[Ecn-sane] proving out SCE

[Ecn-sane] proving out SCE

[Pete Heist]

[-- Attachment #1: Type: text/plain, Size: 616 bytes --]

I’m starting to read Jon’s proposal for modifying TCP CC for SCE (https://github.com/dtaht/bufferbloat-rfcs/blob/master/sce/ELR%20Proposal%203%20(TCP).txt <https://github.com/dtaht/bufferbloat-rfcs/blob/master/sce/ELR%20Proposal%203%20(TCP).txt>), and get the (perhaps misplaced) feeling that it’s not that big a stretch to implement, and that proving basic efficacy may not be that hard. I’d like to do what I can to help before IETF, so please do post what needs doing, and meanwhile I’ll just keep reading.

Pete

p.s.- at least we don’t have to justify bifurcating Internet congestion control


[-- Attachment #2: Type: text/html, Size: 1021 bytes --]

Re: [Ecn-sane] proving out SCE

[Jonathan Morton]

> On 16 Mar, 2019, at 12:11 pm, Pete Heist <pete@heistp.net> wrote:
> 
> I’m starting to read Jon’s proposal for modifying TCP CC for SCE (https://github.com/dtaht/bufferbloat-rfcs/blob/master/sce/ELR%20Proposal%203%20(TCP).txt), and get the (perhaps misplaced) feeling that it’s not that big a stretch to implement, and that proving basic efficacy may not be that hard. I’d like to do what I can to help before IETF, so please do post what needs doing, and meanwhile I’ll just keep reading.

It'd be really nice to have a complete proof-of-concept implementation that we can pull real numbers out of, even if it's just a bare-minimum spitball in ns-3 or the like.

There exists an implementation of COBALT in ns-3 now, though I doubt it's mainlined.  Adding a raw SCE threshold to that, or to the Codel implementation that *is* mainlined, should be easy.  Maybe adding support to NewReno in it would also be feasible?  Which option for sender feedback would be easiest in that context?

> p.s.- at least we don’t have to justify bifurcating Internet congestion control

That is certainly a major motivating factor.

 - Jonathan Morton

Re: [Ecn-sane] proving out SCE

[Jonathan Morton]

Also, a version of tcptrace that can plot SCE activity would be valuable.  A plot of SCE marking proportion observed on the forward path, and some indication of any feedback path that exists.

 - Jonathan Morton

Re: [Ecn-sane] proving out SCE

[Pete Heist]

[-- Attachment #1: Type: text/plain, Size: 855 bytes --]

Seeing the marks visually would be useful, even if it’s just to look at how the AQM marks it for now. It looks like tcptrace currently puts the text “CE” and “CWR CE” in the time-sequence graphs. This is where you’d want to see “SCE”?

Otherwise, I could extract the bits from Go and use some plotting lib or make a csv. I’ve had to parse pcaps before (https://github.com/heistp/wanonpcap <https://github.com/heistp/wanonpcap>). I just don’t know if you need to see the ECN bits alongside window sizes or other things...

> On Mar 16, 2019, at 3:58 PM, Jonathan Morton <chromatix99@gmail.com> wrote:
> 
> Also, a version of tcptrace that can plot SCE activity would be valuable.  A plot of SCE marking proportion observed on the forward path, and some indication of any feedback path that exists.
> 
> - Jonathan Morton
> 


[-- Attachment #2: Type: text/html, Size: 1408 bytes --]

Re: [Ecn-sane] proving out SCE

[Jonathan Morton]

> On 16 Mar, 2019, at 6:21 pm, Pete Heist <pete@heistp.net> wrote:
> 
> This is where you’d want to see “SCE”?

That might be one way to do it, but since the steady state is supposed to be 50% SCE and existing TCPs won't even react to it, it'll start looking very dense very quickly.

I was thinking you could add it as an extra plot (a windowed average of proportion), or incorporate it into the throughput plot.  Or colour-code the segment bars.  Some sort of graphical representation, anyway.

 - Jonathan Morton

Re: [Ecn-sane] proving out SCE

[Pete Heist]

> On Mar 16, 2019, at 5:49 PM, Jonathan Morton <chromatix99@gmail.com> wrote:
> 
> That might be one way to do it, but since the steady state is supposed to be 50% SCE and existing TCPs won't even react to it, it'll start looking very dense very quickly.
> 
> I was thinking you could add it as an extra plot (a windowed average of proportion), or incorporate it into the throughput plot.  Or colour-code the segment bars.  Some sort of graphical representation, anyway.

Right, a windowed average of the proportion is really what we want to see. I’ll see what I can come up with...