W3C home > Mailing lists > Public > ietf-http-wg@w3.org > October to December 2013

Re: Question on flow control for a single file transfer

From: Roberto Peon <grmocg@gmail.com>
Date: Mon, 4 Nov 2013 16:08:53 -0800
Message-ID: <CAP+FsNfTdCiaD0wDCdKuxYC63ws1d0-sb5BNA72ORdHQsSB3Cw@mail.gmail.com>
To: Peter Lepeska <bizzbyster@gmail.com>
Cc: Mike Bishop <Michael.Bishop@microsoft.com>, Amos Jeffries <squid3@treenet.co.nz>, HTTP Working Group <ietf-http-wg@w3.org>
I'll point out again, that the proposal of ignoring flow control at any
time implies that one MUST have a synchronization event when one
transitions from ignoring to adhering to flow control. With the proposal of
ignoring when only having one stream, *which is the common case for the
beginning of browsing*, it would cause de-synchronization, possible
deadlock, etc.

Even if I liked the proposal, which I don't, it would need to be completed
with adequate synchronization for the transition from 1 stream to multiple
streams, and vice versa (what happens when you go from 2-1-2 streams?).


Dealing with this would require additional complexity far in excess of
simply following the flow control semantic as it is defined today.

-=R


On Mon, Nov 4, 2013 at 2:20 PM, Peter Lepeska <bizzbyster@gmail.com> wrote:

> Okay I get the misunderstanding now.
>
> I'm not proposing that the sender determines that there is just one active
> stream end to end. It just checks to see if it knows of other active
> streams. If it does not, then it ignores flow control. If in fact there are
> other active streams end to end, then it will not be in the "ignoring flow
> control" state for long.
>
> Peter
>
>
> On Mon, Nov 4, 2013 at 2:15 PM, Mike Bishop <Michael.Bishop@microsoft.com>wrote:
>
>>  No, the sender *doesnít* know if thereís a single stream end-to-end.
>> It knows whether thereís a single stream on the first hop, and it doesnít
>> know whether that hop is the only hop.
>>
>>  If a client is directly connected to a server and thatís the only
>> request that will ever be made, then youíre correct -- you may get some
>> improved performance by disabling flow control, and we provide that option
>> in the spec.  But the server, at least, doesnít know whether itís dealing
>> with a proxy or a direct client.  The client can inform its next hop
>> whether it wants flow control on that hop.  If that next hop is a proxy,
>> itís the proxyís decision whether it wants flow control, and thatís an
>> independent choice.
>>
>>  Sent from Windows Mail
>>
>>   *From:* Peter Lepeska <bizzbyster@gmail.com>
>> *Sent:* Monday, November 4, 2013 1:17 PM
>>
>> *To:* Amos Jeffries <squid3@treenet.co.nz>
>> *Cc:* HTTP Working Group <ietf-http-wg@w3.org>
>>
>>
>>
>>
>> On Mon, Nov 4, 2013 at 1:02 PM, Amos Jeffries <squid3@treenet.co.nz>wrote:
>>
>>> On 2013-11-05 05:05, Peter Lepeska wrote:
>>>
>>>> Amos,
>>>>
>>>> I agree what you said, but again only when there is more than one active
>>>> stream. Again, HTTP 2 flow control is harmless at best when there is
>>>> only
>>>> one active stream.
>>>>
>>>
>>>  Part of my point was that there is absolutely no way to determine that
>>> one active stream cases existence all the way along the path. Middleware
>>> exists (whether it is visible to the endpoints or not) and the "single
>>> stream" may be sharing any HTTP hop with one or more other streams.
>>>
>>> "At best" this single stream will be able to avoid contention in the
>>> more common cases where it ceases being a single end-to-end stream at some
>>> middel hop. So no I think the best-case is rather better than you are
>>> saying.
>>
>>
>>  The sender knows if there are other active streams at the time it has
>> data to send.
>>
>>
>>>
>>>
>>>
>>>
>>>> But you don't have to believe me. Just setup a test a browser that does
>>>> flow control and add a few % loss and 200 ms latency and see whether you
>>>> are able to download large files faster with flow control on or off. The
>>>> flow control off case should never lose, assuming the loss/latency are
>>>> regular and your test is long enough.
>>>>
>>>
>>>  At what size data frames? and what relative size of TCP and HTTP layer
>>> buffer sizes? over how many hops?
>>>
>>> In the grand scheme of HTTP, single client going to single server, with
>>> a single stream and nothing in between is a rather rare occurance. Just
>>> like it is a rather rare and artificial occurance to see only a single
>>> isolated TCP connection today.
>>
>>
>>  Those questions will not impact your results. When there is available
>> buffering at the TCP layer, HTTP flow control makes one of two decisions --
>> send now or send later. When flow control is disabled the answer is always
>> send now. Therefore no HTTP flow control will always be as fast or faster.
>>
>>  Actually, I don't know for sure but I'd bet the single stream case is
>> the most common from an overall bytes sent perspective due to http
>> streaming movies from services like Netflix. In any case, there are many
>> uses of HTTP that involve one at a time file transfers.
>>
>>  Peter
>>
>>>
>>>
>>> Amos
>>>
>>>
>>>
>>
>
Received on Tuesday, 5 November 2013 00:09:20 UTC

This archive was generated by hypermail 2.3.1 : Tuesday, 1 March 2016 11:11:19 UTC