Volume 7, Issue 1 -
Spring/Summer 1999
Volume 6, Issue 3
Fall 1998
Volume 6, Issue 2
Spring/Summer 1998
Volume 6, Issue 1
Winter 1998
Volume
5, Issue 4
Fall 1997
Volume
5, Issue 3
Summer 1997
Volume
5, Issue 2
Spring 1997
Volume
5, Issue 1
Winter 1997
Volume
4, Issue 4
Fall 1996
Volume
4, Issue 3
Summer 1996
Volume
4, Issue 2
Spring 1996
Volume
4, Issue 1
Winter 1996
Volume
3, Issue 4
Fall 1995
Volume
3, Issue 3
Summer 1995
Volume
3, Issue 2
Spring 1995
Volume
3, Issue 1
January 1995
Volume
2, Issue 4
October 1994
Volume
2, Issue 3
July 1994
Volume
2, Issue 2
April 1994
Volume
2, Issue 1
January 1994
Volume 1, Issue 4
October 1993
Volume
1, Issue 3
July 1993
Volume
1, Issue 2
April 1993
Volume
1, Issue 1
January 1993
|
Projects on the Delta
With 570 nodes and a peak performance of 32 GFLOPS, the Intel Delta at
the CRPC's Caltech site has been highly effective for several different
research projects. The following projects described below are a few
examples of how researchers have used this resource.
Aviation Gridded Forecast System
Leslie Hart, Young Chun, Tom Henderson,
Bernando Rodriguez, Frank Tower, Science and Technology Corporation,
Boulder, CO
Advancements in real-time forecasting have immediate benefits to both
the domestic economy and the national defense. Due to the National
Weather Service's new radar and satellite-based sensing systems, the
unprecedented quantity of data provided for these models has resulted in
the need for increased computing capability. As a parallel machine, the
Delta has provided this capability at an order of magnitude that is more
cost-effective than traditional vector supercomputers.
Using CRPC-allocated time, researchers in this project used the Delta to
test their conversion of sequentially oriented weather models to
parallel forms. Their objective is to develop a parallel weather
analysis and forecasting system with a 15 km resolution and a three-hour
turnaround time on a 24-hour forecast. Researchers hope to improve the
resolution to four km while maintaining or improving the three-hour
turnaround time.
Other challenges to developing an efficient parallel weather analysis
and forecast system include achieving portability on several parallel
machines and specifying the memory architecture and algorithms to
dictate data movement requirements. Research in operating systems and
compilers has also been pursued to supplement these developments.
Concurrent Solver for the Euler Equations
Dan Meiron, Eric Van de Velde,
John Hofhaus, Paul Hardy, California Institute of Technology
This project involves the development and use of a numerical scheme to
integrate two-dimensional Euler equations in an annulus. The numerical
scheme will be used as a basis for high resolution simulations of
inviscid and viscid fluid flows. From this simulation, researchers can
examine the late-time dynamics of vortex structures that emerge due to
the special nature of energy transfer in two-dimensional fluid flows.
With sufficient spatial resolution provided by the use of the Delta,
several issues can be investigated in the examination of these vortex
structures. These issues include the mechanisms of an irreversible
approach to negative temperature states in the two-dimensional flows and
the nature of energy distribution at equilibrium. Several theories that
have been advanced regarding these issues can now be tested with the
help of the simulations being developed through this project. The
algorithms being used are well suited for use on parallel machines.
In addition, axisymmetrical three-dimensional flows can be simulated by
modifying these algorithms slightly. These three-dimensional flows have
been of use lately to study the possible singular behavior of the three-
dimensional equations of motion. A significant difference between two-
dimensional and three-dimensional flow is the possibility of vortex
amplification through vortex stretching. It is not yet known whether
divergent vorticities can be attained in a finite time as a result of
such stretching. The computational capabilities of the Delta are
providing sufficient resolution to address these issues.
A third project involves the translation of the parallel algorithms
described above to the Fortran M language for parallel computation.
Fortran M is a dialect of Fortran 77 developed by CRPC researchers Mani
Chandy and Ian Foster. The language adds a few extensions to the Fortran
77 language which allow one to construct parallel programs in which
message passing can be performed in a way which is independent of the
underlying architecture. In this way it is possible to guarantee that
parallel programs written in Fortran M will perform deterministically.
In collaboration with Chandy and a group of undergraduate students, this
group has been developing "templates" in the Fortran M language that
allow one to write parallel code for various common data distributions
and have all aspects of communications hidden in the low level details
of the underlying template. This approach to parallel programming has
the advantage that, as long as one interfaces with the data structures
and utilities provided through the template, parallel programs in
Fortran M appear at the source level to be more or less identical with
their sequential counterparts. The group is in the process of
constructing templates for several common regular data distributions
used in scientific computing. The scientific problems described above
both possess data distributions that are amenable to this approach.
Currently they are implementing these codes on networks of workstations
but plan to port these codes to the high-performance SP1 machine at
Argonne.
Numerical Simulations of Quantum Gravity Using Random Surfaces
Paul
Coddington, Enzo Marinari, Mark Bowick, Leping Han, Geoff Harris,
Syracuse University
String theories are quantum field theories in which the fundamental
particles are tiny one-dimensional strings, rather than points with no
dimension. There has been great theoretical interest in string theories,
since together they provide a long-awaited quantum theory of gravity, as
well as reproduce the standard quantum model of the other fundamental
forces of nature, and thus provide a possible TOE (Theory of
Everything). However, string theories have a major problem--calculations
are usually analytically intractable. Methods are currently being
developed to alleviate this problem by doing numerical calculations
using computer simulation.
String theory calculations involve integrating over all possible two-
dimensional surfaces swept out by the string in some higher dimensional
space-time. In order to compute this integral numerically, the surfaces
are discretized as a triangulated mesh. The integral is then
approximated by a sum over a large number of different meshes, which are
obtained by making random changes to the mesh throughout the
calculation, using a Monte Carlo method. The mesh is thus referred to as
a dynamically triangulated random surface.
Currently, the research group is running simulations on the Intel Delta
and on networks of workstations by using the trivial parallelism of
averaging the results of independent simulations on different
processors. However, this can only be done effectively for small meshes.
The group is currently working on a data parallel algorithm for larger
meshes. Since both the data and the algorithm are dynamic and irregular,
this is a challenging problem, which requires parallel algorithms for
graph coloring, graph partitioning, load balancing, adaptive mesh
generation, as well as the Monte Carlo update.
Parallelization of an All-electron Density-functional Program
Peter
Nordlander, Richard Smalley, Liang Lou, Rice University
This group is developing a parallel version of software program that
calculates the electronic and geometric structures of polyatomic
systems, specifically semiconductor and transition metals as well as
hollow-cage carbon structures, new materials known as metcars. The
parallelization of the program arises from the use of a recently
introduced multi-center numerical integration scheme that is inherently
suited for vectorized parallel processing. This numerical integration
scheme is based on the discretization of the one-electron Schršedinger
equation on a grid of sampling/integration points. This scheme has
yielded the most streamlined code out of several possible approaches for
partitioning the space and setting up the integration points.
The sequential version of this program has been implemented on several
different workstations and used in the study of clusters of various
types and sizes, such as C60 structures (Buckminsterfullerene, or "Bucky
balls"), small gallium arsenide clusters, NH3-cluster chemisorption
systems, GaAs clusters, and some organometallic cage structures.
Although program performance is very high on sequential computers, a
single CPU limits the calculations involved in accurately describing the
electronic structures; these calculations involve several hundred basis
functions. The Delta, however, provides sufficient computational power
to run these functions. This has allowed the researchers to
systematically study cluster systems with extended size range.
Table of Contents
|