|
Program Manager: Dr. Dennis Healy (MTO)
The Discovery and Exploitation of Structure
in Algorithms (DESA) Program seeks a
dramatic reduction of expert human programmer
involvement in porting critical high-performance
algorithmic capabilities into new and
upgraded computing platforms and into
new classes of processing hardware. Design
tools are sought to enable programmers
without platform-specific expert ise (e.g.,
application experts) to produce algorithm
implementations that are optimized to
a target platform. As new platforms emerge
requiring substantially different algorithm
implementations in order to make best
use of their architectural characteristics,
DESA tools will enable these implementations
to be produced by application practitioners
without investment in platform specific
expertise—thereby enabling rapid
and inexpensive porting of legacy software
to new host machines. Similarly, DESA
tools are sought to allow application
practitioners to optimally partition
hardware-firmware in algorithm implementations
on new hybrid platforms and to produce
application-specific integrated circuit
(ASIC) designs exploiting new technological
capabilities.
DESA is prototyping major advances
in software engineering and processor
(ASIC) design by demonstrating mechanisms
for producing libraries of algorithmic
primitives that are both optimized and
portable. In the first phase of the program,
researchers demonstrated a tool that
enabled power efficient implementations
of the Fast Fourier Transform that were
more efficient then expert-developed
implementations and required 1/80th the
development time. Performers developed
an architecture and implementations of
the Fast Multipole Method, which outperformed
previous Gordon Bell prize winning machines. In
the program’s second phase, high-level
algorithm specification tools will be
produced to allow application domain
experts (e.g., engineers in signal processing
or fluid dynamics) to specify algorithmic
library modules.
Successful realization of the DESA
vision will lower DoD system costs and
decrease design cycle time currently
entailed by extensive hand-tuning of
software module, firmware, and ASIC designs
for the most computationally intensive
mathematical operations in DoD systems.
The primary areas of potential national
security impact for the tools to be produced
by this program are in digital signal
processing applications (e.g., communication
systems, radar, and other sensing systems)
and computational physics (e.g., computational
fluid dynamics for aerodynamics and combustion
modeling, and computation electromagnetics
for low-observable vehicle design).
|
