Dataflow computing models, languages, and machines for intelligence computations

Abstract

Dataflow computing, a radical departure from von Neumann computing, supports multiprocessing on a massive scale and plays a major role in permitting intelligence computing machines to achieve ultrahigh speeds. Intelligence computations consist of large complex numerical and nonnumerical computations. Efficient computing models are necessary to represent intelligence computations. An abstract computing model, a base language specification for the abstract model, high-level and low-level language design to map parallel algorithms to abstract computing model, parallel architecture design to support computing model and design of support software to map computing model to arcTiitecture are steps in constructing computing systems. This paper concentrates on dataflow computing for intelligence computations and presents a comparison of dataflow computing models, languages and dataflow computing machines for numerical and nonnumerical computations. The high level language-graph transformation that must be performed to achieve high performance for numerical and nonnumerical programs when executed in a dataflow computing environment are described using the DCBL transformations and applied to the Lisp language. Some general problems in dataflow computing machines are discussed. Performance evaluation measurements obtained by executing benchmark programs in the ETL's nonnumerical dataflow computing environment, the EM-3, are presented.

A Reliability and Performance Model For Static Data Flow Machines

“…Dennis [1,11,26, 27] has identified three basic issues regarding the development of an ideal architecture for data flow machines. The first one is to achieve a high performance/cost ratio, the second is to match the ratio with technological progress, and the third is to offer better programmability in application areas.…”

“…The data flow architecture can be classified as static architecture or dynamic architecture. The static scheme was first proposed by Dennis [1,11,27], and has been used by various research architectures such as the Texas Instruments (TI) DDP and LAU architectures. The Dynamic scheme is used in Arvined's research group at MIT [28] and at Manchester University [11,28].…”

“…A block diagram of the architecture is shown in Figure 1. It consists of the memory units, the processing section, the arbitration network and the distribution network [1,11]. Each memory unit is a collection of blocks of storage locations.…”

“…Data flow and very large-scale integration (VLSI) offer two mutually supportive approaches towards the design of future supercomputers. When very high speed computations are needed, data flow machines may be relied upon as a promising solution in which extremely parallel processing is achieved [1,2,3,4,5]. In contrast to Von-Neumann architecture, data flow machines possess many processors that work in parallel.…”

“…In this paper, three different data flow machines are considered. These machines are the Massachusetts Institute of Technology (MIT) machine, the data-driven processor (DDP) and the LAU static data flow machine, a 32-processor system built at the Department of Computer Science, ONERA/CERT, France [1,11].…”