[Rpm] [Bloat] *under*bloated networks

Dave Taht dave.taht at gmail.com
Sun Nov 28 13:24:12 EST 2021


On Sun, Nov 28, 2021 at 7:42 AM Neal Cardwell <ncardwell at google.com> wrote:
>
>
>
> On Mon, Nov 22, 2021 at 10:35 PM Dave Taht <dave.taht at gmail.com> wrote:
>>
>> In the last two weeks I have found two dramatically underbuffered Gbit
>> fiber networks.
>>
>> This one appears to have about a 400 full size packet uplink buffer (5ms)[1]
>>
>> https://imgur.com/a/Bm9hdNf
>>
>> It was pretty remarkable to see how well multiple tcp flows still
>> achieved close to the full rate with such a small fixed size queue,
>> eventually.

And baseline rtts not increasing more than 5ms before seeing a drop with cubic.

https://imgur.com/a/S5v2EF3

>>
>> A single bbr flow can't crack 150mbits: https://imgur.com/a/DpydL5K.
>
>
> Thanks, Dave. The single-flow BBR upload case is interesting.
>
> I took a look at the packet traces in the later thread:
>   https://www.reddit.com/r/eero/comments/qxbkcl/66_is_out/hltlep0/
>
> For the single-flow BBR case it seems that....
>
> (1) the BBR(v1) flow is running into a 300 Mbps bottleneck rate (visible in the slope of the green ACK line in the zoomed-in trace, attached).

This particular bottleneck is capable of a gbit, according to the 8
flow tests. A single cubic flow peaks at 300Mbit,
in earlier tests.

> (2) the BBR(v1) flow is achieving an average rate a bit above 150 Mbps because it repeatedly runs into receive window limits (the yellow line in the zoomed-out trace, attached). The frequent receive window limits mean that the flow spends a lot of time unable to send anything, thus leading to lower average throughput.

In all honesty, I haven't looked at BBR, + long rtts, and rates >
100Mbit often enough. (I actually have never had more than a 35Mbit
uplink to the internet in my whole life.)  What I see here is a big
sack block, and that pause, and yes, the window getting smaller... And
universally, this fellah's link kicking flows out of slow start at
around 100Mbit... but what does it mean?

this server is running 5.11.0-40-generic #44-Ubuntu SMP, if there is
something I can tune.

Most speedtest sites use 8+ flows on the upload portion of the test...
>
> cheers,
> neal
>


-- 
I tried to build a better future, a few times:
https://wayforward.archive.org/?site=https%3A%2F%2Fwww.icei.org

Dave Täht CEO, TekLibre, LLC


More information about the Rpm mailing list