Optical free-space interconnection network for parallel computers

Dror G. Feitelson*, Larry Rudolph, Eugen Schenfeld

*Corresponding author for this work

Research output: Contribution to journalConference articlepeer-review

Abstract

The authors present an architecture for a crossbar-like, free-space interconnection network. The design follows principles and guidelines derived from an architectural study of massively parallel computers. The design has reconfigurability. Each processor can change its connections dynamically, independently of the connections made by other processors. Scalability is up to a very large number of processors (more than 10,000). Distributed operation means that no central controller is used to coordinate the network's interconnection pattern. Prudent use of the 3-D space. The design meets these goals by using a two level communication scheme formed by a local interconnection network on a single board that connects a set of processors by a fast switching crossbar. The number of processors connected is relatively small (up to 100), which enables the fast switching time (as low as nanoseconds). A global interconnection network between the boards is based on a routing switch associated with each processor. The switch is implemented by a dynamic hologram. A static hologram on each board is used to deflect the light beams towards the destination processors.

Original languageEnglish
Pages (from-to)233-234
Number of pages2
JournalProceedings of SPIE - The International Society for Optical Engineering
Volume1359
StatePublished - 1990
Event1990 International Topical Meeting on Optical Computing - OC '90 - Kobe, Jpn
Duration: 8 Apr 199012 Apr 1990

Fingerprint

Dive into the research topics of 'Optical free-space interconnection network for parallel computers'. Together they form a unique fingerprint.

Cite this