Where is the deck ?

The speaker achieves full throughput at a Gbit and also claims there is no packet loss at that speed. There is. Or there is nearly infinite delay in the stack. Loss is a required signal for that for cubic. BBR on the other hand estimates round trip delay.

An insightful take on Intent-Based Networking (IBN) from the perspective of automation. Great work, James!

great presentation Naveen. Well done.

Good presentation, but he is wildly out of touch about setting up a virtual data center deployment. Virtualization has never been easier with tools like containerlab. I don't know anyone still buying Cisco 1921s and Cat 3750s on eBay to build a lab anymore..

very informative

Excellent summary and very entertaining.

Python is so unstructured I did not able to learn it. It always derail my basic c# programming skills.

King 🔥

Good Stuff Kam. Good stuff. You nailed it.

Thanks for the overview between the languages!

I've encountered stuck routes before but never realized what was happening until now.

Very interesting presentation on current optical communication technologies and their real world applications. Just a small remark, at 0:9:44 the speaker probably meant to say Terahertz (THz) in place of Gigahertz (GHz). As Channel13 = 191.3 THz, in DWDM ITU-T grid channels

Good stuff

Brilliant presentationn. Thank you so much.

In case of Single Homed Hosts in Fabric A , Spines acting as data center gateways would also rewrite tge ESI values?

Danke!!!

What a great video, I am blasted. Thank you Kam I should revise some topics

Thank you very much for sharing this useful information.

Where can we get the slides?

Jason's surname is Livingood.

I think it might be better terminology to use term "fronthaul" for in premise "transit" networks which are from the OSP demarc (e.g. Cable MODEM) to the per room radio heads. Backhual is too confusing here, at least to me. Then there are three fronthaul choices in priority order, fiber fronthaul, copper fronthaul and wireless fronthaul. Copper & ethernet is going got top out at 10Gb/s. The signal loss of 100Gb/s over copper is in dB per inch, roughly 4 meters for a passive cable. Signal loss over fiber is .5db per kilometer and is independent of modulation and will scale to petabits/sec. So, make sure while you're installing communications fronthaul conduits to consider future fiber pulls too. The per room location doesn't really need a wall outlets as the fiber can be connected to wireless radio in the ceiling. It's sorta analogous to irrigation systems where the physical aspects of the landscape and plants should influence both the sprinkler heads placements and how the PVC is laid and zoned. Placing the radios (and antennas) properly and using a proper fronthaul design will help a lot.

How to handle lots of flows through a queue? Focus on the elephants? kzfaq.info/get/bejne/j6-DqM2KyNTGg30.html

I love the candid style and truth!

Great talk, as always. The linearity of IPv6 growth isn't because of unhealthy (lack of) growth; it's because we are in the linear portion of the logistic growth. We are past the point of the exponential growth. You're absolutely correct that most traffic for most hosts most of the time isn't going across the planet. But we still have the long tail ... and we must preserve the long tail to fight an overly monopolistic nature of the net

Promo'SM

IPv6 has the 24 bit flow label field for this exact reason.

Excellent speaker with fantastic info! Thank you Geoff Huston and NANOG!

I'm curious how NVLink and/or XGMI fits into your DC architecture. Can you comment?

Great talk! It's inspiring me to re-learn some Go since I haven't touched it in about six years.

Very interesting! Thanks!

If you use IPv6 you can add entropy in the src/dest addresses

How many customers will KEA serve? Will the server cope with ISP clients for example 300k?

Shame NANOG for allowing this silly cryptocurrency-promotion. None of what is proposed here is of any use. In a "permissioned" system a blockchain is at best a detail, and likely a very poorly performing database. Just use postgres! In a "non-permissioned" one you need a crytpo coin (to fund it / pay miners), and with proof-of-waste you're pouring petrol on our climate woes. Below vid is well worth a watch if anyone doubts. Sigh. kzfaq.info/get/bejne/qtmgermT0rPRhnU.html

❤

Terrible presentation. I agree with the Q+A guys: software should validate, run sanity check and determine the chance of success like if all gRPC/gNOI calls are working and their interface versions, what is really supported. Smirk on his face - just ugh Google.

Great presentation! Thank you to Leon Adato, hope to see more of him

thanks God this was not a sales presentation… what a negative person…. “ you are going to hate this…” 🤦🏻‍♂️

👍

Okay, now I just realized I know nothing about networks. I have an interview tomorrow for the position of a Network Systems Engineer though I am just a newbie in this field. I applied for Network Support Technician and I was called to be interviewed for Network Systems Engineer. I'm nervous, but I'll see what I can do in 24 hours. I pray to succeed. Fingers crossed

gNMIC is written in Golang

Right, it has been a decade... I am still getting to grips with how the time flew by. Both networking and software engineering worlds have come such a long way in that time. This presentation sums up much of the learnings that we have come to in that period, thank you so very much for that. The second Q+A question is excellent in so many ways. Yes, the promise of all interfaces being data-modeled is very much an idealization of the real world. On the other hand, in order to make a dent in the status quo, which OpenConfig certain has, you need to be bright-eyed. On the point of "still building HTTP REST interfaces", this is where the choice of YANG does not seem as such a bad idea: what is wrong with using YANG 'rpc's and 'action's to define those interfaces and use RESTCONF to automatically generate their projection into the HTTP/REST world?

Thanks Chris, great talk, hope to see more like it. another big advantage of Clos fabrics not transiting traffic through edge nodes is that you really separate the failure domain of the edge nodes from the fabric/transit nodes. would be cool to see more about modern BGP neighbor autodiscovery extensions for fabrics, & even RIFT as a new fabric routing protocol there's also a fascinating "Spineless Data Centers" paper (that I can't link to in YT comments) by Harsh, Jyothi, & Godfrey on non-Clos DRing designs that are designed for massively distributed server-server compute flows within the datacenter, vs. folded Clos which is better suited to egressing traffic out of the fabric via a low number of border routers.

Merci d'avoir pronouncé "Clos" correctement Chris!

Is Automattic ever going to implement IPv6?

Wow that's interesting. This could also make p2p video calls with large audiences possible.

very well done! You won another subscriber with this panel :)

Great talk, thank you!

Again honored to share the stage with a bunch of very smart fellow network automation engineers who were able to help compare and contrast the various use cases, pros and cons for Python and Go. Also don't forget to check out the GoLang tutorial that Dan presented at N90 and the more in depth break down for Go, Python and Rust that Clause presented also at N90!

Thanks for putting Go on the network engineer's map, Daniel!