Digital signal conditioning on multiprocessor systems

Abstract

An important application area of modem computer systems is that of digital signal processing. This discipline is concerned with the analysis or modification of digitally represented signals, through the use of simple mathematical operations. A primary need of such systems is that of high data throughput. Although optimised programmable processors are available, system designers are now looking towards parallel processing to gain further performance increases. Such parallel systems may be easily constructed using the transputer family of processors. However, although these devices are comparatively easy to program, they possess a general von Neumann core and so are relatively inefficient at implementing digital signal processing algorithms. The power of the transputer lies in its ability to communicate effectively, not in its computational capability. The converse is true of specialised digital signal processors. These devices have been designed specifically to implement the type of small data intensive operations required by digital signal processing algorithms, but have not been designed to operate efficiently in a multiprocessor environment. This thesis examines the performance of both types of processors with reference to a common signal processing application, multichannel filtering. The transputer is examined in both uniprocessor and multiprocessor configurations, and its performance analysed. A theoretical model of program behaviour is developed, in order to assess the performance benefits of particular code structures and the effects of such parameters as data block size. The transputer implementation is contrasted with that of the Motorola DSP56001 digital signal processor. This device is found to be much more efficient at implementing such algorithms on a single device, but provides limited multiprocessor support. Using the conclusions of this assessment, a hybrid multiprocessor has been designed. This consists of a transputer controlling a number of signal processors, communicating through shared memory, separating tiie tasks of computation and communication. Forcing the transputer to communicate through shared memory causes problems, and these have been addressed. A theoretical performance model of the system has been produced. A small system has been constructed, and is currently running performance test software.