Petascale: A Scalable Buffer-Less All-Optical Network for Cloud Computing Data Center

Abstract

Optical packet switching is considered as a long-term solution for future data center networks due to their technological strengths, including high throughput, fine switching granularity, and excellent flexibility. However, most prior optical designs are either difficult to scale, costly to solve the packet collision, or inflexible to exploit the multiple wavelengths to increase the end-to-end connectivity. To this end, we present PETASCALE, a scalable, lossless, and high-performance optical data center network. By embedding full-mesh into a bipartite graph, PETASCALE is able to scale beyond 60 000 servers while achieving up to 2.5 and 2 times saving in a number of switches and cables, respectively, compared with fat tree topology. PETASCALE leverages the negative acknowledgment (NACK) and retransmission scheme to solve the packet collision, thus eliminating the complex and costly buffers. Then by leveraging a state retention mechanism, the contention can be notified over multiple hops in near real time. To further improve the bandwidth utilization, a wavelength routing algorithm is proposed to restrict the packet collision domain within the source pod. Moreover, PETASCALE designs a novel switch structure to support both the multi-hop NACK and wavelength routing. Our simulation results show that PETASCALE is able to reduce the end-to-end latency by more than 50% compared with an electrical fat tree network. For throughput, PETASCALE can deliver about 89% bisection bandwidth of a non-blocking network under uniform traffic pattern. While under the local traffic pattern, it can even achieve higher throughput than fat tree and other optical designs.