Re: TCP limitations (was Re: ISCSI: Urgent Flag requirement violates TCP.)

[Vern Paxson <vern@ee.lbl.gov>]

> Could you please elaborate - or send me a reference?

The formula is just a summary of how TCP's congestion control, rate-halving
on loss / linear increase on no loss, behaves if each packet has an independent
probability p of being lost.  The fact that the packet has to be retransmitted
doesn't enter into the math - the extra bandwidth that retransmission requires
is a minor effect compared with the rate-halving.

What really kills you (for high performance over a large bandwidth-delay
product path) is the linear increase phase.  It takes a *whole* lot of
RTTs to ramp up your rate to a very high speed, and if during any one of
those RTTs you suffer a loss, you're back to the halved rate (or worse)
you were trying to recover from.

		Vern