New Moonshot talk accepted!
With the emergence of switchdev as the canonical NIC-switch
representation comes the view that we need to expose each switch port as
a singular netdev.
There is, however, an impedance mismatch between that model and a few
important use cases.
Take for example the need to scale performance of high speed ports;
where a reasonable approach is to spread a single NIC-switch port's
traffic across multiple PCIe devices. A single netdev representation
doesnt cut it for that scenario.
There are other use cases of which involve hierarchies of VMs/containers
where more of these issues emerge - for these use cases there have been
discussions to use 2 or 3 netdev layers; however, even there some
challenges emerge.
Anjali Singhai Jain and Sridhar Samudrala make the arguement that the
switchdev port representor as is is not the best fit for these use
cases.
In this talk they are putting forth a proposal that they feel will
cater to various use case needs while maintaining strong control over
the resources proposed by the switchdev model.
More info at:
https://www.netdevconf.org/0x12/session.html?composing-and-configuring-comp…
cheers,
jamal
New workshop accepted.
Roopa Prabhu will chair a discussion on Open NOS for ASICs.
Various topics on Linux support for switch ASICs will be discussed:
- New hardware support
- Hardware Resource management updates
- Scaling routing fib and bridge forwarding database
- Building Network virtualization solutions: E-VPN
- Network Configuration Management, Debugging and
troubleshooting
- Futures
More info at:
https://www.netdevconf.org/0x12/news.html?workshop-roopa-prabhu-open-linux-…
cheers,
jamal
We are pleased to announce our Netdev 0x12 keynote
speaker: Van Jacobson
For most people involved in the networking world, Van will
need very little introduction for his work, amongst many,
in TCP/IP network performance and scaling of the internet.
For the young lads amongst us who may be oblivious of his
accomplishments, Van is what you folks typically refer to
as a G.O.A.T(Greatest Of All Time) and some of us would
refer as a living legend[1]. A few of his accomplishments
are listed on his Internet Hall of Fame Induction bio:
https://www.internethalloffame.org/inductees/van-jacobson
This announcement space is too small for us to enumerate his many
accomplishments and current passions - but we wanted to highlight
one buzzword in particular:
You would not have EBPF today if Van was not there[2].
At the moment Van is involved, at Google, in a passion of
his - Network Performance Scaling.
Having Van at this conference is both a privilege and an honor as
he has seen a lot of ideas take wing from the early days of the
internet to the specialized roles around Data Center, Web, IOT and
more that we see today.
We wish to learn from this sage.
To paraphrase European Communication Magazine[3]:
"When Van Jacobson has something to say, people listen."
And Van has something he wants to say to us at Netdev conf 0x12:
We need to teach Network Interface Cards about time.
What kind of surgery do we need on the kernel NIC interfaces?
What kind of features do NIC vendors need to provide?
Come and listen to Van.
cheers,
jamal
[1]https://www.merriam-webster.com/dictionary/living%20legend
[2]http://www.tcpdump.org/papers/bpf-usenix93.pdf
[3]https://www.eurocomms.com/features/analysis/8238-could-content-centric-ne…
We are pleased to announce our first Platinum sponsor,
HAProxy Technologies!
Thank you HAProxy Technologies - your kindness is what
makes Netdev conf happen. We are grateful not just to your
sponsorship but also for all the contributions made by
your organization to keeping Linux Networking on the map!
More information:
https://www.netdevconf.org/0x12/news.html?platinum-sponsor-haproxy
cheers,
jamal
New accepted talk.
From the same folks at MIT who brought you the idea
of Congestion Manager (Linux being able to plugin
different congestion control algorithms) comes an
exciting(my emphasis!) idea to bring even more modularity
into Linux TCP. CM concepts that were harder to put into
the kernel are now possible.
Akshay Narayan et al discuss Congestion Control Plane (CCP).
CCP is a new way of separating sender side TCP into control
(sitting in user space) and datapath (sitting in the kernel).
Control state from the kernel is used by algorithms in user space.
User space algorithms use this information to control the
kernel’s congestion window or pacing rate.
The talk will describe the details of the design principles used,
kernel refactoring made, libccp which exposes user API, and
experimental results from the implementation.
More info at:
https://www.netdevconf.org/0x12/news.html?talk-akshay-narayan-et-al-restruc…
cheers,
jamal