Re: [apps-discuss] I-D Action: draft-nottingham-http-browser-hints-00.txt

[Brian Pane <[e-mail]>]

On Thu, May 19, 2011 at 5:51 PM, Mark Nottingham <[e-mail]> wrote:

> Also, header compression only works well if you have many requests
> on a connection; IME we're not there yet.

My experience has been that the number of requests per site (scheme+authority)
often is large enough to make header compression interesting, but that number
divided by ~6 (a modern browser's connection parallelism) usually isn't.  Thus I
see header compression being viable in combination with another change: hinting
to the client that it should use pipelining and a smaller number of connections.
But yes, at that point we're halfway to reinventing SPDY.

> Also, there is more in the draft than just reducing request header sizes; while
> that might be the most beneficial case for you, I actually started the draft for
> the pconn-ip behaviour.

The pconn-ip hint looks good.  The only complication I can think of is that the
client implementer will have to decide whether to manage the maximum size
of its persistent connection pool on a per-IP-addr or per-site basis.  E.g., if
static.example.com and images.example.com resolve to the same address,
a typical browser today might open N concurrent connections to static and
another N connections to images.  If the client honors the pconn-ip hint for
both those sites, it needs to decide on a suitable number of concurrent
connections.  Some possibilities that come to mind are:
  N (may reduce concurrency)
  Some fixed number P > N
  M*N, where M is the number of sites that resolve to the same addr
        (that the client knows about)
None of those is really ideal, though.

Finally, some notes on max-conns: I think it's a good hint, as it leads server
side applications away from having to use tricks like domain sharding.  And
for servers that can handle a lot of concurrent connections, the ability to tell
the client that it's okay to open some very large number of persistent
connections
reduces the need to come up with a pipeline hinting standard.  The catch, I
think, is that the clients most in need of a hint to increase their connection
limit are specific old browser versions, still in widespread use, that
aren't likely
to ever be updated with support for a new browser hint standard.

-Brian