Continuous Models for Communication Density Constraints on Multiprocessor Performance

Reviewer: Borivoje Djokic

This well-written paper presents a continuous model for communication density in large multiprocessor arrays. The authors use the criterion of information density to measure the difficulty and cost of building a machine's interconnect technology; they justify this choice by pointing out that information density does have a physical maximum based on material properties and information theory. The paper discusses the implications of identifying communication density as a fundamental quantity. After reading the section that quantifies the degree by which communication density must increase as the number of processors grows (for finite machines) one can much better understand the constraints involved in building a machine. The authors have achieved the main goal of their research: to provide the computer designer with the necessary theoretical background and a model for reasoning about machine tradeoffs when only general requirements are known and before a network topology has been chosen. The authors' statements are very general, as they are concerned with the functionality of the network rather than its design. I consider this to be a strength of their approach. The reader should have some knowledge of multiprocessor architecture, interprocessor communication, and information theory. The authors present their material effectively and in a logical and understandable way.