1420 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 19, NO. 10, OCTOBER 2001 WDM Packet Routing for High-Capacity Data Networks Qimin Yang, Student Member, IEEE, Keren Bergman, Member, IEEE, Member, OSA, Gary D. Hughes, and Frederick G. Johnson Abstract—We present experimental and numerical studies of a novel packet-switch architecture, the data vortex, designed for large-scale photonic interconnections. The selfrouting multihop packet switch efficiently scales to large port counts ( k) while maintaining low latencies, a narrow latency distribution, and high throughput. To facilitate optical implementation, the data-vortex architecture employs a novel hierarchical topology, traffic control, and synchronous timing that act to reduce the necessary routing logic operations and buffering. As a result of this architecture, all routing decisions for the data packets are based on a single logic operation at each node. The routing is further simplified by the employment of wavelength division multiplexing (WDM)-encoded header bits, which enable packet-header processing by simple wavelength filtering. The packet payload remains in the optical domain as it propagates through the data-vortex switch fabric, exploiting the transparency and high bandwidths achievable in fiber optic transmission. In this paper, we discuss numerical simulations of the data-vortex performance and report results from an experimental investigation of multihop WDM packet routing in a recirculating test bed. Index Terms—Optical networks, optical packet switch, wave- length division multiplexing (WDM) optical packet routing. I. INTRODUCTION W ITH THE ever-growing demand for higher data rates and a wide variety of services, a fully transparent optical-switch network element presents an attractive way to overcome the electronic speed bottleneck. However, all-optical implementations encounter considerable challenges in the processing and buffering of optical data, which are essential to ensuring switch performance and to providing the neces- sary services and protections. Therefore, most optical-switch architectures actually employ hybrid optical-to-electronic tech- nologies where the strengths of optics such as transparency and Manuscript received January 23, 2001; revised May 30, 2001. This work is sponsored by the Defense Advanced Research Projects Agency and by the Na- tional Security Agency through an agreement with the National Aeronautics and Space Administration under Contract 960 199, and by the National Science Foundation under contract ECS98-00 401. Q. Yang is with the Electrical Engineering Department, Princeton University, Princeton, NJ 08544 USA (e-mail: qyang@ee.princeton.edu). K. Bergman was with the Electrical Engineering Department, Princeton Uni- versity, Princeton, NJ 08544 USA. She is now with Tellium, Inc., Oceanport, NJ 07757 USA. G. D. Hughes is with the Laboratory of Physical Science, College Park, MD 20742 USA. F. G. Johnson was with the Laboratory of Physical Science, College Park, MD 20742 USA. He is now with Little Optics, Inc., Annapolis Junction, MD 20701 USA. Publisher Item Identifier S 0733-8724(01)08794-1. high bandwidth can be fully exploited while the weaknesses can be bypassed in the electronic domain. Nevertheless, issues such as contention resolution, buffering, scalability, and latency remain the key challenges in photonic packet switching [1]–[4]. To achieve high throughputs, a packet switch generally pro- cesses multiple packets simultaneously. Therefore, whether the architecture is based on a single hop or multiple hops, contention is inevitable when multiple packets are competing for the same output port. To solve the contention problem, either buffering or deflection techniques can be implemented. Simple deflection methods without buffers (hot-potato routing) usually introduce severe performance penalties in throughput, latency, and latency distribution [5]. Current optical buffering techniques configured as optical fiber delay lines in a traveling or a recirculating geom- etry do not have random access capability [6], [7]. The straight delay-line buffers can introduce large timing errors and the re- circulating rings are generally bulky and expensive. Therefore, it is still impractical to employ optical buffers within a switching network element, especially for very large-scale switches. In recent research, the use of wavelength division mul- tiplexing (WDM) in packet switches has been explored to achieve greater flexibility. Technologies such as wavelength routing (WR) and wavelength conversion (WC) are proposed to solve the difficulty of optical buffering [8]–[12]. These techniques are very attractive in maintaining the throughput performance as well as the optical data transparency without resorting to optical buffers. The wavelengths are exploited as logical buffers within the existing architectures. Generally, however, the additional WR and WC devices add performance penalties and make such systems complex and costly. In this paper, we present experimental and numerical studies of a new architecture, the data vortex, as a candidate for large-scale packet switching [13]. The multihop packet switch tightly couples the deflection method and a virtual buffering mechanism to achieve hardware simplicity, scalability, and high throughput. The timing and control algorithm permits only one packet to be processed at each node in a given clock frame and, therefore, the need to process contention resolution is elimi- nated. The wavelength domain is additionally used to enhance the throughput and to simplify the routing strategy. Numerical studies of the traffic flow within the switch architecture have shown that the data vortex can efficiently scale to greater than 10 000 ports while maintaining a low packet switching latency and a narrow latency distribution. The rest of the paper is organized as follows. Section II pro- vides an overview of the data-vortex switch architecture, and in Section III, the numerical studies of the system performance are 0733–8724/01$10.00 © 2001 IEEE