> > TCP is still bursty over high bandwidth-delay product links while inflating
> > the congestion window. Basically everything in the window is trasmitted at
> > once, and when ACKs arrive, the process repeats with a larger window. Thus
> > larger windows cause the burstiness to proceed for longer amounts of time.
>
> What do you mean by "trasmitted at once?" TCP does not send the whole window
> all at once. Do you mean the case when after the link is idle for some time,
> TCP can send the whole window at once? I think this has been discussed before
> and the problem is an implementation bug. Joe Touch has a note on this.
>
> Or do you mean something else? Note that congestion window (cwnd) is inflated
> based on incoming ACKs. The rate of increase is determined by the rate of
> incoming ACKs. And for every ACKs, it can send up to 3 segments. This is
> the burstiness. And cwnd is reduced back to 1 (or 2 or 3 segments depending
> on whether the implementation uses larger slow start cwnd) after the link is
> idle.
>
> One exception is when there is segment dropped. But as I mentioned
> previously, I think it is not as bad as it used to be. There are other
> issues, say ACK compression, ... But I think a larger window does not make
> this worse than before. Please elaborate.
I am talking about burstiness in slow start, while the congestion window is
growing. When doing slow start, the small initial congestion window is
transmitted fairly rapidly, then waits for ACKs. When the ACKs arrive, they
arrive fairly rapidly as well, and the congestion window grows and again the
same process occurs. If and when the window grows large enough that an ACK is
received for the start of the window right after we send the last window
chunk, the burstiness stops. But if the window never grows this large, then
this equilibrium won't set in, and the connection remains bursty.
I believe that is the type of burstiness Mukul was referring to. This is
exacerbated when using high b*d links because many segments can be in the
pipe, necessitating a larger window.
In any case, no, TCP over high b*d links is not bursty at equilibrium, but
with too small a window the connection appears very bursty unless rate-based
clocking is used. Neither do large windows cause burstiness - too small
windows cause burstiness. But it takes longer for the ideal window size to be
reached when windows are large.
-- Zachary Amsden [email protected] (650) 933-6919 09U-510 Core Protocols
This archive was generated by hypermail 2b29 : Fri Feb 25 2000 - 01:22:26 EST