Datacenter optics : requirements , technologies , and

Over the past decade, datacenters and their networks have become technology enablers for a number of internetbased applications. As of today, most of the popular internet applications, from the traditional search, online interactive maps, social networks, video streaming, and the Internet of Things, are running in datacenters. The pivotal role played by the datacenter will be further heightened by a wider adoption of cloud computing, where a significant portion of computing and storage is migrating into shared datacenters. This is already occurring at a rapid pace today with a number of large cloud providers leading the way. This has resulted in a dramatic increase of datacenter capability. For example, the bisection bandwidth (BW) of Google’s datacenter cluster networks have increased by a factor of one thousand over the past decade. A datacenter is essentially a massively parallel supercomputing infrastructure, consisting of clusters with several thousands of servers interconnected together in each cluster (Fig. 1). The early small-scale datacenter networks used copper-based interconnects, but as BW requirements scaled, fiber-optics-based optical interconnects were introduced and quickly established as the most cost-effective, deployable solution to scale out the datacenter. As of today, high BW and low power optical interconnects have been ubiquitously used in datacenters for any link distance beyond a few meters. This Review intends to review over a decade of technology evolution of datacenter optics. Emerging trends and technology options to scale BWs beyond 400 Gb/s will also be discussed. For a typical intra-datacenter network adopting a Clos topology (Fig. 1), a massive number of interconnection links are required to implement the large fan out and corresponding high bisection BW. Thus, the number one consideration for interconnect is the BW cost. To minimize the total interconnection cost, different technologies are adopted at different segments of the network. For example, a backplane printed circuit board (PCB) and copper interconnects are typically used for intra-rack interconnection for a reach of less than a few meters, while fiber-based optical interconnects are used for interconnection between the top of rack (TOR) switch and the edge switch, as well as the between the edge aggregation switch and the spine switch, with a link distance ranging from a few meters up to 2 km. For a link distance of less than 100 m, vertical cavity surface emitting laser (VCSEL) and multimode fiber (MMF)-based technologies have proven to give the best overall link cost (transceiver cost plus fiber cost). Beyond 100 m, however, more expensive single mode fiber (SMF) transmission technologies usually have to be chosen due to the following reasons: (1) the BW of commercially available VCSELs have been limited (about 20 GHz as of today), (2) the BW of an MMF reduces as the distance increases [for example, the BW is limited to be about 20 GHz for 100m of optical multimode (OM3) fiber], and (3) the higher cost of an MMF, although the cost of a VCSEL transceiver is significantly lower than an edge emitting laser-based SMF transceiver. The second important criteriom is power consumption. From an aggregate energy consumption point of view, the power consumption of networking is only a modest piece of the power consumed by a datacenter (less than 10%). But, the power efficiency of optical transceivers is essential for the front panel density (the allowable transceiver size is largely determined by its power envelope). Hence, without power efficient transceivers, there is no optimal way to take advantage of the full capacity of the switch application-specific integrated circuit (ASIC). The third important criterion is serviceability. Since the reliability of the typically used active optical components are not very high, it is better to design the optical transceiver in a way that it can be easily replaced or serviced. In this regard, pluggable optics is preferred over on-board optics, although on-board optics enables a higher front panel density. Due to the need of serviceability, we expect pluggable optics will continue to be the mainstream switch-to-switch optical transceiver choice for the COL 15(5), 120008(2017) CHINESE OPTICS LETTERS May 10, 2017