Parallelizing a Global Climate Model.
Next: Adapting Rigorous Electromagnetic
Up: No Title
Previous: Sparse Cholesky.
GATOR stands for Gas-Aerosol-Transport-Radiation-Chemistry Model, and
is a model of the earth's atmospheric chemistry.
It is part of a larger Earth Systems Model of the earth's climate,
including the ocean (http://http.cs.berkeley.edu/ ssmith/GATOR.html).
The ultimate goal is an integrated system including
- satellties collecting large quantities of sensor data (up to 1 Tbyte/day),
- numerical models using this data to predict the climate up to 50 years
in the future,
- visualization systems to permit scientists to explore this data, and
- a data base to store all the satellite and simulation data.
The data base system is being built by the Sequoia project in the
CS Division. There are a number of issues involved in building this system:
- The computational bottleneck in GATOR is solving the ODEs of chemical
kinetics, for about 90 chemical species participating in about 200
reactions, at every grid cell into which the atmosphere is divided.
The current code was written and optimized for a Cray C90, which has a
much higher memory bandwidth than the cache-based microprocessors
of which most parallel computers are built. As a result, the floating
point performance is very poor. It may be possible to significantly
increase performance by using techniques like prefetching, which we used
on the IBM RS6000/590 to accelerate linear algebra.
- GATOR needs to exchange data with several other components:
atmospheric dynamics, ocean dynamics, the data base, and the visualization
system. All these systems operate on different grids of data, on different
machines, possibly in geographically distant locations, connected over
high speed networks. In order to design this complicated system,
we need detailed performance models of the components, and the networks
and format converters connected them, to identify bottleneck and fix them
before they are built into the architecture.
(Oleg Zakharov, oleg@savanna.berkeley.edu, and
Balazs Kralik, kralik@tiger.berkeley.edu, have signed up for this.)
Next: Adapting Rigorous Electromagnetic
Up: No Title
Previous: Sparse Cholesky.
Boris Vaysman
Sun Jan 14 22:42:23 PST 1996