[Iccrg] Question on RFC 2988 - TCP Retransmission timer

[Mark Allman]

> Nobody is saying that it is safe to have a rule like "use RTO=200ms".
> However, it should be "safe" to trust the RTO calculations, unless
> they predict a very small timeout.  If that is not the case, we should
> IMO fix the RTO calculations, rather than simply ignoring them and
> using a 1s minimum.

The safeness of the RTO is the exponential backoff of the RTO.  IMO.

As I have noted before I think the minimum RTO represents a knob to make
a tradeoff between quicker RTOs and spurious RTOs.  I don't think
determining the correct min RTO is a simple question to answer without
experiments.  (See our 1999 SIGCOMM paper on estimating path properties
and my previous note on the subject for more information.)

> > I'd also wonder why tuning the min RTO is needed (except to recover
> > from quirky packet loss in the few initial RTTs), since modern TCP
> > does a pretty good job of recovering packets without needing timer
> > tuning.
> 
> I'm not sure what you mean by "without needing timer tuning".  RTO
> tuning has been part of TCP since the '80s.

I took Gorry's comment to mean that we have designed a number of loss
recovery techniques that work independently of the RTO (e.g., fast
retransmit, newreno, sack, ...) and so his presumption seems to be (and
my assumption is) that these techniques should be fixing the lion's
share of the loss.  Therefore, trying to somehow 'optimize' the RTO
seems somehow a bit dubious.

allman



-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 185 bytes
Desc: not available
Url : http://oakham.cs.ucl.ac.uk/pipermail/iccrg/attachments/20071025/a28fa576/attachment.bin