[Cake] Cake ingress experiments & policing

[Adrian Popescu]

Hello,

These are some some random thoughts and notes related to some of my
experiments with cake and traffic shaping.

The performance issues observed on the cake setup seemed to be worth
another look. The mirred and ifb setup seemed to have lower performance
than expected.

This was easy enough to test by replacing cake with sfq on the ifb
interface. The setup had the same performance issues as the setup with
cake.

It's most likely not reasonable to expect cake to be able to shape at
1 gbps either if a setup with something with less complexity can't deal
with 300-400 mbps worth of MTU sized packets in the mirred and ifb
setup.

This set of experiments has determined me to consider alternative
options for ingress. A policer seems to be the right choice.

A policer which sets the ECN bit on ECT enabled flows could help. There
doesn't seem to be a documented way of doing this with tc actions &
filters. Would this be possible for IPv6 traffic? The ECN marking on
ECT enabled flows is the missing bit which would make this work for
IPv4. A pedit action may be useful if combined with a match on ECT.
It's something I haven't figured out yet.

Perhaps some tests with a simple flow blind marking policer would be
worth a shot.

Maybe it'd be worth to see other people's setups and even ISP type
setups to figure out what else to try.


Some of my other experiments included wifi ingress shaping. The upload
speed seemed capped at 3-8 mbps while wifi was shaped on ingress. This
appears to be a bug.
Shaping wifi at a rate which is lower than that of the link's total
bandwidth seems to be more difficult since we can't limit on ingress
based on the egress interface.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.bufferbloat.net/pipermail/cake/attachments/20190303/3f63ad48/attachment.html>