W3C home > Mailing lists > Public > public-webrtc@w3.org > June 2012

Re: Target numbers for setup time (Re: Keeping up data channel)

From: Harald Alvestrand <harald@alvestrand.no>
Date: Sun, 24 Jun 2012 14:55:23 +0200
Message-ID: <4FE70E3B.7010700@alvestrand.no>
To: Justin Uberti <juberti@google.com>
CC: Roman Shpount <roman@telurix.com>, public-webrtc@w3.org
On 06/22/2012 08:55 PM, Justin Uberti wrote:
> DataChannel protocol (1) atop SCTP (2) atop DTLS (2) atop ICE (1) has 
> 6 RTT in the handshakes, minimum. That's not counting any signaling or 
> candidate gathering time.
>
> If we want better perf, we need to collapse handshakes. This should be 
> very doable (see Martin's draft on running DTLS handshake over ICE), 
> but I tend to think this is something we can do later.

And - chastising myself since I was the one who renamed the thread 
without transferring it - this being a protocol issue, it belongs in 
RTCWEB, not in WEBRTC..

>
> Just having *any* p2p working in the browser is a huge step from where 
> we are today.
>
> On Fri, Jun 22, 2012 at 1:54 PM, Roman Shpount <roman@telurix.com 
> <mailto:roman@telurix.com>> wrote:
>
>
>     On Fri, Jun 22, 2012 at 1:09 PM, Harald Alvestrand
>     <harald@alvestrand.no <mailto:harald@alvestrand.no>> wrote:
>
>         On 06/22/2012 05:58 PM, Cullen Jennings (fluffy) wrote:
>
>             On Jun 13, 2012, at 5:29 , Randell Jesup wrote:
>
>                     How far down do you think we have to drive the
>                     setup time before you
>                     would not call it "abysmal"?
>
>
>             I'd probably consider above 250 ms abysmal but good news I
>             don't see any problem with getting it down around 100 ms
>             in when both endpoints are in a single country.
>
>         Coast-to-coast US is ~4800 km, so RTT (9600 km) is 32 ms
>         (speed of light is 300 km/msec).
>
>         So, without considering processing time, 3 RTT is 100 msec, 7
>         RTT is "abysmal".
>
>         There are bigger countries than the US, but this will do for a
>         back-of-the-envelope.
>
>
>     IP packets rarely travel in straight lines and typically encounter
>     2-3 routers along the way. In US, from east to west coast, between
>     the data centers, 60-70 ms are typical. Residential to residential
>     are about 80-100 ms. So, 3 RTT in real life are "abysmal" by this
>     standard.
>
>     More reasonable point of view is that normally designed call setup
>     should take 2 RTT (offer, answer and connectivity check in
>     offerer's direction, offerer's connectivity check response and
>     connectivity check in answerer's direction, answerer's
>     connectivity check response). Everything else is adding latency
>     for the sake of bad desing.
>     _____________
>     Roman Shpount
>
>
Received on Sunday, 24 June 2012 12:55:54 UTC

This archive was generated by hypermail 2.3.1 : Monday, 23 October 2017 15:19:28 UTC