Parallel Computing Using Optical Interconnections

by Keqin Li (Editor), Yi Pan (Editor), Si-Qing Zheng (Editor)

Overview

Advances in optical technologies have made it possible to implement optical interconnections in future massively parallel processing systems. Photons are non-charged particles, and do not naturally interact. Consequently, there are many desirable characteristics of optical interconnects, e.g. high speed (speed of light), increased fanout, high bandwidth, high reliability, longer interconnection lengths, low power requirements, and immunity to EMI with reduced crosstalk. Optics can utilize free-space interconnects as well as guided wave technology, neither of which has the problems of VLSI technology mentioned above. Optical interconnections can be built at various levels, providing chip-to-chip, module-to-module, board-to-board, and node-to-node communications.
Massively parallel processing using optical interconnections poses new challenges; new system configurations need to be designed, scheduling and data communication schemes based on new resource metrics need to be investigated, algorithms for a wide variety of applications need to be developed under the novel computation models that optical interconnections permit, and so on.
Parallel Computing Using Optical Interconnections is a collection of survey articles written by leading and active scientists in the area of parallel computing using optical interconnections. This is the first book which provides current and comprehensive coverage of the field, reflects the state of the art from high-level architecture design and algorithmic points of view, and points out directions for further research and development.

Product Details

ISBN-13:	9780792382966
Publisher:	Springer US
Publication date:	10/31/1998
Series:	The Springer International Series in Engineering and Computer Science , #468
Edition description:	1998
Pages:	280
Product dimensions:	6.10(w) x 9.25(h) x 0.24(d)

Read an Excerpt

Chapter 1: Lightning Network and Systems Architecture

Abstract: Lightning is a dynamically reconfigurable, hierarchical, WDM network testbed project designed to interconnect a large number of supercomputers and create a distributed shared memory environment. The hierarchical architecture achieves scalability yet avoid the requirement of multiple wavelength tunable devices per node. Single-hop all-optical communication is achieved: a packet remains in the optical form from source to destination and does not require intermediate routing. The wavelength multiplexed hierarchical structure features wavelength channel re-use at each level, allowing scalability to very large system sizes. It partitions the traffic between different levels of the hierarchy without electronic intervention in a combination of wavelength- and space-division multiplexing. This approach can dynamically vary the bandwidth provided to different levels of the hierarchy. Each node monitors the traffic intensities on its subset of channels and detects any temporal or spatial shift in traffic balance. Bandwidth re-allocation is completely decentralized, any node can initiate it, achieving highly fault tolerant system behavior.

1.1 Introduction

Computer systems have developed rapidly over the past few years. Specifically, processor speeds continue to increase but 1/O capability continues to lag. This is not a new trend yet no general solution to this perpetual problem has emerged. This work describes an approach to support high-performance 1/O both interprocessor communication and remote file system access - with a combination of high speed optical network designed specifically for this situation.

The objective is to develop a scalable technique for clustering: a strategy that is effective (both in performance and price/performance) with both workstation-class processor interconnection and high-performance supercomputer-system level interconnection. The interconnection strategy needs to be flexible to adapt to the severe cost constraints at the low-end and performance requirements at the high-end. A general architecture for such a strategy is defined in this work. The feasibility of this architecture was proven in an experimental testbed known as "LIGHTNING," a joint project involving the State University of New York at Buffalo (Buffalo, NY), University of Maryland (College Park and Baltimore County, MD), Laboratory for Physical Sciences (College Park, MD), Center for Computational Sciences (Bowie, MD), and David Sarnoff Research Center (Princeton, NJ).

The optical network is hierarchical and based on wavelength-division multiple access (WDMA). WDM creates multiple channels, separated in wavelength, to be concurrently transmitted on a single fiber. The channels can be individually accessed, routed and switched. The architecture can be viewed as a tree, where processors reside at the leaves and the internal nodes have wavelengthand spatial-switching capabilities. This class of architecture for processor interconnection was first described in [1]. Section 1.2 briefly describes the structure.

An advantage of this architecture is that bandwidth can be dynamically re-allocated throughout the system to adapt to shifts in traffic patterns. The bandwidth assigned to different levels of the hierarchical system can be dynamically increased or decreased, based on the reference patterns and file activity of the system. However, in comparison to traditional reconfigurable architectures, the reconfiguration of the system is automatic and hidden from the user. The user does not have to logically map an application to a specific topology or inform the operating system at compile- or run-time of its intended communication patterns. LIGHTNING senses the traffic patterns inherent to an application and adapts itself accordingly. This function is hidden from the user and the operating system and avoids placing the burden of understanding the specifics of the system on the programmer. The programmer views the system as a pool of processors and is not involved with process placement. The operating system, during initial process placement and during migration, favors placing a process within the level-1 or level-2 cluster depending on size.

The goal of the testbed currently being developed is to support a (weakly coherent) distributed shared memory environment between the interconnected processors. This is accomplished through a combination of the network architecture, operating system, and network/memory interface design. The goal is not just to provide a high-bandwidth network, but to deliver that bandwidth to the application rather than wasting it in operating system overhead.

This chapter briefly describes the basic architecture and media access protocol in Section 1.2. Section 1.3 describes the support for reconfiguration in greater detail and describes the interaction with the media access protocol. Section 1.4 gives some performance results for the reconfiguration algorithms. Finally, some conclusions are given in Section 1.5.

1.2 Lightning Architecture

This section describes the architecture of the LIGHTNING network. First background information is discussed providing a framework for discussion, and then the architecture is defined. The media access protocol is then presented.

The LIGHTNING architecture was influenced by some of the strengths and weaknesses of the Fat-Tree [2]. LIGHTNING retains the advantages of the FatTree structure - a low latency, scalable interconnection - while avoiding some of the limitations through optical interconnects. Optical networks have many advantages over metallic interconnections such as a relaxed bandwidth distance product, large fan out, low power consumption, reduced crosstalk and immunity to electromagnetic noise (3). Major considerations in determining the best placement for optical connections is the speed mismatch between the optical and electrical components [4] and the performance/cost requirements. Optical fiber has a 30 THz bandwidth [5], much larger than the metallic bandwidth...

Table of Contents

Optical Interconnection Networks and System Architectures.- Lightning Network and Systems Architecture.- Parallel Computing with “Intelligent Optical Networks”.- Scalable Optical Interconnection Networks for Large-Scale Parallel Computers.- The Communication Capabilities of Partitioned Optical Passive Stars Networks.- Otis Optoelectronic Computers.- On Wavelength Assignment in WDM Optical Networks.- Models and Algorithms for Optical Interconnections.- An Abstract Model for Optical Interconnection Networks.- A Unique Design of Fiber-Optic Interconnection Networks and Algorithms.- Fundamental Algorithms for the Array with Reconfigurable Optical Buses.- Computing the Hough Transform on Arrays with Reconfigurable Optical Buses.- Basic Data Movement Operations on the LARPBS Model.- Fast Matrix Multiplication and Related Operations Using Reconfigurable Optical Buses.

Editorial Reviews

A collection of survey articles written by leading scientists in the area of parallel computing using optical interconnections, designed to provide an overview of the state of the art of the field. Concentrates on the potential of using optical interconnections in massively parallel process systems, and their effects on system design and parallel computation. Looks at implementations, design, and evaluation of various optical interconnects; examines data communication schemes, embedding, and mapping methods; and presents models for parallel computing using optical interconnections. Also discusses algorithms design and development, and applications. Annotation c. by Book News, Inc., Portland, Or.

Booknews