dave.taht at gmail.com
Thu Oct 27 12:16:32 EDT 2022
On Thu, Oct 27, 2022 at 8:52 AM Zach Underwood via LibreQoS
<libreqos at lists.bufferbloat.net> wrote:
> Dave recommended that I introduce myself to the list.
Very nice to have you join us. There's been a lot of good traffic this
month in the archives. Feel free to weigh in on any subject!!!
> I am a wifi engineer at Whitesky which is owned by Realpage(yeah I know about the lawsuit, that is not my group). We provide wifi and wired internet services at multiple family housing units. We have 600 sites in the USA and service 135k units. With about 60% being student housing and avg install 1200 APs a week. We range from a 20 unit complex to 5000 unit housing like Harvard graduate housing. Our standard deployment is an example of a 200 unit complex. We will order and install a 10gb wave or EPL from say level3, atnt, uniti or another fiber company between the site and one of our 13 datacenters. In the datacenter we have IX connections, private peering and transit connections. The fiber will connected to a mikrotik crs317_1g_16s_rm to be used as a wan switch. Next we have a pair of ccr2004_1g_12s_2xs.
What version of routerOS? I'm always at pains to say fq_codel was
designed to run at line rate (cheaply), NOT shaped... and that the fq
helps smooth out microbursts, and the aqm is way better than a drop
tail queue, and am very enthused about the routeros 7.2 and later
availability of it, but still struggle to convince people to
"just turn it on" as we did in openwrt, linux, ios, etc so long ago.
> Half of the site vlans are on one router and the other half on the other router. All vlans on each router we NAT(for v4) to a single public ip(so two public ip per site). We also dual stack for v6 and offer dhcpv6-pd on the wired network. After the routers they connect to the nexus 3k core switch and then 10gb links to 3750x that power APs in unit and service wired drops. We do not rate limit or do any type of shaping.
This is one of those things that then, gets done naturally, at the
bottleneck point, which is most likely
the wifi, but might, in a complex campus, be at the up or downlinks. I
was deeply impressed by the ruckus spec
which transliterated reads almost like a direct implementation of the
stuff we described to them circa 2019? (2020?), in what I thought
(then) was a disastrous meeting with their (now ex) CEO. (I refer to
this story here: )
I'm dying to know what they actually implemented, notably if ECN is
present or not. a packet capture of a ping -f -Q 1 -s 1440 from client
to some other point ought (on a packet capture of the other point)
show the CE bits set if they actually directly implemented the stuff
from https://www.cs.kau.se/tohojo/airtime-fairness/ e.g. in wireshark:
ipv6.tclass.ecn == 3
ip.dsfield.ecn == 3
There's been a lot of research as to how much dscp or ecn "washing"
actually existed across the internet. Most of it seemed to be coming
from older, bigger ISPs.
testing udp or tcp diffserv and ecn compliance e2e is also possible
with the flent tool. Are the ecn bits preserved? (-Q 1 or -Q 2)
I use irtt as a means to sweep the diffserv bits.
>At all of our site we are doing a min of 1 ap in each unit. Some of our sites have paid extra and will get two ruckus r650 in say a 2 bedroom 2 story townhouse.
It is in general, my hope, more folk in dense deployments are using
> Zach Underwood (RHCE,RHCSA,RHCT,UACA)
> My website
> LibreQoS mailing list
> LibreQoS at lists.bufferbloat.net
This song goes out to all the folk that thought Stadia would work:
Dave Täht CEO, TekLibre, LLC
More information about the LibreQoS