Open networks and non-OS switches

    How is the network market different from the server? There is a choice in the server market. A server from a global brand, a server from a local brand, a server from a platform manufacturer, a self-assembled server - there is an option for every taste and color.

    The network market (we are talking about enterprise) is much more conservative. There are a couple of dominant companies, there are all the rest. Each vendor has its own OS, its own set of dearly beloved and cherished bugs, the usual methods of circumventing them, its own command line (although some are combined with a single CLI). Everything is mine.

    Some time ago, there was an opinion that the market lacked Linux to accelerate development.





    The server can be purchased with or without an OS, you can choose a configuration for the requirements of applications, you can choose any distribution of the OS or build your own. The choice is rich, and this is what allowed us to develop a new business model, which formed the basis for the success of many Internet companies. Many developments in the software field appeared, the Open Compute project was launched to optimize server hardware.

    And what happened in the network hardware? How were switches managed nearly 20 years ago and how are they now? Nothing has changed




    Years go by, but no changes have taken place. Yes, now the movement towards OpenFlow and SDN has begun, but network vendors continue to offer closed options under the name of open ones. Many vendors develop their Linux-based operating systems, but no one uses just Linux.

    In 2010, a pair of startups, BigSwitch Networks and Cumulus Networks, were formed, who were engaged in the development of such a solution. In 2012, Pica8 joined them. Network OS Obvious difficulties along this path are the need to create an image for a specific switch and complex installation methods, because each switch can have its own bootloader. Complex installation methods To circumvent this limitation, the Open Network Installation Environment (ONIE) was developed








    . In fact, this is a pre-installed micro Linux, which, at the first boot, is looking for a source for installing a full-fledged network OS (NOS). The advantages of this approach:




    1. Environment for installing any network OS.
    2. Automation of switch management.
    3. Managing switches as Linux servers.


    Switch compatibility

    The second important modern trend is the publication of the SDK for working with matrices that were previously marked as 3rd party closed. Mellanox Open Ethernet , Broadcom OF-DPA SDK , Intel is threatening to publish its ONS and tools in the source.

    Distribution to the masses

    It was not without the ubiquitous Facebook, ONIE was given to the Open Compute project to develop a truly open network infrastructure. The principles do not allow Facebook to pay traditional vendors. :)

    Several companies have already made switches and published specifications on the Networking / SpecsAndDesigns website. The

    software part is also not forgotten: General idea




    As part of OCP, the Open Network Linux project was launched , on the basis of which it is proposed to build ready-made distributions for various switches. It's funny that ONL does not directly support packet processing, since it does not initialize the matrix. Only the control part:



    However, publishing the SDK allows you to build the distribution yourself. Just add the packages you need.

    As a result, they want to make a server with matrix drivers from the switch: Switch of the Future Choosing Control Plane






    A large part of modern switches work on PowerPC processors, some work on x86, someone likes MIPS or ARM. In some developments, it is already proposed to place the processor on a separate card, so that you can choose between processors, NOS distributions and available applications.

    The future of

    Facebook does not sit still and has already proposed the concept of a modular switch , calling it OCP Stage II.



    The Control Plane is based on the Group Hug family of microservers (Intel, AMD, ARM to choose from), which allowed integrating the switch management into a single FB system. In addition to managing the OS directly, it became possible to monitor the hardware status of the switch in the same way as the server. Temperature, voltage, condition of fans - everything is seen in the usual way. In general, while some (let’s not point the finger) invent complex switch and server management systems at the same time, others simplify switches to the server level.

    The software part, called FBOSS, is unified with the servers. Matrix management goes through the level of abstraction, the network can be managed as another service. All admins suddenly became network admins :)



    This example is likely to be picked up by manufacturers and we will see more modular designs in the future.

    Our participation

    We decided not to stand aside and made a number of models available as a Bare Metal Switch with a pre-installed ONIE environment.

    The list includes:

    Eos 220 - 48x 1G Base-T, 4x 10G SFP +, dual power supply. Cumulus Linux, Open Network Linux, BigSwitch Switch Light, Pica8 Eos 400 - 48x 10G SFP +, 4x 40G QSFP +, dual power supply are supported . Cumulus Linux, Open Network Linux, BigSwitch Switch Light, Pica8 are supported. Eos 420 - 48x 10G SFP +, 6x 40G QSFP +, dual power supply. Cumulus Linux, Open Network Linux, BigSwitch Switch Light are supported. Eos 520











    - 32x 40G QSFP +, dual power supply. Cumulus Linux is supported. And that's not counting the Eos 410i switch - 48x 10G SFP +, 4x 40G QSFP +, dual power supply. Only Intel ONS is supported , but an extremely flexible matrix. We hope that soon we will be able to announce the availability of sources and tools.






    Also popular now: