Open Networking: A Progress Report

When open networking switches were introduced in 2013, they promised freedom and flexibility for data centers by eliminating proprietary software and enabling greater return on investment. The concept was that in a high-speed, data-intensive world, network scalability and flexibility are critical to ensure that adding new services and hardware is handled as efficiently as possible without compromising network performance, and to empower customers with the ability to tune their networks to their specific business needs. The idea is that traditional closed-code Ethernet switches, delivered on proprietary equipment, limit the ability of data centers managers to develop an environment that can adapt simply and efficiently to the latest market demands.

Open networking was introduced to allow users to gain control of their network with the promise that adding differentiation into data center hardware would enable customers to realize competitive advantages in higher utilization, efficiency, and return on investment in their networking infrastructure.

Three years later, it's worth asking whether open networking has achieved the benefits it claimed and whether it's on the right path toward truly open Ethernet switching.

The simple answer is that open networking has, in fact, made huge strides toward enabling complete flexibility of choice with regard to switch and software vendors. Two models have since emerged beneath the open networking umbrella, offering different routes to achieve separation of software and hardware.

Opening the switch

The first model is the introduction of switch systems that are not locked in to a specific software vendor. In essence, where proprietary switches once dominated, forcing customers to also accept the associated software on top of the hardware, today’s open networking switches are white boxes on which users can install various software options. The drivers and the control are supplied with the white box and the OpenAPI is exposed to allow an interface into any open networking protocol or application -- for example, the Quagga routing suite, the OpenFlow agent, or even vendor-developed apps.

open.jpg

Open networking switches are designed to comply with all available operating systems, whether native Linux, Cumulus Linux, or MLNX-OS. Moreover, when new operating systems like Microsoft SONiC are available, open networking ensures that the switch system can be easily transformed from one software package to another. The installation of such software is enabled via the Open Network Install Environment (ONIE), which was developed by Cumulus Networks and made open source. ONIE comes with the switch system specifically in order to enable the use of any OS.

Compare this to the closed-system vendors such as Arista, Cisco, HP, Extreme, and Brocade, and it becomes clear that their traditional solutions are monolithic, as they only offer their own OS as a software option.

Adding an abstraction layer

There is a second model that also provides freedom and flexibility for data centers. Open networking also supports open hardware that permits an abstraction layer which separates the switch and the operating system. The abstraction layer allows the software that sits on top of it to be used transparently on different types of hardware. That means that if a network deployment uses various types of hardware, they can all be used with the exact same software management, configuration, and monitoring tools. Moreover, any hardware upgrade does not require a change to the software.

There are two primary abstraction layers currently in use: Switch Abstraction Layer (SAI) and switchdev. Each has its supporters and each has its advantages.

SAI is a userspace application of the abstraction layer. The application drives the switch ASIC, bypassing the kernel, which means that there is no need for Linux kernel expertise. It's even possible to change the hardware simply by changing the SAI driver. The primary supporter of SAI has been Open Switch, which was developed by HP and is supported by OCP.

Switchdev, on the other hand, is a kernel implementation of the abstraction layer. The switch state is kept in the kernel instead of an application, meaning that there is no need to rewrite existing Linux tools or to create new ones. In fact, with switchdev, the switch becomes like any other server in terms of management, significantly reducing management and IT costs.

Eliminating lock-in

Networking customers have been locked into vendor proprietary hardware and software solutions for too long. The market for switches and routers has been ruled by a few select vendors, resulting in limited solution flexibility and high acquisition and service costs.

Until the advent of open networking, there was no alternative to buying networking solutions from the few vendors that have dictated the market. This vendor lock-in has been eliminated with the opening of the switch platform and with abstraction layers that allow customers to choose their software independent of their hardware.

Open networking provides data centers of all sizes with the flexibility and freedom to use any software on top of their Ethernet switches, generating optimal performance and scalable solutions that will help them to adapt and thrive in today’s high speed interconnected world.