Efficient parallel processing with optical interconnections

Author

Lili Hai, New Jersey Institute of Technology

Document Type

Dissertation

Date of Award

Spring 5-31-1997

Degree Name

Doctor of Philosophy in Computing Sciences - (Ph.D.)

Department

Computer and Information Science

First Advisor

Mary M. Eshaghian

Second Advisor

John D. Carpinelli

Third Advisor

Franz J. Kurfess

Fourth Advisor

Peter A. Ng

Fifth Advisor

Muhammad Esmat Shaaban

Abstract

With the advances in VLSI technology, it is now possible to build chips which can each contain thousands of processors. The efficiency of such chips in executing parallel algorithms heavily depends on the interconnection topology of the processors. It is not possible to build a fully interconnected network of processors with constant fan-in/fan-out using electrical interconnections. Free space optics is a remedy to this limitation. Qualities exclusive to the optical medium are its ability to be directed for propagation in free space and the property that optical channels can cross in space without any interference. In this thesis, we present an electro-optical interconnected architecture named Optical Reconfigurable Mesh (ORM). It is based on an existing optical model of computation. There are two layers in the architecture. The processing layer is a reconfigurable mesh and the deflecting layer contains optical devices to deflect light beams. ORM provides three types of communication mechanisms. The first is for arbitrary planar connections among sets of locally connected processors using the reconfigurable mesh. The second is for arbitrary connections among N of the processors using the electrical buses on the processing layer and N² fixed passive deflecting units on the deflection layer. The third is for arbitrary connections among any of the N² processors using the N² mechanically reconfigurable deflectors in the deflection layer. The third type of communication mechanisms is significantly slower than the other two. Therefore, it is desirable to avoid reconfiguring this type of communication during the execution of the algorithms. Instead, the optical reconfiguration can be done before the execution of each algorithm begins. Determining a right configuration that would be suitable for the entire configuration of a task execution is studied in this thesis. The basic data movements for each of the mechanisms are studied. Finally, to show the power of ORM, we use all three types of communication mechanisms in the first O(logN) time algorithm for finding the convex hulls of all figures in an N x N binary image presented in this thesis.

Recommended Citation

Hai, Lili, "Efficient parallel processing with optical interconnections" (1997). Dissertations. 1044.
https://digitalcommons.njit.edu/dissertations/1044