DISTRIBUTED PARALLEL-PROCESSING - LESSONS LEARNED FROM A 160-TRANSPUTER NETWORK

The composition and performance of music is a plural activity that com bines the outcomes of a number of procedures, many of which involve fu nctions that operate in parallel. in terms of sound-synthesis operatio ns, a significant number of generative and signal-processing operation s involve a combination of concurrent elements, ranging from the produ ction of simultaneous notes by a single instrument to the superimposit ion of totally independent outputs, where a number of different compon ents contribute to the audio spectrum. The traditional computer proces sor is a serial device, restricted for the most part to the execution of instructions as a single stream of events. Thus processes that requ ire the aggregation of functions executed in parallel must be simulate d by some means of cyclical tasking and data accumulation. In the case of digital audio synthesis and signal processing applications, the re sultant effects on overall processor performance quickly become signif icant, thus limiting the number of individual components that can be h andled in real time. About ten years ago, the Music Technology Group a t the University of Durham started a series of investigations into the construction of computing architectures for audio applications that e mbraced a significant degree of true parallelism, based in the first i nstance on the INMOS Transputer. This article describes some of the mo st important outcomes of this particular Line of investigation, and hi ghlights aspects that hold a particular relevance for future designs o f parallel audio processors.