As IP networks spread, the design challenge for networking SoCs evolves

Perhaps the most pervasive trend in networking today is the seemingly complete victory of IP over rival protocols, and the concomitant spread of gigabit Ethernet. Gigabit Ethernet is replacing other alternatives, even packetizing legacy dataflows and carrying them as encapsulated IP traffic. Along with carrying the traffic, the new fast networks are also having to provide the application-specific services that older networks had evolved. This trend is drawing together disparate networks into a single medium, vastly simplifying facilities management for network operators, but creating a pain for silicon designers.

The highest-profile example may be in data centers, where three different levels of physical networks—Infiniband for clustering, Ethernet for data, and Fibre Channel for storage—are being drawn into a single 10 Gbit Ethernet fabric. The convergence means that the single network must offer all the management and QoS services provided formerly by three very different kinds of networks. That demand, in turn, has implications for network adapter cards, classifiers, and fabrics throughout the network.

The same process seems to be uniting the various networks that serve users of the Internet. In the past, local networks for end-users, the aggregation networks, fiber access networks, and mobile backhaul networks all had very distinct requirements. But according to at least two semiconductor vendors, all of these networks may be converging.

"We are seeing a need for a common feature set at several different layers of the network," said Broadcom senior manager of product marketing for the service-provider segment Jim McKeon. McKeon said that the need for features including MPLS and time synchronization are cropping up across access, backhaul, and aggregation networks. Essentially, switches throughout the network are having to support carrier-class services. Broadcom’s response has been a family of families of Ethernet switch chips, ranging from a 112 Gb/s chip for access networks and aggregation switches to a low-end switch chip with integrated PHYs for multitenant consumer-level fast Ethernet.

Requiring all that functionality across such a range of price and performance points becomes a serious challenge for SoC designers. The high-end switches may require dedicated hardware and substantial memory to meet the demands of, for example, MPLS at wire speed. Yet low-end switches go into very cost-sensitive markets and can’t carry a silicon overhead for performance they don’t need. And Broadcom’s customers, who may be trying to serve all these markets at once, want a single API for all the chip families.

Fortunately for the chip designers, there is still a pattern in where functions occur in the network, according to Broadcom senior director of product marketing Eric Hayes. "At the edge, policy is user-based and generally oriented toward security concerns. As you move nearer the core, policy becomes more QoS-oriented. So while we have a uniform programming model across the families, the ability to apply policy varies from chip to chip."

Switch-fabric vendors aren’t the only ones to see a growing commonality across applications. Xelerated, which bills itself as one of the two surviving NPU vendors, has been very successful serving the mobile backhaul market. But according to the company’s vice president of marketing and business development Thomas Eklund, the company is finding fiber access applications increasingly attractive. Ironically, he says, both markets are facing similar QoS issues, but from different sources.

In the backhaul space, where the networks were often ATM-based and designed to handle latency-intolerant time-domain-multiplexed voice traffic, the challenge is to deal with the increased data services demanded by smart phones and promised by LTE. In fiber access networks, built for sporadic and latency-insensitive Web browsing, the problem is IPTV. Both applications now seem to be converging on similar needs for management and QoS, and those needs seem to be looking a lot like Carrier Ethernet specifications.

There is a parallel between the two convergences, in data centers and in the global network. Perhaps in the long run this foretells an evolution to a single network architecture in which computing and storage resources can be distributed throughout a single network that serves as the molecular bonding of the cloud, as access, backhaul, aggregation, and core. But one suspects there are lots of committee meetings to be survived before that day. In the meantime, SoC designers will be busy enough figuring out how to assemble the right IP and lay out the optimum family to spread a new level of classification and routing services across a lot of legacy applications.