Stanford University engineering professor Nick McKeown expects a new breed of network processors to replace the ASICs used in today’s routers and switches within the next decade. He says he has studied this future communications processor in depth: “And if you squint hard enough, it looks like a RISC processor for networking.”
McKeown is helping drive software-defined networking based on the OpenFlow protocol, with the goal of spawning a whole new set of software applications to manage a variety of simplified switches and routers. If his mission succeeds, it could make running large data centers and enterprise networks simpler and cheaper, and promises to upend the current business model of expensive networking equipment built on complex ASICs and proprietary code.
The new breed of commercial chips McKeown envisions would replace the ASIC solutions used today by vendors like Alcatel-Lucent, Cisco, Ericsson, and Juniper. He says the first such chips will arrive in the next two to three years.
Through collaboration with companies like TI, McKeown has prototyped the new class of component in a research paper. It essentially consists of a parsing engine that translates the ever-growing variety of headers in each data packet, then pushes the packet through a pipeline of execution units that match patterns within the headers and take actions.
“It is a forced match-action, match-action feed-forward pipeline,” McKeown said, noting that the related research paper is undergoing pre-publication review.
According to McKeown’s paper, by increasing chip area and power consumption by just 15 percent, the new chip can handle any of today’s or tomorrow’s communication protocols at the same performance level as current ASICs that only support specific protocols. He expects the large router and switch vendors to replace their ASICs with such chips within a decade and transform themselves into software companies.
“We will see results within ten years, where those vendors will provide control plane software and applications on top,” he noted. Already two or three companies are researching and developing such chips, including startup xPliant, as well as established players like TI and possibly Cavium or Mellanox.
“Commercial chips will be one of the main driving forces behind the OpenFlow project,” McKeown stated. “Existing chip suppliers including Broadcom and Marvell already have switch chips ready to add OpenFlow support — that is what they should do, and they have been involved from the very beginning.”
As the new generation of software and hardware emerges, technology evolution will drive an industry revolution. The current 1.x version of OpenFlow represents a compromise, McKeown pointed out: “Ideally we would have it start with a general-purpose match-action workflow, but we also had to map it onto existing chips — the next generation of technology will be much less protocol-dependent.”
Last year, the Open Networking Foundation (ONF), which oversees the OpenFlow protocol, was busy convening ASIC vendors to form a Forwarding Abstractions working group. The group’s purpose is to narrow the gap between what OpenFlow aims to achieve and the capabilities of existing and planned ASICs.
Now, ONF is setting up a chip advisory board. “We will learn from them what is possible in chip technology, and from there we can derive the possibilities for the next generation of OpenFlow,” McKeown said.
OpenFlow has begun using content-addressable memories as an intermediary for router and switch ASICs, though this approach has functional limitations. As a result, the standard recently started using a technique of matching multiple tables. “An OpenFlow version that supports only a single protocol is still a good while away,” McKeown stated.
On the software side, the code that will bring software-defined networking (SDN) to life is steadily emerging. Startups including Big Switch Networks and Nicira, which has been acquired by VMWare, already have their OpenFlow controller software. Earlier this month, 18 major telecom equipment vendors and software suppliers announced a collaboration on the Open Daylight project, aimed at building an open-source environment for SDN controllers, related APIs, and more.
Industry observers expect the major players to compete to have their code become part of the Open Daylight project. Once sufficiently reliable products hit the market, large companies like IBM will be willing to pay to adopt them and offer integrated services — though that will likely take another two or three years.
McKeown draws an analogy to Posix, the set of API standards that later became Linux. He notes that it took a decade for that operating system, which had different versions, to settle