[Tmrg] Proposal to increase TCP initial CWND

fred at cisco.com (Fred Baker) Tue, 20 July 2010 20:13 UTC

From: fred at cisco.com (Fred Baker)
Date: Tue, 20 Jul 2010 13:13:49 -0700
Subject: [Tmrg] Proposal to increase TCP initial CWND
In-Reply-To: <AANLkTimGWEA6hQYs5rq4aAV2ZzUQqnFyleH2k0ZLs47m@mail.gmail.com>
References: <AANLkTil937lyUzRvUtdqd2qdl9RN7AZ-Mo_cT-dtmqXz@mail.gmail.com> <923833DA-FC31-4D30-96A5-48D4D072DBD3@cisco.com> <AANLkTimFEGfPJo9rjTw6oNPmwuHkEFBmAX1Ps2yYsUWh@mail.gmail.com> <67225862-FF5A-4B3B-B681-C07E5A705B35@cisco.com> <AANLkTilUcGqwMWkMRYfUiXSf5uAXnFPxKfgGlKa-TW-F@mail.gmail.com> <E7E5F841-C9F0-46CD-974C-FA3D38997E9C@cisco.com> <AANLkTinEqcCf1AyTiF2vKuB-duInnh9Y1CJqff0v1Zn8@mail.gmail.com> <C2956C2A-25B2-4363-A896-AA8E6101F3B2@cisco.com> <AANLkTimGWEA6hQYs5rq4aAV2ZzUQqnFyleH2k0ZLs47m@mail.gmail.com>
Message-ID: <B4047E17-E2D6-482E-A9B6-FAE10AF85F5B@cisco.com>

On Jul 20, 2010, at 12:12 PM, Matt Mathis wrote:

> Actually one of the arguments that is being used against SCTP it that
> you have to forfeit the multiple connection IW boost that TCP gets.
> e.g. if we agree on IW=10 and 4 concurrent TCP connections, then to be
> fair we should also allow a single SCTP session to start at 40.....
> Although this may be logically consistent, I can't bring myself to
> advance it as a serious position.

To be honest, I think the example I gave yesterday would suggest that anyone who thinks that they can presume that the queue between the residential gateway and its ISP (or, more to the point, the queue between the ISP and the residential gateway) is absolutely empty at the point in time they want it to be is smoking something.

> Especially without raising IW, multiple TCP connections has a huge
> advantage over SCTP, and the people who care have already noticed.  We
> aren't going to change their minds by whining about it.   A corollary
> is that the current too small IW standard is actually one of the huge
> barriers to SCTP deployment.

Well, some have suggested bumping cwnd by IW each time a new stream is opened. Some variation on that logic might make sense, such as setting cwnd to 10 if you are opening the first stream and bumping it a little each time a new stream is opened. I'll argue that cwnd is a measure of available capacity (you're moving one cwnd/rtt) and the fact that you opened a new stream doesn't imply that the capacity has increased.

I think your argument is (a) unresponsive to the point raised ("what happens to the sessions you are competing with"), and (b) neglects issues over the past few years that have led service providers and the US FCC in the direction of doing something to protect the competing traffic from applications and application companies (google, yahoo, BitTorrent, etc) that don't give a hoot about anybody but themselves. Consider the reason that the LEDBAT working group came into being. I'd really be interested in a response that is responsive to the question.

> Thanks,
> --MM--
> The best way to predict the future is to create it.  - Peter Drucker
> On Tue, Jul 20, 2010 at 1:50 PM, Fred Baker <fred at cisco.com> wrote:
>> On Jul 20, 2010, at 10:20 AM, Matt Mathis wrote:
>>> As I said earlier: the applications community has already usurped the
>>> transport's ability control congestion.  We will not regain control
>>> unless we pick IW that meets some global optimal criteria for a
>>> reasonable level of parallelism (e.g. 4 connections), and forbid
>>> legacy applications that are too aggressive from moving to new stacks.
>> Or, unless we convince them to use SCTP and put the connections in separate streams. There are good solutions there.
>> http://www.ipinc.net/IPv4.GIF