Abstract

The growing popularity of cloud and multimedia services is increasing the traffic volume that each datacenter (DC) needs to handle. As a consequence, the serious bottlenecks in DC networks in terms of both capacity and energy consumption need to be addressed. DC networks typically consist of edge tier, aggregation tier and core tier, which interconnect different servers within a DC as well as provide the interfaces to the Internet. Current large-scale DC network architectures are based on the fat-tree three-tiers topology [1] and on electronic switches, which are not able to scale to meet future traffic requirements in a sustainable manner. Reducing the power required by the inter- and intra-rack communication inside DCs through use of optical technology opens a way to solve this problem. However, the current optical switching technologies are not able to support the dynamic DC traffic, and hence, new optical interconnect architectures are needed. Several optical switching architectures have been recently proposed to replace the aggregation and core tiers of current DC networks with a high-capacity optical switch [1][2]. However, the largest amount of energy in current DC networks is consumed in the edge tier by the electronic top-of-the-rack (ToR) switches. Therefore, the optical switching in the aggregation-core tier doesn't solve the problem. Moreover, the majority of the optical DC network architectures proposed so far can be categorized as optical circuit switching or optical packet switching. Unfortunately, there are certain limitations associated with these optical switching technologies for their application in DC networks. Namely, optical circuit switching architectures are not able handle the bursty and highly variable DC traffic while optical packet switching usually makes use of electronic buffers, which limit scalability and increase energy consumption. To reduce the energy consumption in the edge tier, we propose a novel optical broadcast-and-select architecture at the ToR. In this architecture, each server is equipped with an optical network interface (ONI) and is connected to the other servers in the same rack through a N×2 coupler, where N represents the number of servers in the rack. In addition, in order to provide both fine switching granularity and high scalability, we propose the use of the elastic optical networking paradigm [3]. Consequently, each ONI will be equipped with a bandwidth variable transceiver (BVT), which provides the ability to tune wavelength and change dynamically the number of the occupied spectral slots. In this way, the capacity can be varied from 1 Gb/s to 100 Gb/s and beyond on a per-server level. On the other side, the inter-rack communications are handled by a large singlesided optical core switch. One of the commercially available single-sided switches is fabricated using the beam steering technology [4], where the maximum number of switch ports available so far is 192 and a 500-port matrix is under development. Larger single-sided switches can be realized by combining several stages of smaller switch matrices. The results of a preliminary study show that the proposed architecture is able to significantly reduce the energy consumption with respect to other solutions [1][2].

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call