Barefoot "Tofino": 6.4 Tbps whitebox switch silicon?
a lot of PR fluff, but this may be of interest: http://www.wired.com/2016/06/barefoot-networks-new-chips-will-transform-tech... https://barefootnetworks.com/media/white_papers/Barefoot-Worlds-Fastest-Most... Based on their investors, could have interesting results for much lower cost 100GbE whitebox switches.
On Wednesday, June 15, 2016, Eric Kuhnke <eric.kuhnke@gmail.com> wrote:
a lot of PR fluff, but this may be of interest:
http://www.wired.com/2016/06/barefoot-networks-new-chips-will-transform-tech...
https://barefootnetworks.com/media/white_papers/Barefoot-Worlds-Fastest-Most...
Based on their investors, could have interesting results for much lower cost 100GbE whitebox switches.
Where is the price tag? Why would you think it is inexpensive? I do think p4 is very interesting, but is it really much different from openflow? Which...umm ... Did not succeed in the market.
On Wed, Jun 15, 2016 at 8:31 PM, Ca By <cb.list6@gmail.com> wrote:
On Wednesday, June 15, 2016, Eric Kuhnke <eric.kuhnke@gmail.com> wrote:
a lot of PR fluff, but this may be of interest:
http://www.wired.com/2016/06/barefoot-networks-new-chips-will-transform-tech...
https://barefootnetworks.com/media/white_papers/Barefoot-Worlds-Fastest-Most...
Based on their investors, could have interesting results for much lower cost 100GbE whitebox switches.
Where is the price tag? Why would you think it is inexpensive?
I do think p4 is very interesting, but is it really much different from openflow?
Great question! OpenFlow enabled decoupling of the control plane from the data plane, whereas P4 allows you to define the data plane. Check out this popular P4 program that defines the data plane of a L2/L3 switch - https://github.com/p4lang/switch/tree/master/p4src OpenFlow is a protocol, whereas P4 is a programming language. You can use OpenFlow controller to control a P4 data plane - Check out this OpenFlow agent on top of a P4 data plane - https://github.com/p4lang/p4ofagent Check out this blog post which explains the differences between P4 and OpenFlow quite well - http://p4.org/p4/clarifying-the-differences-between-p4-and-openflow/ More questions? Shoot me an email :-) Which...umm ... Did not succeed in the market.
I do think p4 is very interesting, but is it really much different from openflow? Which...umm ... Did not succeed in the market.
P4 is quite different from OpenFlow. The various generations of the OpenFlow specification assume a fixed-format packet header with an ever increasing number of header fields (from 12 in OF 1.0 to 41 in OF 1.4), a fixed set of packet-processing actions, and so on. P4 allows programmable parsing and flexible match-action processing. In P4, the program tells the data plane how to act; in OpenFlow, the way the data plane can act is “baked” in advance. So, OpenFlow 1.x can be one of many *programs* you can write in P4, whereas many other P4 programs can exist. For some discussion of the differences, see Clarifying the differences between P4 and OpenFlow http://p4.org/p4/clarifying-the-differences-between-p4-and-openflow/ <http://p4.org/p4/clarifying-the-differences-between-p4-and-openflow/> — Jen
Looks very promising! On Thu, Jun 16, 2016 at 6:21 AM, Eric Kuhnke <eric.kuhnke@gmail.com> wrote:
a lot of PR fluff, but this may be of interest:
http://www.wired.com/2016/06/barefoot-networks-new-chips-will-transform-tech...
https://barefootnetworks.com/media/white_papers/Barefoot-Worlds-Fastest-Most...
Based on their investors, could have interesting results for much lower cost 100GbE whitebox switches.
-- Sincerely yours, Pavel Odintsov
On 16 June 2016 at 06:21, Eric Kuhnke <eric.kuhnke@gmail.com> wrote:
Based on their investors, could have interesting results for much lower cost 100GbE whitebox switches.
Why lower cost? The BOM isn't the expensive part, the code is the expensive part. Only way I see this happening, is if we get open source routing suite for the box, i.e. 0 cost software. If you're thinking of writing your own routing suite, even if your requirements are trivial, it's still probably take 2-3 years and +2MUSD in salaries, and then maintenance +300kUSD/year in salaries. Need quite significant annual unit number scale to make it cheap. I'm quite fascinated by the idea of doing something really novel in routing suite space, but I don't see how it could possibly work commercially. How many customers would there be for licensing COTS routing-suite when costs are millions annually to develop it for general use-case. -- ++ytti
Saku, I agree completely. Isn't this what Arista did? They coded from like 2004 to 2008 before launching EOS using commercial chipsets. You seem to really understand routing software, so I would love to hear your take on Arista EOS. On Thu, Jun 16, 2016 at 3:19 AM, Saku Ytti <saku@ytti.fi> wrote:
On 16 June 2016 at 06:21, Eric Kuhnke <eric.kuhnke@gmail.com> wrote:
Based on their investors, could have interesting results for much lower cost 100GbE whitebox switches.
Why lower cost? The BOM isn't the expensive part, the code is the expensive part. Only way I see this happening, is if we get open source routing suite for the box, i.e. 0 cost software.
If you're thinking of writing your own routing suite, even if your requirements are trivial, it's still probably take 2-3 years and +2MUSD in salaries, and then maintenance +300kUSD/year in salaries. Need quite significant annual unit number scale to make it cheap.
I'm quite fascinated by the idea of doing something really novel in routing suite space, but I don't see how it could possibly work commercially. How many customers would there be for licensing COTS routing-suite when costs are millions annually to develop it for general use-case.
-- ++ytti
On Thu, Jun 16, 2016 at 1:19 AM, Saku Ytti <saku@ytti.fi> wrote:
On 16 June 2016 at 06:21, Eric Kuhnke <eric.kuhnke@gmail.com> wrote:
Based on their investors, could have interesting results for much lower cost 100GbE whitebox switches.
Why lower cost? The BOM isn't the expensive part, the code is the expensive part. Only way I see this happening, is if we get open source routing suite for the box, i.e. 0 cost software.
It shouldn't be long before we see open source routing suites (Bird, Quagga, GoBGP, etc) running on Linux (ONL, OS10, OpenSwitch). There is a P4 program that implements the Switch Abstraction Interface (SAI), which provides a common interface, device independent, interface to merchant silicon. http://p4.org/p4/an-open-source-p4-switch-with-sai-support/ A quick way to do interesting things with the programmable data plane is to use P4 to augment the basic switching / routing behavior provided by SAI: moving resources from layer 2 tables to layer 3 tables, adding telemetry, adding additional control capabilities, etc.: http://blog.sflow.com/2016/06/programmable-hardware-barefoot-networks.html
participants (8)
-
Ca By -
Colton Conor -
Eric Kuhnke -
Jennifer Rexford -
Pavel Odintsov -
Peter Phaal -
Prem Jonnalagadda -
Saku Ytti