Re: New Version Notification for draft-montenegro-httpbis-speed-mobility-02.txt

["Poul-Henning Kamp" <[e-mail]>]

In message <[e-mail]
ndeploy.ntdev.microsoft.com>, Rob Trace writes:

>A new version of I-D, draft-montenegro-httpbis-speed-mobility-02.txt

I have read this draft, and I do like the thinking, even if not all
the details.

With one small tweak, this prosal has the potential to solve what
I see as one of the really big high-performance issues: load-balancing
at Tbit/s rates.

The tweak I seek is:

	Once created, all requests on a given stream MUST have the
	same "Host:" header.

This does not prevent a client (or proxy) from using multiple streams
with same "Host:" header to get parallelism, but it means that
load-balancers in front of hosting farms only need to examine the
first request on any stream to find the "Host:" header, after that,
it can just shift the stream through.

Another way to think of this, is to make the stream-ID also
work as a hop-to-hop flow-label.

With that said...

A lot of the benefit thus gained will be eaten up because taking a
SPDY stream apart is a performance nightmare, particularly at Tbit/s
rates, but at least the load-balancers will only need to take the
first part of the SPDY streams apart.

I seem to recall Willy and Adrien working on using serialization/encoding
to save header transmission cost, instead of just blindly throwing
things at gzip.

Put something like that on top of Bobs draft instead of SPDY, and
Tbit/sec starts to look realistic.

Poul-Henning

PS: I was happy to see that Guy didn't find any significant improvement
from the gzip static dictionary, I thought I had done something wrong
when I got that result.

-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
[e-mail]         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe    
Never attribute to malice what can adequately be explained by incompetence.