Re: [hybi] Frame size

[Jamie Lokier <[e-mail]>]

Mike Belshe wrote:
> 
>    On Mon, Apr 19, 2010 at 4:33 AM, Jamie Lokier <[1][e-mail]>
>    wrote:
> 
>    Thomson, Martin wrote:
>    > You make a good case... for SCTP :p
> 
>      Yes, absolutely!
>      SCTP would be a much better match for the problem.
>      Meanwhile, we have to work over TCP.
>      Remember that Google's tests show you can only communicate from
>      many
>      client sites over:
>        - TCP port 80, using HTTP
>        - TCP port 443, using CONNECT
>        - No other TCP ports.
> 
>    Are you talking about the Google WebSocket test?  If so, that is not
>    quite right.
>    The Google Websockets test shows that using HTTP over non-port 80 has
>    more success than HTTP over port 80.  We believe this is due to
>    proxies.

Yes, that test.  Yes it shows more success over non-port 80, but there
is still a significant minority of sites where non-port 80/443 did not work.

That is what I mean by "you can only communicated from many client
sites [...] no other TCP ports".  Because there are many client sites
for which that is true according to the Google WebSocket test.

Sorry if my language was unclear.

>    The problem with SCTP is NAT.  NAT services generally only work for
>    TCP and UDP.  And what percentage of the net uses NAT today?
>    Everyone.  I suppose SCTP might be deployable those using IPv6.

Well, and that problem of sites where non-port 80/443 are unavailable.
I don't see them all letting SCTP through...

NAT is a problem, but there is also something called SPI, "stateful
packet inspection".  That is, firewalls which don't translate
addresses, but still track connections and block unrecognised
connections or those originated in the wrong direction.

-- Jamie