Comparative study of multiprocessor systems based on transputers and DSPS for image processing
Loading...
Date
1993
Authors
Kumarasamy, Seelan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The ever increasing requirement for high speed image processing has motivated much
research in parallel implementations of common image processing functions. This work
considers the implementation of low and intermediate level image processing functions
exhibiting differing data dependency characteristics on multiprocessor networks based on
transputers and Digital Signal Processors (DSPs). A number of parallel programming
paradigms and multiprocessor network architectures have been examined to determine the
factors that affect speedup and scalability performance of the algorithms and to identify
optimal solutions. A basic routing harness for dynamic processor farms and a routing
harness designed to exploit low-level parallelism and eliminate redundant copying have
been implemented and performance compared over a wide group of workloads,
distributions and network and workpacket sizes. The Fast Fourier and Hough transforms
have been implemented on transputer farms and other network architectures and the
degree to which parallelism and hardware resources may be exploited compared. Several
parallel solutions reported recently have been studied. Attention is given to the data
distribution and results compilation phases of the parallel solutions as this sometimes
neglected consideration can limit the performance of the implementation and the
adaptability of parallel solutions derived for non-transputer systems. System memory
requirement for the various implementations are also discussed. A hybrid processing
element with a DSP as a co-processor to a transputer, interfaced via shared memory, has
been designed to study the effect, on the algorithms studied, of enhancing the
computational capability of the transputer. Such an architecture yields minimal benefits
for 20 FFr computation, as the ratio of computation to communication capability is not
balanced. With the subimage Hough Transform and other severely computationally
bound implementations however, the hybrid network coupled with the efficient routing
harness leads to outstanding performance for farmed applications.
Description
Keywords
Multiprocessor system , Transputers