On Thu, May 11, 2023 at 7:32 AM Mark Tinka <mark@tinka.africa> wrote:
On 5/11/23 13:25, Vasilenko Eduard via NANOG wrote:
Sandvine is not representative of global traffic because DPI is installed mostly for Mobiles. But Mobile subscriber is 10x less than fixed on traffic – it is not the biggest source. Moreover, Mobiles would look better growing because the limiting factor was on technology (5G proposed more than 4G, 4G proposed much more than 3G) – it would probably would less disruptive in the future.
Fixed Carriers do not pay DPI premiums. And rarely share their traffic publicly. Sandvine could not see it.
And QUIC is making the the DPI model so loved by MNO's quickly irrelevant.
<plug> What we have been doing with libreqos is add in FQ + AQM (sch_cake), while pioneering some nifty new eBPF traffic monitoring abilities like being able to do "mice and elephants" plots, and more importantly sample TCP RTT statistics using a variant of kathie nichols' "passive ping" (pping) utility. This latter feature is proving very useful in diagnosing problems deeper in the network. Quic does respond to good ole queuing delays, packet drop, and in some cases ecn, but I am going to miss pping as it deploys more. It is the FQ that helps the most on videoconferencing and voip traffic. Libreqos is free software, working as a bridge, you can plug it in between any two points on your network, and on cheap (350 bucks off of ebay) xeon gold hardware easily cracks 25Gbits while shaping with a goal of cracking 100Gbits one day soon. demo here: https://payne.taht.net , running a variety of flent based stress tests, tcp up and downloads, the rrul test, etc. The interesting part to me, is that *real* traffic actually looks nothing like that on a per subscriber basis. Here´s a screen-movie example of real-world netflix queuing depth and delay: https://www.youtube.com/watch?v=C-2oSBr2200 Our new "piano roll" mice and elephants plotter is pretty neat, if I can say so myself. ... What we see from those ISPs that have shared their data from their Libreqos deployment is that the principal usage of major bandwidth is movie streaming, and there is essentially no difference in average usage between a 50Mbit residential plan and a gbit plan and only barely discernible at 25 vs 50. Given how new this codebase is we do not have any long term statistics for traffic growth or decline, or other patterns. But I kind of hope more folk here fire it up in their labs, at least. Please drop in on our chat channel #libreqos:matrix.org if you need help getting it setup. Sampling as we can at 10ms is a very different view of the net from 5 minute averages. </plug> Specifically plugging because I would like to understand CDN traffic patterns better, and do not have much data as yet on that side, collected this way.
Mark.
-- Podcast: https://www.linkedin.com/feed/update/urn:li:activity:7058793910227111937/ Dave Täht CSO, LibreQos