A Hardware Testbed for Distributed Heterogeneous Adaptive Computing
Christopher Bachmann
Abstract
Adaptive computing systems are composed of reconfigurable hardware
elements that are capable of adapting to various computation
requirements. Using these reconfigurable hardware elements, adaptive
computing systems provide significant performance improvements over
conventional non-reconfigurable systems. However, completely
reconfigurable systems, such as those based entirely around FPGAs, become
very unsuitable for implementing complex applications. This is due to
both the time to develop algorithms and then implement them, as well as
the FPGAs being slower than processor's internal computational units when
implementing the same functions. The solution, being developed by the
MATCH project, is to merge various commercial-off-the-shelf (COTS)
components into one cohesive unit. Such components would include not only
FPGAs, but DSPs and embedded processing elements as well. This
allows the advantages of each element to be combined, such as the ability
of the FPGAs to deal with more custom processing. The primary
goal of the MATCH project is to explore methods for bridging the gap
between the high level of applications developers, and the low level
control required by the adaptive computing environment. This would
include new developments in compiler technologies to take advantage of
such a heterogeneous environment. In order to test the new theories and
ideas of the MATCH system developers a testbed was developed including
the necessary custom components targeted by the MATCH system developers.
The actual testbed itself consists of a Motorola embededded board, an
Annapolis Microsystems WILDCHILD, a Transtech DSP board, and a Force
5V Sparc based board. This thesis will present prior work that has
been done in the area of adaptive computing, the testbed that we have
developed, and the results of various benchmarks on the testbed,
including matrix multiply on the various components.
Gzipped Postscript version of the paper