netdev - Re: Cake not doing rate limiting in a way it is expected to do

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <875zpvvsar.fsf@toke.dk>
Date:   Tue, 28 May 2019 11:31:40 +0200
From:   Toke Høiland-Jørgensen <toke@...hat.com>
To:     Akshat Kakkar <akshat.1984@...il.com>,
        netdev <netdev@...r.kernel.org>, lartc <lartc@...r.kernel.org>
Cc:     cake@...ts.bufferbloat.net
Subject: Re: Cake not doing rate limiting in a way it is expected to do

[ adding cake list ]

Akshat Kakkar <akshat.1984@...il.com> writes:

> Cake is expected to handle traffic in 2 steps :
> First is on the basis of host
> Second is within every host, on the basis of flow
>
> So, if I limit traffic to 20Mbps shared across 2 host A & B,
> Following are various scenarios, expectation and observations
> 1. If either A or B is downloading, they will be getting speed of 20Mbps
> Observation: Meeting with expectation
>
> 2. If both A & B downloads (single download each), each will be
> getting speed of 20Mbps
> Observation: Meeting with expecation but its very jittery (around
> 20%), i.e. speed varies from 8Mbps to 12 Mbps. If I use fq_codel speed
> is same BUT jitter is very less (around 1%).
>
> 3. Now if A starts 3 downloads, and B is still having single download,
> A each download should be around 3.3 Mbps and B should be around
> 10Mbps
> Observation: Around 5 Mbps for each download with lot of jitter, i.e.
> no advantage of having CAKE!!!
>
> Linux Kernel 4.20
> For case 3, output of command : tc -s class show dev eno2
>
> class htb 1:1 root leaf 8003: prio 1 rate 20000Kbit ceil 20000Kbit
> burst 200Kb cburst 1600b
>  Sent 688474645 bytes 455058 pkt (dropped 0, overlimits 381196 requeues 0)
>  rate 19874Kbit 1641pps backlog 21196b 14p requeues 0
>  lended: 382532 borrowed: 0 giants: 0
>  tokens: 1260573 ctokens: -9427
>
> class cake 8003:44f parent 8003:
>  (dropped 3404, overlimits 0 requeues 0)
>  backlog 9084b 6p requeues 0
> class cake 8003:516 parent 8003:
>  (dropped 3565, overlimits 0 requeues 0)
>  backlog 0b 0p requeues 0
> class cake 8003:590 parent 8003:
>  (dropped 3023, overlimits 0 requeues 0)
>  backlog 4542b 3p requeues 0
> class cake 8003:605 parent 8003:
>  (dropped 1772, overlimits 0 requeues 0)
>  backlog 7570b 5p requeues 0

Could you please share some more details of your setup? The output of
`tc -s qdisc show dev eno2`?

How are you running the download tests? Is this over the internet, or in
a controlled setup? What's your actual line rate? Are you using ingress
shaping as well?

It looks like you are using HTB for rate limiting on top of CAKE? Why?

-Toke