Re: [Ecn-in-quic] ECN-in-Quic w/ ABE

So as one of the people who worked on and evaluated ABE....

There has not yet been wide deployment, but the laboratory tests referenced in the ABE draft cover a wide range of scenarios and did include cubic with ABE (and that's mentioned in the ID).  There's kernel code for BSD that tracks the latest draft:https://svnweb.freebsd.org/changeset/base/331214, the patch for linux cubic really was not much different.

I think it is hard to see any real danger arising from the use of ABE. Consider this: Cubic has used a beta (backoff) factor of 0.8 for a long time, in response to loss, before it was tuned down to 0.7. This may not have been an ideal configuration, but neither has the Internet melted, not has Linux become a useless OS in this period.

We recommend 0.8 for standard Reno-like congestion control and 0.85 for Cubic (although, in our experiments, 0.8 also works fine for Cubic)*ONLY*  in response to ECN, i.e. when the queue is normally smaller than when we see loss. This is, altogether, a way less aggressive behavior than what Cubic had for a long time in Linux.

I'd really encourage we use this rather than the loss-equivalent model - mainly because it is a significant step towards making ECN useful and allowing it to work to reduce latency - which is surely important to apps like QUIC.

Happy to say more (if needed).

Gorry

On 22/03/2018, 19:07, Ingemar Johansson S wrote:
>
> OK. Missed that one. I guess my conclusion may then be a bit premature.
>
> /Ingemar
>
> *From:* Ecn-in-quic [mailto:ecn-in-quic-bounces@ietf.org] *On Behalf 
> Of *Praveen Balasubramanian
> *Sent:* den 22 mars 2018 19:56
> *To:* rs.ietf@gmx.at; ecn-in-quic@ietf.org
> *Subject:* Re: [Ecn-in-quic] ECN-in-Quic w/ ABE
>
> QUIC as specified in drafts does not use Cubic but I expect most 
> implementations to use it. Winquic is doing Cubic already. Are you 
> proposing that quic cc should do the ABE experiment by default? Has 
> its safety been proven by a wide deployment? I asked this q in tcpm 
> and the answer was no.
>
> ------------------------------------------------------------------------
>
> *From:* rs.ietf@gmx.at <mailto:rs.ietf@gmx.at>
> *Sent:* Thursday, March 22, 2018 6:47 PM
> *To:* ecn-in-quic@ietf.org <mailto:ecn-in-quic@ietf.org>
> *Subject:* [Ecn-in-quic] ECN-in-Quic w/ ABE
>
> Just a quick note to those present here:
>
> ABE (Alternative Backoff for ECN) is past WGLC in TCPM, allowing a 
> differntiated CC control between ECN and Loss - to incentivise the 
> deployment of ECN.
>
> Since Quic uses Cubic, the guidance given there for differential 
> reaction between loss and ECN marks would be directly applicable...
>
> Any thoughts?
>
> Best regards,
>   Richard
>
> _______________________________________________
> Ecn-in-quic mailing list
> Ecn-in-quic@ietf.org <mailto:Ecn-in-quic@ietf.org>
> https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.ietf.org%2Fmailman%2Flistinfo%2Fecn-in-quic&data=04%7C01%7Cpravb%40microsoft.com%7Cc0bd42e05fcb4b68184d08d5902553ba%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C1%7C636573412355153091%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwifQ%3D%3D%7C-1&sdata=77SDofa4ECl4rdaA1p%2F72nT0WFDWPhrRtWDu7aCYZFM%3D&reserved=0 
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.ietf.org%2Fmailman%2Flistinfo%2Fecn-in-quic&data=04%7C01%7Cpravb%40microsoft.com%7Cc0bd42e05fcb4b68184d08d5902553ba%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C1%7C636573412355153091%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwifQ%3D%3D%7C-1&sdata=77SDofa4ECl4rdaA1p%2F72nT0WFDWPhrRtWDu7aCYZFM%3D&reserved=0>
>
>
>
> _______________________________________________
> Ecn-in-quic mailing list
> Ecn-in-quic@ietf.org
> https://www.ietf.org/mailman/listinfo/ecn-in-quic