Re: [pmtud] Improvement for the current PMTUD mechanism

At 03:23 p.m. 10/10/2005, John Heffner wrote:

> > >Using the algorithm specified in section 7.2 of
> > >draft-gont-tcpm-icmp-attacks-04, imagine the following scenario.  Some
> > >packets take a path with pmtu 1500, and every nth packet takes a path with
> > >pmtu < 1500.  So the first n-1 packets get through fine and are
> > >acked.  The nth packet will be lost and an ICMP PTB message sent back.
> > >The segment is fast-retransmitted and will get through, causing the
> > >delayed response to the ICMP to become no response.  This will continue
> > >indefinitely, with a loss rate of 1/n.
> >
> > Good point. I think the simple answer to this issue is to disable "fast
> > retransmit" when there's a pending ICMP message. What do you think?
>
>This concerns me because it could have real side effects beyond pmtud.  This
>might be a topic for tcpm.

Actually, this cannot have interoperation problems. IIRC, TCP congestion 
control is not mandatory.
And even then, fast retransmit is not specifically congestion control. If 
anything, I'm proposing to slow down the packet rate, rather than 
increasing it. (i.e. being more conservative).

> > (As a side note, I'm not sure the behavior you describe would happen in
> > practice. Particularly because of TCP's slow-start algorithm. As there are
> > only a few segments "in flight" just after connection establishment or a
> > period of idleness, in many cases there won't be enough duplicate ACKs for
> > fast retransmit to be triggered. Anyway, as the fix to the point you raised
> > is pretty simple and inoffensive, that's the easier way to go).
>
>It does feel like a corner case, probably difficult to trigger.  But I think
>it would be hard to prove it never happens, particularly if you consider all
>possible future behaviors of TCP (with modified slow start or retransmit
>algorithms).

Will address this in the next revision of my draft.

Kindest regards,

--
Fernando Gont
e-mail: fernando@gont.com.ar || fgont@acm.org

_______________________________________________
pmtud mailing list
pmtud@ietf.org
https://www1.ietf.org/mailman/listinfo/pmtud