[Cake] Cake throughput levels off during longs tests

[Cake] Cake throughput levels off during longs tests

[Toke Høiland-Jørgensen]

[-- Attachment #1: Type: text/plain, Size: 296 bytes --]

So I ran a couple of tests of Cake on the Archer-c7 on newest openwrt
trunk with sqm-scripts set to 30/30 Mbps. The attached graphs are the
result of equivalent fq_codel and cake settings.

As you can see, the cake throughput levels off over long test durations.
Anyone has any idea why?

-Toke


[-- Attachment #2: cake-30Mbit.png --]
[-- Type: image/png, Size: 195842 bytes --]

[-- Attachment #3: fq_codel-30Mbit.png --]
[-- Type: image/png, Size: 174944 bytes --]

[-- Attachment #4: rrul-2015-10-22T165703.686658.fq_codel_30_30_Mbps_ethernet_kau_long_designated_d.flent.gz --]
[-- Type: application/vnd.flent.data.gzip, Size: 337116 bytes --]

[-- Attachment #5: rrul-2015-10-22T170722.147606.cake_30_30_Mbps_ethernet_kau_long_designated_drive.flent.gz --]
[-- Type: application/vnd.flent.data.gzip, Size: 399899 bytes --]

Re: [Cake] Cake throughput levels off during longs tests

[Jonathan Morton]

> On 22 Oct, 2015, at 18:50, Toke Høiland-Jørgensen <toke@toke.dk> wrote:
> 
> As you can see, the cake throughput levels off over long test durations.
> Anyone has any idea why?

Well, it shouldn’t be doing that.

Can you see any trends in the marking patterns - anything that might reveal what’s happening in codel?

 - Jonathan Morton

Re: [Cake] Cake throughput levels off during longs tests

[Toke Høiland-Jørgensen]

Will re-run tomorrow and take captures. Fwiw, this seems to be specific to the Archer box, though... Don't see the same behaviour on the (x86) test bed. Timer issue in the shaper, perhaps?

On 22 October 2015 18:06:57 CEST, Jonathan Morton <chromatix99@gmail.com> wrote:
>
>> On 22 Oct, 2015, at 18:50, Toke Høiland-Jørgensen <toke@toke.dk>
>wrote:
>> 
>> As you can see, the cake throughput levels off over long test
>durations.
>> Anyone has any idea why?
>
>Well, it shouldn’t be doing that.
>
>Can you see any trends in the marking patterns - anything that might
>reveal what’s happening in codel?
>
> - Jonathan Morton

Re: [Cake] Cake throughput levels off during longs tests

[Toke Høiland-Jørgensen]

Toke Høiland-Jørgensen <toke@toke.dk> writes:

> So I ran a couple of tests of Cake on the Archer-c7 on newest openwrt
> trunk with sqm-scripts set to 30/30 Mbps. The attached graphs are the
> result of equivalent fq_codel and cake settings.
>
> As you can see, the cake throughput levels off over long test durations.
> Anyone has any idea why?

Right, so did some more experiments, and this is definitely related to
CPU usage. I did a new build of openwrt with an updated toolchain, and
now the archer runs fine with Cake at 30 Mbps symmetrical. However, it
starts to struggle at 50 Mbps; but only after around three minutes of
continuous load.

See the comparison graphs here:
https://kau.toke.dk/experiments/cake/archer/cake-30-50.png

That directory also contains the full Flent data files and packet dumps.

So the question is now, of course, is there anything that can be done
about this, or should we simply accept that the box is out of oomph? The
way it behaves does strike me as weird...

-Toke

Re: [Cake] Cake throughput levels off during longs tests

[Sebastian Moeller]

Hi Toke,

On Oct 29, 2015, at 16:59 , Toke Høiland-Jørgensen <toke@toke.dk> wrote:

> Toke Høiland-Jørgensen <toke@toke.dk> writes:
> 
>> So I ran a couple of tests of Cake on the Archer-c7 on newest openwrt
>> trunk with sqm-scripts set to 30/30 Mbps. The attached graphs are the
>> result of equivalent fq_codel and cake settings.
>> 
>> As you can see, the cake throughput levels off over long test durations.
>> Anyone has any idea why?
> 
> Right, so did some more experiments, and this is definitely related to
> CPU usage. I did a new build of openwrt with an updated toolchain, and
> now the archer runs fine with Cake at 30 Mbps symmetrical. However, it
> starts to struggle at 50 Mbps; but only after around three minutes of
> continuous load.

	Could this be some sort of thermal throttling? The plot reminds me a bit of the currently popular plots of smartphone performance/CPU-frequency over time under load over time… I do not know whether the archer’s cpu actually can throttle… If it would be pure CPU-ooomph shouldn’t it be there from early on, 200 seconds seem too long to run out of resources like cache?

> 
> See the comparison graphs here:
> https://kau.toke.dk/experiments/cake/archer/cake-30-50.png
> 
> That directory also contains the full Flent data files and packet dumps.
> 
> So the question is now, of course, is there anything that can be done
> about this, or should we simply accept that the box is out of oomph? The
> way it behaves does strike me as weird…

	Maybe cooling it artificially during the test might reveal whether the slow down is temperature dependent? I would test myself but I do not own an archer C7…

Best Regards
	Sebastian


> 
> -Toke
> _______________________________________________
> Cake mailing list
> Cake@lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/cake