Skip to main content
eScholarship
Open Access Publications from the University of California

Performance analysis of two classes of data flow computing systems

  • Author(s): Thomas, Robert Eugene
  • et al.
Abstract

Dataflow may be thought of as a language-oriented approach to the design and organization of computing machines. A particular reason for considering dataflowas a basis for the design of a machine is that it offeres theoretically sound semantics for achieving highly parallel deterministic computation. Whether or not dataflow becomes a viable alternative to conventional (von Neumann) machine organization may well depend on the level of performance obtained by hardware based on dataflow principles. Since the fundamental concepts of dataflow are different from those of the von Neumann model, it is believed that conventional predictors of performance are unsuitable for dataflow machines. As a beginning for understanding dataflow performance, this thesis develops a method to analyze the execution of dataflow programs to determine how time is utilized. In order to substantiate the analytical results a simulation model was constructed on which dataflow programs can be executed to observe relative time and resource utilization. The results of the study are a set of parameters and equations which roughly approximate the computation time required for certain types of dataflow programs on two different classes of dataflow architectures - the preallocated and the dynamically allocated machines.

Main Content
Current View