[Ecn-sane] fq_codel_fast ECN signals vs number of flows

[Ecn-sane] fq_codel_fast ECN signals vs number of flows

[Pete Heist]

Plots of congestion signals for increasing numbers of iperf3 flows:

https://www.heistp.net/downloads/fqf_nflows/

Most of it is expected. SCE signals start earlier than CE. As you increase the number of flows the number of CE signals increases, and SCE signals decrease until eventually all packets have CE markings and SCE is not used.

It’s still surprising though that with one flow that there’s really just one CE signal then that’s it, yet the rate is limited without drops. Is that somehow caused by TSQ?

Re: [Ecn-sane] fq_codel_fast ECN signals vs number of flows

[Jonathan Morton]

> On 20 Mar, 2019, at 8:49 pm, Pete Heist <pete@heistp.net> wrote:
> 
> It’s still surprising though that with one flow that there’s really just one CE signal then that’s it, yet the rate is limited without drops. Is that somehow caused by TSQ?

Yes, that's expected in Linux if the bottleneck queue is local to the sender.

 - Jonathan Morton