Re: [conex] Expiration of credits

[Mirja Kühlewind <[e-mail]>]

Hi,

one more question:

If I send a large number of credits in Slow Start, lets say 50, and than have 
a quite moderate TCP like Reno with drop tail. That means I usually see only 
on loss everytime my cwnd exceeds the link capacity. Lets say my max cwnd is 
100 the i reduce to 50 after a loss and it takes another 50 RTT until I see 
the next loss.

That means it take 50*50=2500 RTTs or 1500*50*50=3.75MByte until my credits 
are gone even if I don't send any further ConEx markings. Why should I send 
any real ConEx marking in this situation (and pay twice)?

But with ConEx we actually would like to have the real ConEx signal because 
only this reveals the current congestion level on the link!

Mirja


On Wednesday 26 October 2011 18:33:26 John Leslie wrote:
> Mirja Kuehlewind <[e-mail]> wrote:
> > I guess a cap is even more complicated. An expiration would somehow
> > depend on RTT but the num of credit you want to send (at once)
> > depend on the congestion control mechanism in the endsystem.
>
>    IMHO you should never depend upon a credit lasting more than one
> RTT.
>
>    That said, credits are intended for start-up -- and in many cases
> currently, you will need more than one RTT to infer that a packet
> was dropped (because ECN marking wasn't enabled at the bottleneck).
> There may be cases where you should decide whether to issue additional
> credits while waiting for a timeout -- and there are certainly cases
> where you _don't_ want to issue such additional credits.
>
> > If I increase the cwnd by 100 packet in one RTT I might want to
> > send 100 credits
>
>    I cannot imagine such a case. Credits need never exceed the greatest
> plausible loss. I cannot imagine increasing cwnd by 100 when I expect
> 100% packet loss.
>
>    IMHO, credits are not generally appropriate for any situation where
> you are increasing cwnd -- but I'm sure _someone_ will find a case
> where it would help... :^}
>
> --
> John Leslie <[e-mail]>