Scientific
Computation
Hardware is the lowest
level in the CFD 'food chain' but it is by no means the least
important.
Graphics & Stream Processors
Graphics processors (GPUs)
and their relatives are special purpose but cheap commodity
processors that (fortuitously) are well suited for the types
of algorithms used in scientific computation.
-
In 2006 we were
seeing about 3x faster performance from low end (<$100) GPUs
compared to a regular desktop CPUs.
-
We will be testing
the IBM Cell processor in 2007 using the PlayStation 3 video
game from Sony.
-
Basic scientific
operations (similar to BLAS operations) can be carried out
on the GPU using our GPU-BLAS libraries.
-
A demo program that
uses the GPU for some off the calculations is
here.
-
E-mail if you would
like to try our GPU-BLAS library.
Parallel Processing
-
Our internally
developed codes use MPI and latency hiding techniques to
operate efficiently on all parallel machines - even those
with very slow (100 Mbit Ethernet) interconnection networks.
-
Our codes use
special versions of Krylov solvers that minimize global
communication and implicit parallel synchronization points.
-
All our algorithms
are chosen because they parallelize efficiently. For
example mesh adaptation via mesh movement is both more
parallel and maintains load balancing better than other mesh
adaptation algorithms (such as h or p refinement or point
insertion and removal).
Parallel Facilities
-
The laboratory uses
the DoD supercomputing facility in Alaska and NSF Teragrid facilities for our very large computations.
-
We maintain a
608 Processor cluster (Cyclops) for the College of
Engineering.
-
The group built a
Desktop Supercomputer (DC) call Orion.
Why the top500.org
web site (listing the fastest supercomputers) is popular and
well funded but is a dead weight on the parallel
processing world (diatribe here).
Back to
top
