Re: [Cerowrt-devel] cerowrt-3.10.32-9 released

[Dave Taht <dave.taht@gmail.com>]

On Tue, Mar 18, 2014 at 11:22 AM, Török Edwin <edwin@etorok.net> wrote:
> On 03/18/2014 04:21 PM, Dave Taht wrote:
>> Regrettably the SQM system on the wndr series of hardware maxes out on
>> CPU at about 50Mbit down, 10Mbit up, or any combination thereof (e.g
>> 25/25 works). If you want to apply this code at higher rates, routing
>> hardware with more "oomph" is needed.
>>
>> I would be interested in a rrul test of your 50Mbit system. My tests
>> of verizon at 25/25 showed them well managed on the up, far less well
>> managed on the down, so in your 50Mbit design you might want to merely
>> control the down with SQM.
>
> Host: Linux 3.14-rc5, AMD FX(tm)-8350 Eight-Core Processor, cpufreq set to performance, cpb disabled
> Router: NETGEAR WNDR3700v2, 3.10.32-9
> Target: OpenBSD 5.4, QEMU Virtual CPU version (cpu64-rhel6), 3300.54 MHz
>
> Host <-> Router connected via gigabit ethernet.
>
> I confirmed that my target can do >50Mbit to another server, netperf shows 128*10^6 bit/s / 108 * 10^6 bit/s.
>
> SQM with link-layer none settings:
>  * download speed: 46000 kbit/s, upload speed 50000 kbit/s (~95% of measured speed on ISP's speedtest site).
>  * fq_codel (default)
>  * simple.qos
>  * link-layer: none
>  * results: <sqm_on_none.png>
>
> SQM off: <sqm_off.png>
>
> SQM with overhead Ethernet overhead 30: <sqm-on-30.png>
> SQM with overhead Ethernet overhead 22: <sqm-on-22.png>
>
> Should I also run a rrul46/rrul46compete test?

If you can test ipv6 too, that would be great.

> Best regards,
> --Edwin

At 8ms of induced extra latency without SQM there doesn't seem to be
much point in running it on your platform. It does look like you are tail
dropping...

A thought would be to leave it
off and try running your link at 100Mbit rather than gige.

# ethtool -s ge00 advertise 0x008

and see if fq_codel alone can break up bursts better.

I note that linux 3.14 TCP is now so highly debloated that it is really hard to
compare the results we get with it with any TCP before it - the TSO offload
fixes, tcp small queues, etc, have really shortened and made more accurate
the control loops.

second note is that the wndr can only forward packets at about 330Mbit
without firewall rules. Add in the firewall rules and you are looking at
sub 120mbit forwarding performance.

I am jealous of your link. :)


-- 
Dave Täht

Fixing bufferbloat with cerowrt: http://www.teklibre.com/cerowrt/subscribe.html