development• Open source• Currently ~800 unit tests, ~15 system tests are bundled with the ESMF

distribution, can be run in non-exhaustive or exhaustive modes• Results of nightly tests on many platforms are accessible on a Test and

Validation webpage• Test coverage, lines of code, requirements status are available on a Metrics

webpage• Exhaustive Reference Manual, including design and implementation notes, is

available on a Downloads and Documentation webpage

• Development is designed to allow users clear visibility into the workings and status of the system, to allow users to perform their own diagnostics, and to encourage community ownership


www.esmf.ucar.edu

Port Status

• SGI• IBM • Compaq• Linux (Intel, PGI, NAG, Absoft, Lahey)


www.esmf.ucar.edu

Outline



www.esmf.ucar.edu

ESMF Key Accomplishments• Public delivery of prototype ESMF v1.0 in May 2003• Monthly ESMF internal releases with steadily increasing functionality • Completion of first 3 coupling demonstrations using ESMF in March

2004– NCAR CAM with NCEP SSI– NCAR CAM with MITgcm ocean– GFDL B-grid atmosphere with MITgcm ocean– All codes above running as ESMF components and coupled using

the framework, codes available from Applications link on website– Other codes running as ESMF components: MOM4, GEOS-5– Less than 2% lines of source code change

• Delivered ESMF v2.0 in June 2004• 3rd Community Meeting to be held on 15 July 2004 at NCAR


www.esmf.ucar.edu

ESMF PrioritiesStatus• Components, States, Bundles, Fields mature• On-line parallel regridding (bilinear, 1st order conservative) completed• Other parallel methods, e.g. halo, redist, low-level comms implemented• Comm methods overloaded for r4 and r8• Communications layer with uniform interface to shared / distributed memory,

hooks for load balancingNear-term priorities• Concurrent components – currently ESMF only runs in sequential mode• More optimized grids (tripolar, spectral, cubed sphere) and more regridding

methods (bicubic, 2nd order conservative) from SCRIP• Comms optimization and load balancing capability• IO (based on WRF IO)• Development schedule on-line, see Development link, SourceForge site tasks


www.esmf.ucar.edu

Outline



www.esmf.ucar.edu

Next Steps– Integration with data archives and metadata standardization efforts,

anticipate collaboration with Earth System Grid (ESG) and European infrastructure project PRISM

– Integration with scientific model intercomparison projects (MIPs), anticipate collaboration with the Program for Climate Model Diagnosis and Intercomparison (PCMDI), other community efforts

– Integration with visualization and diagnostic tools for end-to-end modeling support, anticipate collaboration with the Earth Science Portal (ESP)

– ESMF “vision” for the future articulated in multi-agency white paper on the Publications and Talks webpage


www.esmf.ucar.edu

ESMF Multi-Agency Follow-on• 3 ESMF FTEs at NCAR slated to have ongoing funding through

core NCAR funds• NASA commitment to follow-on support, level TBD• DoD and NSF proposals outstanding• Working with other agencies to secure additional funds


www.esmf.ucar.edu

Outline



www.esmf.ucar.edu

ESMF Overall

• Clear, simple hierarchy of data classes• Multi-use objects mean that the same object can carry

information about decomposition, communications, IO, coupling• Tools for multithreading, cache blocking, and load balancing are

being integrated into the architecture• Objects have consistent naming and behavior across the

framework


www.esmf.ucar.edu

The Benefits

• Standard interfaces to modeling components promote increased interoperability between centers, faster movement of modeling components from research to operations

• The ability to construct models hierarchically enables developers to add new modeling components more systematically and easily, facilitates development of complex coupled systems

• Multi-use objects mean that the same data structure can carry information about decomposition, communications, IO, coupling – this makes code smaller and simpler, and therefore less bug-prone and easier to maintain

• Shared utilities encourage efficient code development, higher quality tools, more robust codes


www.esmf.ucar.edu

More Information

ESMF website: http://www.esmf.ucar.edu

AcknowledgementsThe ESMF is sponsored by the NASA Goddard Earth Science Technology Office.

Documents

NSF NCAR | NASA GSFC | DOE LANL ANL | NOAA NCEP GFDL | MIT | U MICH Architecture of the Earth System Modeling Framework GMAO Seasonal