Quantized BvND: A Better Solution for Optical and Hybrid Switching in Data Center Networks

Abstract

Data center network continues to grow relentlessly in the amount of data traffic it has to "switch" between its server racks. A traditional data center switching architecture, consisting of a network of commodity packet switches (viewed as a giant packet switch), cannot scale with this growing switching demand. Adding an optical switch, which has a much higher bandwidth than the packet switch but incurs a nontrivial reconfiguration delay, to a data center network has been proposed as a costeffective approach to boosting its switching capacity. However, to effectively do so, we need to meticulously schedule the optical switch. In fact, we are dealing with two very different scheduling problems here, namely hybrid switching and standalone optical switching, depending on whether or not there is effective cooperation between the optical switch and the packet switch during their respective scheduling processes. In this work, we propose a solution that performs better than the respective state of art solutions for both scheduling problems. Our solution outperforms by a wide margin all existing optical switching solutions in terms of throughput, yet its computational complexity is comparable to those of others. Our solution also has the best properties of both Eclipse and Solstice, the state of the art hybrid switching solutions. Eclipse and Solstice have different advantages: Eclipse has better throughput performance but incurs a much higher computationally complexity than Solstice. Our solution gets the better of both worlds: it delivers almost the same throughput performance as Eclipse, yet incurs a similar computational complexity as Solstice.