On Wed, 9 Jan 2002, Gorry Fairhurst wrote:
> on 9/1/02 11:34 am, Andrei Gurtov at gurtov@cs.Helsinki.FI wrote:
>
> > In my opinion, connections using SACK and Limited transmit can afford having
> > a conservative RTO since they suffer non-spurious timeouts infrequently.
> >
> > RFC2988 is a good starting point to provide a conservative enough timer
> > suitable for wireless environment. The next step could be to use
> > non-decreasing version of the timer, i.e. to use only RTT samples that
> > increase the RTO value. This way the RTO timer becomes the estimate of the
> > highest RTT value observed on this connection so far.
>
> But, what about path changes, from a slow/long path to a short delay path?
> --- surely there must be some grounds for reducing the RTO. Even if, the
> tendency is to fix the RTO much larger than the RTT measurements, MAX seems
> excessive...
I know it sounds overly conservative, but nothing better comes to mind
while thinking of a suitable rxmt timer for GPRS with fairly stable 1
s RTT with eventual 10 sec delay spikes. A possible option is to reset the
timer to estimate based on actual RTT when a non-spurious timeout is
encountered. Still, this may be a better approach than hard-coding some
(arbitrary) fixed minimum RTO.
Andrei
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 09:12:29 EST