Optical packet switches that scale to thousands of input/output ports might find their application in nextgeneration datacenters (DCs). They will allow interconnecting the servers of a DC in a flat topology, providing higher bandwidth and lower latency in comparison with currently applied electronic switches. Using a simple analytic model that allows computing end-to-end latency and throughput, we show that optical interconnects that employ a centralized (electronic) controller cannot scale to thousands of ports while providing end-to-end latencies below 1 µs and high throughput. We therefore investigate architectures with highly distributed control. We present a strictly non-blocking wavelength division multiplexing architecture with contention resolution based on wavelength conversion. We study the packet loss probability of such architecture for different implementations of the contention resolution functionality. Furthermore, we show that the proposed architecture, applied in a short link with flow control, provides submicrosecond end-to-end latencies and allows high load operation, while scaling over a thousand ports.

No abstract

High-speed, submicrosecond-latency, largeport-count (thousands) optical packet switches (OPSs) for intercluster communication networks can become a key element in the deployment of cloud-oriented largescale data centers. In this work we numerically investigate the performance of a large-port-count wavelength-division multiplexing (WDM) OPS based on a Spanke-type architecture with highly distributed control. We analyze it under a data center traffic model to determine its suitability for this type of environment. Results indicate that the proposed architecture can be scaled to 4096 ports while providing packet loss below 10 −6 and latency under 1 μs, with a total switching capacity over 55 Tbits∕s. Additionally, we propose and analyze two WDM OPS architectures. The first one detects and processes small and large-sized Ethernet packets with two parallel switches. The second architecture includes multiple receivers to decrease packet losses and latency while using very limited electronic buffers. Results indicate that both techniques can lead to substantial improvements. In terms of packet loss and latency, they allow up to 40% higher input load with respect to the original WDM OPS architecture.

Data centers have to sustain the rapid growth of data traffic due to the increasing demand of bandwidth-hungry internet services. The current intra-data center fat tree topology causes communication bottlenecks in the server interaction process, power-hungry O-E-O conversions that limit the minimum latency and the power efficiency of these systems. In this paper we numerically and experimentally investigate an optical packet switch architecture with modular structure and highly distributed control that allow configuration times in the order of nanoseconds. Numerical results indicate that the candidate architecture scaled over 4000 ports, provides an overall throughput over 50 Tb/s and a packet loss rate below 10 −6 while assuring sub-microsecond latency. We present experimental results that demonstrate the feasibility of a 16x16 optical packet switch based on parallel 1x4 integrated optical cross-connect modules. Error-free operations can be achieved with 4 dB penalty while the overall energy consumption is of 66 pJ/b. Based on those results, we discuss feasibility to scale the architecture to a much larger port count.

We investigate the impact of switching architectures that scale to thousands of ingress and egress nodes on the node control. We give an example of an architecture that is highly scalable while supporting low latency