Re: multiplexing -- don't do it

[Jon Leighton <[e-mail]>]

On 04/04/2012 06:05 PM, Mike Belshe wrote:
>
>
> On Wed, Apr 4, 2012 at 1:58 PM, Peter Lepeska <[e-mail] 
> <mailto:[e-mail]>> wrote:
>
>     As long as SPDY is sent over TCP, it also suffers from HOL
>     problems, just not as bad as pipelining.
>
>     I think SPDY (or whatever the HTTP 2.0 muxing protocol is) should
>     framed in such a way that if running over a protocol like SCTP,
>     that solves the HOL problems, we should be able to take advantage
>     of it. Due to gzip compression of headers, even if the transport
>     allowed me to grab messages out of order, I'd still have to wait
>     for all packets prior in order to decode the HTTP headers.
>
>
>
> I think people have confusion about layering on top of transport 
> protocols.  Any time you have an app protocol that wants to take 
> advantage of new transport features you *MUST* change the definition 
> of how the app protocol is bound to the lower level transport.  This 
> absolutely applies to HTTP and TCP/SCTP (not talking about SPDY yet). 
>  For example, RFC2616 does *not* specify how to use HTTP over SCTP, 
> and a whole I-D exists for that: 
> http://tools.ietf.org/html/draft-natarajan-http-over-sctp-01  This I-D 
> defines one possible binding between HTTP and SCTP, but others could 
> exist too.
>
> Why is this necessary?  Well, TCP has a single, bidirectional stream. 
>  SCTP doesn't have bi-directional streams at all, and instead only has 
> multiple, uni-directional streams.  So, if you want to leverage the 
> new feature (streams) you need to define how you're going to bind onto 
> it.  It's trivial, but doable.
>
> If we had SCTP, we wouldn't do MUX and SPDY.  It's redundant.  That's 
> not to say that HTTP/2.0 is worse with multiplexing, its just to say 
> that we wouldn't need to consider it if the transport already had it. 
>  But SCTP is not viable today or even within the next decade.
>
> Google sponsored a project at the Univ of Delaware where they 
> investigated SPDY over SCTP already.  A couple of bindings for SPDY 
> over SCTP were considered:  use stream 0 for a control stream, and 
> send the headers over that stream.  This has the nice property that it 
> binds the stateful compression to a single, in-order stream.  Also, as 
> you can imagine, it introduces a small amount of HoL there.  Another 
> solution was to use N SCTP streams, with M mux'd SPDY streams on top 
> of it.  Other such bindings could be considered.  Overall, SCTP was 
> too immature to really benchmark.  The FreeBSD and Linux 
> implementations of SCTP have many problems already.  The UoD guys 
> might want to comment, I found the whole thing non-conclusive.

For anyone interested, here's a link to a summary of the work 
http://www.eecis.udel.edu/~leighton/SPDY.html 
<http://www.eecis.udel.edu/%7Eleighton/SPDY.html>. Under the conditions 
tested, SPDY over SCTP was clearly faster than SPDY over TCP. This is 
true regardless of which mapping was used. Though the data is not 
provided in the summary, using SCTP stream 0 as a control stream for 
SPDY headers was only very slightly faster than sending headers on the 
same SCTP stream as the data. There were some issues with the linux 
implementation of SCTP that had to be worked around (most since fixed), 
but I can't comment on any issues with SCTP on FreeBSD as we decided 
against trying to port Chrome and the flip_in_mem_edsm_server to 
FreeBSD. The results were inconclusive with regard to how best to map 
SPDY to SCTP. Regarding the benefit of using SCTP to avoid HOL blocking, 
the results are pretty straightforward - SPDY performed better over SCTP 
than TCP. Increasing the delay and loss increases SCTP's advantage.

> My recommendation:
> Designing HTTP/2.0 for SCTP is a mistake and should NOT be a 
> requirement.  SCTP is not a viable transport over the Internet today, 
> and will not be in the foreseeable future.  When it is available, an 
> appropriate binding for HTTP/2.0 can be determined, trivially, and we 
> can worry about it then.  This is the same approach that was taken for 
> HTTP with http://tools.ietf.org/html/draft-natarajan-http-over-sctp-01.
>
> Mike
>
>
>
>     Peter
>
>
>     On Wed, Apr 4, 2012 at 9:07 AM, Patrick McManus
>     <[e-mail] <mailto:[e-mail]>> wrote:
>
>         On Wed, 2012-04-04 at 07:02 +0000, Poul-Henning Kamp wrote:
>         > In message <[e-mail]
>         <mailto:[e-mail]>>, Willy Tarreau writes:
>         >
>         > >> I'm starting to get data back, but not in a state that
>         I'd reliably
>         > >> release. That said, there are very clear indicators of
>         intermediaries
>         > >> causing problems, especially when the pipeline depth
>         exceeds 3 requests.
>         >
>         > I always thought that the problem in HTTP/1.x is that you
>         can never
>         > quite be sure if there is an un-warranted entity comming
>         after at GET,
>
>         its not uncommon to have the consumer RST the whole TCP
>         session when
>         asked to recv too far beyond the current request it is
>         processing. For
>         some devices "too far" appears to be defined as "any new
>         packet". I
>         presume some variation of this is where Will's data point
>         comes from.
>         (Often 3 uncompressed requests fit in 1 packet).
>
>         That class of bug sounds absurd, but its really a pretty
>         common pattern.
>         As an example: hosts that fail TLS False Start (for which I
>         understand
>         second hand that Chrome needs to keep a black-list), react
>         badly because
>         there is TCP data queued when they are in a state that the
>         expect their
>         peer to be quiet. Same pattern.
>
>         The lesson to me is that you want to define a tight set of
>         functionality
>         that is reasonably testable up front - and that's what you can
>         depend
>         widely on later. Using anything beyond that demands excessive
>         levels of
>         pain, complexity, and cleverness.
>
>         (and all this pipelining talk as if it were equivalent to spdy
>         mux is
>         kind of silly. Pipelining's intrinsic HOL problems are at
>         least as bad
>         of an issue as the interop bugs.)
>
>         -Patrick
>