Re: [tcpm] initcwnd updates?

[Mark Allman <[e-mail]>]

> if I am not mistaken there is already data that shows IW=10 and
> unpaced easily leads to interesting microburst events in switches.

I am OK with pacing larger windows.  But, I will also say that an
adaptive scheme covers the sort of situation you describe.  I.e., it
gives a way to adapt to a lower IW if IW=10 is causing lots of
problems.

> This is IW, if your losing a packet during the IW the network is
> already in a bad state, do not make it worse by increasing the
> amount of unacknowledged traffic to enter flight.

Right.  Agreed.  But, we have to start somewhere (the IW).  And, if
we don't invent a better system then we're going to keep making the
IW higher over time.  That's how things work.  And, the IW is going
to sometimes be an issue---regardless of where we set it.  Right now
our process is someone proposes a higher IW, there is a little
evidence gathered and then we decide either "it's in general OK" or
"let's not increase it".  And, we decide that for all hosts and
connections and situations.

The proposal is ... We know there is no ideal one-size-fits-all for
IW across situations and contexts.  So, why continue to pretend that
we can pick such a number?  Why don't we focus on something like the
anticipated impact of starting a connection on the network?  I.e.,
saying something like "if we use IW=X then we anticipate---based on
previous experience---that the with probability 0.95 this IW will
not lead to loss".

> Rather than state this as a percentage perhaps
> state it as any loss during IW MUST be acted upon.

As far as I am aware nobody has ever said loss during the IW (or any
W) would not be acted upon as congestion.  Clearly when we directly
observe congestion then we'll react to that in the standard manner.
So, of course what you say is true.

> (Note that 5% of 10 packets is only half a
> packet, but I believe this number is being used over mutliple
> IW's, so thats every other IW=10 experience loss, which would be a
> poor situation.)

No- not that at all.  Both 9040 and what I said earlier involve a
two stop process that doesn't measure the packet loss rate.  (1)
Does the connection experience any loss in the IW (could be 1
packet, could be all IW packets)?  And, then (2) What percentage of
connections have loss that sort of loss?

So, say you have 100 connections (to make the math easy).  And, say
the IW is 10 packets.  So, in total that is 1000 packets in the IWs
of 100 connections.  So, the 95% could be triggered by as few as 5
packet losses (1 loss in 5 connections) or as many as 50 packet
losses (losing all 10 packets in 5 connections).  So, your packet
loss rate will range from 0.5% to 5%.

So, this is not at all anything like one out of two connections.
This is 1 out of 20 connections (5%).

And, note, I didn't mean to suggest 95% was the right number.  I
think I specifically said this would be the decision point.  I am
not arguing for 95%.  But, let's not confuse things that using this
95% would tolerate a loss in every second connection.

allman