Hybrid tree: a scalable optoelectronic interconnection network for parallel computing

Abstract

The performance of parallel computer systems is increasingly limited by constraints imposed by interconnects and this limitation will inevitably become more serious as capability of each processing node increases. Optics has already proved its worth in telecommunications and more recently in interconnecting computers and computer peripherals. We propose an optical hybrid tree network which is essentially a cluster of optical binary trees interconnected with a fat tree near the root. Binary trees have favorable features such as constant node degree, small node degree, scalability, etc., however, in a conventional binary tree, the traffic towards the root of the tree is very heavy and hence a fat tree structure in which branches get thicker towards the root has been employed in computers such as the CM-5 from Thinking Machines. Use of optical channels results in increased bandwidth per channel and we suggest that the high bandwidth of optical interconnects and channels enables the use of binary tree topology for several levels of a tree network without congestion or traffic problems. We demonstrate that optical binary trees work satisfactorily for 7 to 10 levels higher than binary electrical trees. A 1024 node parallel machine can be built using binary trees alone if the interconnection network uses optical technology for the links and routers. We present the design of optical hybrid trees and perform a comparison of electrical and optical binary trees.