[Tmrg] Proposal to increase TCP initial CWND

fred at cisco.com (Fred Baker) Tue, 20 July 2010 15:13 UTC

From: fred at cisco.com (Fred Baker)
Date: Tue, 20 Jul 2010 08:13:19 -0700
Subject: [Tmrg] Proposal to increase TCP initial CWND
In-Reply-To: <AANLkTilUcGqwMWkMRYfUiXSf5uAXnFPxKfgGlKa-TW-F@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>
Message-ID: <E7E5F841-C9F0-46CD-974C-FA3D38997E9C@cisco.com>

I'm sure that for the user in question, user completion time was faster with a higher initial window, for the simple reason that the first few RTTs are fully used.

That's non-responsive to the question that Lachlan raised yesterday and which I commented on, which is "how does that affect other traffic".

On Jul 19, 2010, at 8:25 PM, Jerry Chu wrote:

> On Mon, Jul 19, 2010 at 7:11 AM, Fred Baker <fred at cisco.com> wrote:
>> "Typical" varies by vendor and software implementation. On Cisco products, the default IOS configuration gives 40 buffers in the queue.
> Thanks for the data. So a "typical" dialup link can easily accommodate a couple
> of simultaneously IW=10 flows, but will likely drops pkts for the more
> aggressive
> services that open more than 3 connections, each will full load.
> We recently setup a small testbed to try to evaluate the effect of
> IW=10 in various
> scenarios. The test results are still preliminary but like what we expected, in
> many contentious cases pkt retransmission rates are higher for IW=10 compared
> to IW=3. But in most cases the "user completion time" for IW=10 still beat out
> that of IW=3. It looks like pkt drops don't necessitate lower performance.
> In the IW=10 cases the saving in the roundtrips more than compensate for the
> extra roundtrips to recover lost pkts. Here every pkt and flow are accounted for
> so there is no hidden "victim" flow.
> Jerry
>> On Jul 19, 2010, at 1:12 AM, Jerry Chu wrote:
>>> On Sat, Jul 17, 2010 at 11:53 PM, Fred Baker <fred at cisco.com> wrote:
>>>> This is essentially the point of concern I raised to them; I also raised the point that what works well in the broadband Internet is quite a bit different than Internet access in parts of the world that have slower typical access rates.
>>> I'd like to point out Google's services are not restricted to just
>>> broadband users. Also from our experiements
>>> some of the greatest latency improvments of a larger IW seem to have
>>> come from users behind slow or highly
>>> multiplexed links.
>>> Do you know how much buffer space a typical dial-up (e.g., 56Kbps)
>>> access router provides?
>>> Jerry
>>>> On Jul 17, 2010, at 11:27 PM, Lachlan Andrew wrote:
>>>>> - An increase in retransmissions from 3.94% to 5.79% seems very
>>>>> significant, if that rate reflects the loss rate.  Although described
>>>>> as a 1.85% increase, it seems likely to cause competing traffic to
>>>>> slow down by a factor of  sqrt(5.79/3.94).  Moreover, that is the
>>>>> increase in loss when (presumably) a *single* flow on that bottleneck
>>>>> increases its initial window.  How can we estimate the impact of *all*
>>>>> flows increasing their initial windows?
>>>> http://www.ipinc.net/IPv4.GIF
>> http://www.ipinc.net/IPv4.GIF