LinkedIn has launched a standards effort to support webscale data centers - highly standardized facilities which must be able to grow massively whiile still remaining responsive.
The proposed OpenFabric standard has emerged from the social media company’s efforts to develop networks for its own needs. These resulted in its Project Altair hyperscale architecture (described in a LinkedIn blog post), and its own switches including a design called Pigeon. Now LinkedIn is sharing the protocols developed in its hyperscale architecture, it was announced at the DCD>Middle East last week.
New webscale protocols
”As we scale our data center infrastructure we felt the need to build, grow, optimize and manage our data center fabric in a more effective and disruptive way around specific requirements and functions,” principal network architect Shawn Zandi said in an exchange with DCD. General purpose networks would not do this, so LinkedIn developed its own.
Conventional IP networks, including the Internet, were designed to link up diverse autonomous systems, which might have different functions and policies. Data centers on the other hand are managed centrally, and can be built with identical hardware throughout, allowing for simplification, said Zandi:
“We had different objectives and felt relying on conventional control planes brings limitations that are not necessarily required anymore for today’s networks,” he told DCD. ”We also need a self-defined network that can be built automatically without a need for configuration and to some extent self-heal without a dedicated operations team to punch commands and troubleshoot basic operations of the network.”
LinkedIn redeveloped its data center networks from physical to application layer, in Project Altair: ”It involved changing our physical topology and hardware, bandwidth and latency as well as load-balancing and our security model around containerization and LPS (LinkedIn Platform as a Service),” said Zandi.
LinkedIn’s data centers, in particular its flagship facility in Oregon which opened in 2016, are now built with a spine-and leaf topology.
Altair provided LinkedIn with a disaggregated model, and OpenFabric is a layer on that, which creates an “autonomous, self-defined and self-healing network that can be optimized the way it was not possible before.”
LinkedIn is publishing drafts of the basic concepts of OpenFabric, through the Internet Engineering Task Force (IETF) the organization that is responsible for Internet standards including TCP/IP, so that the entire community can both enable, and be enabled by, this work. OpenFabric includes a modification of the IETF’s long-established IS-IS protocol, designed to better suit the needs of big spine-and-leaf - or webscale - data centers.
This new protocol is independent of LinkedIn’s Altair designs and hardware, and is available for anyone to implement or use on their own webscale hardware, but it’s not the end of the OpenFabric story, said Zandi: “That proposal is just one piece of the OpenFabric project.”