Re: [conex] Expiration of credits

[John Leslie <[e-mail]>]

Mirja K?hlewind <[e-mail]> wrote:
> 
> one more question, we need to answer for the TCP modifications draft:
> 
> In the draft we propose a way how to use the credits bits in Slow Start.
> We assume that the credits are valid for a couple of RTTs.

   The credits, IMHO, are intended to cover for one RTT plus processing;
but nobody knows how long one RTT is...

> But I assume that those credits should not be valid for ever, as
> otherwise I could just send credits whenever I have 'congestion
> volume tokens' available and not care about any conex markings
> later on.

   I doubt that would prove useful.

   It is conceivable that someone somewhere might attempt to match
"credits" to congestion over a longer period; but I can't see this
would accomplish anything useful to the party counting them.

   A more likely practice would be to accrue an allowance up to a
maximum (which might be as low as a half-dozen packets worth or
might range up to thousands), and check for a "user" exhausting
that allowance. It is far too early, IMHO, to say what these rules
might be.

   Down the road a bit, your upstream may want to traffic-shape so
as to fit within a contractual allowance at each peering point. The
maximum you could accrue might be a function of traffic patterns
at these peering points.

> I believe we want to avoid this case. But I don't think that any
> kind of expiration of the credits bit is addressed yet (in the
> abstract mechanism draft)...?

   I don't believe that an "expiration" rule will prove helpful;
and I certainly don't believe we should try to define one at this
stage of the design.

   And I'm not sure we _could_ attach a useful definition to "valid"
here -- packets considered to be part of an "abusive" flow may get
dropped somewhere along the path, but that would be a function of
the "flow", not the individual "credits", and I can't imagine a
practical way to "expire" the "credits".

--
John Leslie <[e-mail]>