|
The Software
Engineering Working Group (SEWG) met during the 9th Annual CCSM Workshop on 8
July 2004 from 1:00 p.m. to 5:00 p.m. The following is a summary of that
meeting.
Douglass Post, Los Alamos National Laboratory, opened the meeting with a
presentation on the DARPA High Productivity Computer Systems (HPCS) Initiative.
He described his role as application analyst for HPCS and discussed
characteristics and typical lifecycles of successful projects. His presentation
also included a discussion of the "coming crisis" in high performance computing,
due to difficulties in validation of extremely complex models.
John Drake described recent and ongoing activities in the SciDAC project:
Collaborative Design and Development of the Community Climate System Model for
Terascale Computing.
After an overview of participants, science goals, and project goals, he
discussed progress and plans for each of the CCSM component models, as well as
general CCSM software engineering issues, such as vectorization, performance
portability, parallel I/O, and single
executable CCSM.
Activities for each component model were divided into software engineering, in
particular optimization and performance portability, model development,
including both new algorithms and new physical processes, and science
applications. Details follow.
Coupler: MCT (Model Coupling Toolkit) and MPH3 (Multi-Processor handshaking)
were used to develop CPL6, the version of the coupler in CCSM3. Ongoing
development includes vectorization, support for single executable CCSM (e.g.,
resolving name conflicts), and support for atmospheric and ocean (biogeo)
chemistry.
Atmospheric Model: Vectorization, load balancing, and communication
optimizations were developed, shown to improve performance, introduced into
CAM3, and released as part of CCSM3. Gas-phase atmospheric chemistry with
emissions, deposition, transport, and photochemical reactions for 89 species
were developed and the impact of different levels of approximations evaluated.
Subgrid orography was shown to produce realistic precipitation, snowcover,
runoff, and river discharge, without requiring high-resolution dynamics.
Ongoing tasks include additional optimization on the vector systems,
optimizations for high resolution, continued development of a version of the
model with subgrid orography, and continued development of a chemical atmosphere
model. Planned new model development activities include renewed development of
the Semi-Lagrangian spectral dynamics. Software engineering plans include
blocked/ESMF-compliant dynamics, utility layer standardization, and the
introduction of parallel I/O options developed as part of ZIOLIB. Planned
science applications include using subgrid orography with IPCC simulations and
high-resolution downscaling of selected IPCC runs.
Land Model: Land model was completely rewritten to add vectorization.
Decomposition algorithm was completely modified to significantly improve load
balancing. River runoff history output was modified so that it appeared on the
RTM grid rather than the model grid. Ongoing tasks include incorporation of new
biogeochemistry, (CLM3-CN and CLM3-CASA'), an urban model, and dynamic land use.
Ocean Model: Vector modifications were introduced into the CCSM3 ocean model.
Studies showed that high-resolution runs were able to resolve eddies visible
in
observations. Other studies examined (a) the impact of iron enrichment, (b)
chlorophyll distribution during La Ninas and El Ninos, and (c) simulating global
flux of DMS from the ocean to the atmosphere, all using POP. HYPOP development
continues, with recent progress on the vertical coordinate system and a new time
stepping algorithm. Work also continues on the LANL ecosystem model.
Sea Ice Model: The new incremental remapping scheme proved to be three times
faster than MPDATA, resulting in a total model speedup of about 30 percent, and
was added to CSIM, the ice model in CCSM3. Vectorization modifications were
first introduced into CICE3.0, and then ported into CCSM3/CSIM. An Arctic Ocean
Model Intercomparison Project (AOMIP) run, using global, 0.4 deg, coupled POP2.0
and CICE3.1_beta on the Cray X1, finished 1948-2002. Ongoing activities include
sensitivity analysis and parameter tuning test of the CICE code using Automatic
Differentiation (AD)-generated derivative code.
Cecelia Deluca described ongoing progress with the ESMF project, including the
release of ESMF Version 2.0 on 23 June 2004. The release includes software for
representing and manipulating modeling components, states, bundles of fields,
fields, grids, and arrays, as well as utilities for time management,
configuration, and logging. Major limitations of the current distribution are
that it only supports logically rectangular grids; only supports sequentially
executing components; and does not yet have framework-managed threading enabled.
She also described interoperability experiments recently completed, which
demonstrate how the framework can be used to create prototype applications out
of never-before coupled model components. Prototype applications include CAM/NCEP
SSI, CAM/MITgcm, and GFDL B-grid atmosphere/MITgcm.
Tom Henderson discussed plans for CCSM model unification. The objective of this
effort is to bring together the stand-alone CAM and the full CCSM so that only a
single CCSM model needs to be supported. This would greatly reduce the overhead
necessary to maintain two separate and similar codes.
The meeting wrapped up with an open discussion of an evaluation plan for using
ESMF in the CCSM. Some of the issues that came up were MPMD versus SPMD, how to
handle name conflicts in SPMD, and what the appropriate code and strategy for
evaluation would be.
|
