Re: [Cake] fairness vs RTT

["Toke Høiland-Jørgensen" <toke@toke.dk>]

Pete Heist <pete@eventide.io> writes:

>> On Apr 16, 2018, at 11:23 PM, Toke Høiland-Jørgensen <toke@toke.dk> wrote:
>> 
>>> I remember that fairness behavior at low RTTs (< 20ms) needed to be
>>> either improved or documented, and don’t see anything about that in
>>> the man page in the tc-adv repo thus far. Summarizing the host
>>> isolation results at
>>> http://www.drhleny.cz/bufferbloat/cake/round2/#hostiso_cake_{rtt}_{qos-id}
>>> <http://www.drhleny.cz/bufferbloat/cake/round2/#hostiso_cake_%7Brtt%7D_%7Bqos-id%7D>:
>>> 
>>> RTT: fairness (1.0 == perfect fairness)
>>> ---
>>> 100us: 2.22
>>> 1ms: 1.7
>>> 2ms: 1.6
>>> 3ms: 1.42
>>> 5ms: 1.31
>>> 8ms: 1.16
>>> 10ms: 1.12
>>> 20ms: 1.02
>>> 40ms: 1.017
>> 
>> Erm, what's the metric and which data source are you looking at here?
>
>
> Subject changed...
>
> The clients were as follows:
> Client 0- 1 stream up
> Client 1- 12 streams up
> Client 2- 1 stream down
> Client 3- 12 streams down
>
> It looks like what I used before was Client 3’s “TCP Download Sum avg” divided by Client 2’s “TCP Download avg” from the srchost/dsthost tests. The data’s in a table here (see column “Client 3 Mean / Client 2 Mean”):
>
> https://docs.google.com/spreadsheets/d/1e06ZfHSSmecJx9sPU2s2g2GYCS18cMIhBN8PXf1jwaM/edit?usp=sharing <https://docs.google.com/spreadsheets/d/1e06ZfHSSmecJx9sPU2s2g2GYCS18cMIhBN8PXf1jwaM/edit?usp=sharing>
>
> I was calculating by hand before, so the numbers are slightly different, but that’s the idea.
>
> Now, these tests were done at 500Mbit between two cabled APU2s, so we
> could just be running out of CPU on this hardware at lower RTTs. There
> are CPU stats included, and a “Flent Data Files” section. Is it
> possible to tell from this if CPU is the problem? The highest median
> client load I see looks to be 0.77 for the 40ms tests, for example,
> with a mean of 0.63.

Well, the CPU usage meter is just summing that first line of /proc/stat;
so yeah, individual CPUs can definitely be 100% loaded. And the fact
that the test only achieves a total of ~200 Mbps rather than the 500
would indicate that this is the case...

-Toke