Network for Sale

Philippe Strauss philou at safe-host.net
Thu Feb 22 15:06:13 UTC 2001


On Wed, Feb 21, 2001 at 01:23:31PM -0800, Paul A Vixie wrote:
> 
> > > RTT doesn't drive performance, (bw*delay)-loss does.
> > 
> > FWIW, the heavily interactive apps are more affected by RTT than they are
> > by bandwidth. Network games are the new TELNET. They despise varying
> > latency levels, and are generally oblivious to bandwidth. Your point is
> > still mostly valid, in that the only thing they hate more than varying
> > latency is packet loss, but if the network isn't losing packets then RTT
> > does affect "perfomance" for the heavily interactive apps.
> 
> i wasn't talking about network performance in general.  (and i would never,
> ever recommend (bw*delay)-loss as a routing metric!)  this is in the context
> of so-called global server load balancing.  RTT may, or may not, matter in
> the decisions such a system must make ("serve this client from which proxy?")

when I was talking (writing) about RTT, it's the variation above
the minimal RTT which is interesting as a congestion indicator.
still, it's the case were each node use FIFO queueing which
is ''easily'' analysed (refer to vern paxson papers amongst other).

RTT in itself is nothing, analysing RTT dynamic over time is very interesting.
But in any case, packet drop rate as an impact of a much higer order of
magnitude than RTT variation.

But interesting RTT variation happen usually before packets get dropped.
again, presuming FIFO queueing. Per-flow queueing may give a more blurred
picture.

-- 
Philippe Strauss, safehost sa

En offrant aux regards trop de drame humain, la technologie nous a
desensibilisés, tant sur notre propre souffrance que sur celle des autres.




More information about the NANOG mailing list