------- start of forwarded message (RFC 934 encapsulation) -------
From: Marc Slemko <marcs@znep.com>
Sender: owner-end2end-interest@ISI.EDU
To: Tony Li <tli@juniper.net>
cc: end2end-interest@ISI.EDU
Subject: Re: ftp://ftp.ietf.org/internet-drafts/draft-gettys-webmux-00.txt
Date: Mon, 14 Sep 1998 13:32:15 -0700 (PDT)
On 14 Sep 1998, Tony Li wrote:
> marcs@znep.com (Marc Slemko) writes:
>
> > Are they just looking at all HTTP traffic, or actually comparing HTTP/1.1
> > and 1.0 requests? Most HTTP requests are still 1.0 at this time,
> > apparently because no one at Netscape sees the advantages of HTTP/1.1.
>
>
> Must be time to start advocating (painful tho it may be) IE. When will IE
> support 1.1?
IE4 does support HTTP/1.1. Well, mostly. It obscenely doesn't support
some basic parts of it like Content-Types, and it doesn't do pipelining,
but it does do HTTP/1.1 persistent connections and chunking (except for
the cases where it is broken and doesn't handle it; not sure if any are
left).
It is not intuitively obivious if supporting pipelining would have a
significant impact on the network view of the protocol. It is clear that
it woudl in theory, but it is a question of how often the conditions where
it would have an impact actually happen.
As far as I am aware, Navigator 5.x will have at least basic support for
HTTP/1.1, but it has been slow in coming and several Netscape people I
have talked to have been resistent to it and did not see any benefit.
There is a lot of room for real-world numbers based on traffic analysis,
including things like:
- what % of requests are 1.0? what % are 1.1?
- what % of 1.0 requests are persistent connections? How many of
the ones that aren't can't be done persistently with 1.0
due to the lack of a content length; the implication here is
that they could be persistent using 1.1 chunking.
- what % of requests do and don't have a content length?
- how many connections does the average client open at once?
- how many requests per connection does the average client
make? What is the difference between non-persistent 1.0
connections (obviously 1 req/conn), persistent 1.0 connections,
and persistent 1.1 connections?
- how does a varying timeout for persistent connections impact
the average number of request per connection?
- how localized is a single client's traffic at a given time?
ie. how many servers will a client talk to within a small
period of time that could represent one "page view"?
- how many 1.0 look like they are something (eg. cache validation)
that would benefit greatly from the merging of segments that
can result from pipelining?
- what % of servers are actually talking 1.1?
Most of these numbers can be grabbed from traffic dumps without too much
effort, a few require more complex analysis and extrapolation. I haven't
seen any research presenting the results of such studies, and I think it
could be very useful.
------- end -------