- From: Willy Tarreau <w@1wt.eu>
- Date: Tue, 21 Oct 2014 16:08:43 +0200
- To: Greg Wilkins <gregw@intalio.com>
- Cc: Roberto Peon <grmocg@gmail.com>, Mark Nottingham <mnot@mnot.net>, RUELLAN Herve <Herve.Ruellan@crf.canon.fr>, Amos Jeffries <squid3@treenet.co.nz>, "ietf-http-wg@w3.org" <ietf-http-wg@w3.org>
Hi Greg, On Wed, Oct 22, 2014 at 12:36:30AM +1100, Greg Wilkins wrote: > Willy, > > I actually think that literal without/never indexed should be favoured over > literal indexed or at least on event footing, because by nature the literal > indexed headers are sent only once and then used by index, while the > literal without/never indexed headers are sent literally every time they > are used. OK, thanks for explaining. > Thus I would think something like: > > Indexed Header: > +---+---+---+---+---+---+---+---+ > | 1 | D | Index (6+) | > +---+---------------------------+ Here we still have the 20% waste I wanted to address : ~48/256 values will never be sent (those for static entries >16). That's where I wanted to recover that many values (5.5 bits) which are clearly missing from other encodings. > Literal Never Indexed: > +---+---+---+---+---+---+---+---+ > | 0 | 0 | 0 | D | Index (4+) | > +---+---------------------------+ > > Literal Not Indexed: > +---+---+---+---+---+---+---+---+ > | 0 | 0 | 1 | D | Index (4+) | > +---+---------------------------+ > > Literal Indexed: > +---+---+---+---+---+---+---+---+ > | 0 | 1 | 0 | D | Index (4+) | > +---+---------------------------+ Two of them can then be increased by 1..1.5 bit by recovering the unused encoding above. > Max Dynamic Table size change: > +---+---+---+---+---+---+---+---+ > | 0 | 1 | 1 | Index (5+) | > +---+---------------------------+ And this one looks like a huge waste to me, it only happens once in a long while and kills 32 possible values that we'd better use for each individual header. In practice, 80/256 code points are unusable for each header, that's roughly 33% wasted encoding space :-/. > I believe that most 62 static slots can easily be filled with values that > will be used. That's where I'm really doubting to be honnest. > I'll come up with a suggestion for this tomorrow. OK great! Cheers, Willy
Received on Tuesday, 21 October 2014 14:10:33 UTC