The issue of intra-point-of-presence (PoP) traffic flows is a real hot potato. Literally.
Internet Service Providers (ISPs) with infrastructures architected toward north-south traffic flows are increasingly seeing patterns shift toward east-west. Driven by a dramatic increase of streaming content, ISPs’ access infrastructures are being choked with third-party content, leaving these carriers with the burden of saturated bandwidth and congested edge switches, along with the need to fund expensive network upgrades without a discernible return on investment. The evolution of software-defined networking (SDN) and the introduction of white boxes (or generic, commodity OpenFlow switches), however, are now offering operators inexpensive architectural alternatives.
Targeted toward the mass market, on-demand video services such as YouTube, Netflix and Hulu, along with numerous CDN players, are often attracting multiple simultaneous viewers per household, with each individual accessing distinct content served by separate providers. Extend this trend to even a small geographic area and the consequences can be catastrophic. Even while applying “hot potato” routing philosophies — the practice of dumping traffic to the target service provider’s ISP or network as soon as possible — serving ISPs are still saddled with transport and switching costs at the access layers until the packets in question reach, and ultimately hairpin into and out of, provider edge (PE) routers. This is true even when an “eyeball network,” as they are known, has its content stored, cached and distributed locally in the same geo-region as the serving ISP.
With some models suggesting up to 60 percent of traffic is unnecessarily traversing PE routers, it is now becoming apparent that network operators can realize 60 percent cost savings, or more, by adopting alternative approaches to handling east-west IP packet flows.
Consider an example: A Netflix subscriber and a customer of ISP-A in Los Angeles selects a documentary to watch. The request to establish a video stream is handed off at ISP-A’s L.A. PoP, like that aforementioned hot potato, to ISP-B — the ISP of Netflix hosting partner Amazon Web Services (AWS). This connection ultimately serves the AWS us-west-2 data center in Boardman, Ore., by way of a quick stop in Seattle. With ISP-B participating in the same scorching spud-like shenanigans, initial responses are peeled off at their PoP in Seattle and forwarded on to the ISP-A’s core network by way of their local PE router.
Eliminating a torrent of traffic extending across ISP-A’s core, however, a Netflix content distribution network (CDN) partner, ISP-C in this example, is engaged and a server in neighboring Orange County, California, starts streaming the end user’s show to their target device. While these east-west packets are, for all intents and purposes, using no bandwidth on the ISP-A’s core network itself, they must traverse (hairpin) an IP-aware device capable of eBGP peering between Autonomous Systems and are therefore consuming valuable port, processing and switching fabric resources on the carrier’s provider edge (PE) router.
In the current architecture, the PE router provides two key functions for CE traffic flows, namely traffic and route aggregation, or default routing. Reducing capital expenditures at PoP locations, while facilitating an ever-increasing need to serve high-bandwidth interconnect traffic, demands an intelligent approach to edge aggregation. High-quality, commodity Ethernet switches in the form of white boxes are now being considered viable alternatives to expensive and complex edge routing resources. Such switching fabrics can be used to directly connect peering ISP CE routers, lifting the burden of supporting east-west traffic flows from the serving ISPs’ edge routing functions.
To achieve this, however, these new switches must be capable of performing potentially complex flow aggregation and path selection at the IP layer (Layer 3), handling the majority of CE-bound traffic while leaving more complex forwarding decisions — or north-south packets — to the existing routing platforms. With a cost-effectiveness that comes at a price, however, these devices must operate without the overhead of a fully functional PE router.
Forming a new aggregation layer between the access and edge infrastructure, commodity switches can be employed at interconnect points between alternative ISP CE routers and the host ISP’s PE routers. Deployed without an Ethernet control plane, thereby avoiding instabilities that have plagued carrier-scale Ethernet implementations, these switches will instead be managed by a centralized or hierarchical SDN controller. Countering the classic switch-centric view of legacy Ethernet control planes, the SDN controller can operate at the IP layer and with a global, loop-free, topological view of the entire network.
When architected for elastic compute environments, the SDN controller is almost infinitely scalable, enabling it to achieve on its own the route processing and path computation typically performed by numerous individual network elements. Establishing itself as an internal or external BGP peer with PE and CE routers, by virtue of a standard route reflector function, the controller can learn the best path for each destination in the network without announcing routes or participating in the packet-forwarding process. The controller is simply a routing information base (RIB) repository for the distributed array of switches.
While the switches themselves must hold the forwarding information base (FIB), the size of this table is dramatically smaller than that of the controller’s RIB or the northbound PE router’s RIB/FIB. Indeed, the abundant amount of ternary content-addressable memory (TCAM) materializing in these white boxes is more than adequate to support the data-plane connectivity needs of this aggregation layer. Populating the forwarding tables in the switches from the route discovery mechanism operating in the controller can be achieved by a number of standardized (or standardizing) protocols, such as OpenFlow, ForCES, NetConf or I2RS.
Replacing the classic CE-PE path with one purely in the CE domain, this new network architecture achieves the goals of replicating PE functionality in terms of route and traffic aggregation, without the cost and complexity typically associated with port utilization and packet processing on these devices. Furthermore, despite the SDN controller and additional switches introducing more physical components and functional elements into the overall infrastructure, the configuration load of the network itself is actually reduced and even simplified through centralization.
The promise of programmability must start with a problem. Changes in the traffic patterns in large carrier networks have challenged existing planning models, design rules and their accompanying economics, finally resulting in the need to examine the benefits of software-defined networking in carrier infrastructures. Infrastructure and Internet service providers are holding another hot potato, only this time they can’t afford to pass on it.