Technical Report


Pipelining performance of structured dataflow networks Public Deposited

Downloadable Content

Download PDF


Attribute NameValues
Alternative Title
  • System architectures for interconnecting large numbers of processors are being widely studied [AG82,TH75,TR82]. Of particular interest in such architectures is the exploitation of concurrency among processors. This concurrency can be parallelism, in which different parts of a single data case are processed at the same time, or pipelining, in which different processes are carried out simultaneously on successive data cases. Many problems involving file and vector processing can be viewed as pipelining problems. Both structured and arbitrarily interconnected networks of processors have been proposed. The asserted advantage of structured networks is that they are easier to comprehend, control, and so utilize effectively (arguments analogous to those made for structured programming [DA72,KO74]). The arguments for arbitrary networks are that they allow greater flexibility of interconnection and so can be more efficient in many cases. Also, network architectures are being studied incorporating processor activation through data flow [AC82,AG82,DA82,TR82,WA82] and through flow of control [AN75,GA82,TH75]. In all of these cases, attention must be given to developing languages for expressing procedure interconnections and rules for allocating procedures to processors so as to obtain maximum concurrency. This paper presents a particular approach to specifying procedure interconnection and allocation. The major result is that, within stated assumptions: networks constructed using a small set of structured process connectives can achieve at least as good throughput (pipelining performance) as arbitrarily interconnected networks.
Resource Type
Date Issued
Academic Affiliation
Rights Statement
File Format



This work has no parents.