Facebook announces its open data-center switch
By Stephen Lawson, IDG News Service | June 18th, 2014
The company wants to open up the platform for development and use by others.
Facebook has taken networking into its own hands, building a switch to link servers inside its data centers, and wants to make the platform available to others.
The sprawling social network provider needs maximum flexibility to keep up with user growth and roll out new services, said Jay Parikh, the company’s vice president of infrastructure engineering. He announced the switch, code-named Wedge, at the Gigaom Structure conference in San Francisco. Facebook had said in May of last year that it was working on a switch with the Open Compute Project (OCP).
The Wedge switch is being tested in Facebook’s production environment and the company hopes to have it working in full production mode and released through OCP by the end of this year, said Najam Ahmad, Facebook’s director of network engineering.
Networking is the final element of its infrastructure that Facebook is tackling by itself and through OCP, having already built servers and storage. Though few companies have the resources available within Facebook, ultimately making the switch design available through OCP could rock a networking industry that is still dominated by Cisco and a few other makers of specialized gear.
The Wedge switch is a “top of rack” switch, the kind built into each rack of servers in a data center to link those servers to the rest of the infrastructure. It has 16 40-Gigabit Ethernet ports and can be expanded to 32 ports, Parikh said. Facebook built its own chassis — painted Facebook blue — that is optimized for cooling and can have dual AC or DC power supplies.
A commercially available networking ASIC (application-specific integrated circuit) powers the switching functions. But the heart of the platform is a micro server based on the Open Compute Project’s “Group Hug” design, Parikh said.
“This is no longer a switch … it is just another server,” Parikh said.
Separating the software and the hardware of networking platforms is one of the principles behind software-defined networking (SDN), which is being implemented across the networking industry in various standards-based and proprietary ways. Facebook went beyond that concept, separating the various components of the software as well as of the hardware in the switch. That gives developers more freedom to innovate in the future, Parikh said.
There won’t be a software equivalent to the Wedge switch, Facebook’s Ahmad said. Most of the company’s networking software is internally developed for its own use, though it will open source some components. On Wednesday, the company also announced its Linux-based software for the switch, code-named FBOSS, but that’s not intended to run on the switch in outside implementations.
“The idea is that you can build your own software distribution and run it,” Ahmad said. By open sourcing the hardware, Facebook also will allow companies such as SDN startup Cumulus Networks to put their code on Wedge-based switches, Ahmad said. Facebook wants to foster an ecosystem of many companies offering either software or hardware for networks, and enterprises it has met with through OCP are very interested in that approach, he said.
When the OCP switch project was announced last year, Intel, Broadcom, VMware and Cumulus were among the participants.
Facebook wants to free both its own engineers and others to innovate more quickly on networks to keep up with changing requirements, Parikh said. The company not only wants to reach billions more consumers who don’t yet have Internet access, but also is expanding its focus beyond its traditional platform and into additional services such as Instagram, Messenger and WhatsApp, he said.
“We can’t let our business slow down because we can’t find the stuff that we need,” Parikh said.
The company wants more control over its network in two areas. For its connections to Internet service providers around the world, it wants a more flexible way to find and use the best route. For the so-called “east-west” traffic among its own servers, Facebook wants more control over how data moves around based on the needs of particular applications, Ahmad said.
Traditional networking gear doesn’t do that as well as Facebook wants, and to get new capabilities, the company has to wait for the vendor to implement them, he said. “That is very limiting in terms of what you can do,” Ahmad said.
Top-of-rack switches were the easiest place to start, but Facebook thinks that in time, the same programmable approach will apply all the way into the core of a wide-area network, Ahmad said. The company’s also testing an “edge fabric” for the meeting point of Facebook’s network and ISPs, using a central controller instead of traditional routers to determine routes, he said. That has allowed the company to boost the utilization of its network resources to more than 90 percent and eliminated backlogs in delivering Facebook traffic, the company says.