Re: [nsp] DS3 rtt

From: Kai (kai-cisco-nsp-trap@conti.nu)
Date: Fri Feb 04 2000 - 12:44:43 EST


At Friday 12:17 PM 2/4/00 , someone wrote via private email:

>Kai,
>
> > If this is a link transporting public IP traffic, you can safely assume
> > that all Windows machines using it have a measly 8K window, throttling
> > the performance to roughly 8K per RTT (in reality its slightly more),
> > yielding you 533KB/s @15ms and 320KB/s @25ms.
>
>Can you explain your parenthetical comment that the performance is slightly
>more than window/RTT?

I was still thinking about that hastily made comment, and I think I am
quite incorrect.

I believed that the optimum would be a RTT shortened by the time it
takes to transmit the entire 8K window "in flight".

In reality, the RTT would be measured from the first segment of those 8K ,
until the beginning of the first segment of the next 8K going out.
If the receiver implements delayed ACK, or doesn't even send an ACK for
every segment received, the returning (delayed, or spaced-out) ACKs will
then actually "clock" the transmission of every segment in the "next" 8K,
quickly leading to dispersion of packet transmission over time: the first
segments that add up to 8K would go over the wire tightly spaced, possibly
back-to-back. But immediately beginning with the next segments, the
receiver will see ACKs in not so evenly spaced manner and send out
segments accordingly, but never exceeding windowsize/RTT.

It would be interesting to see how the dispersion develops over time,
even in a tightly controlled environment with a known, 99.9% predictable
propagation-RTT.
My guess is that the dispersion pattern (density of packets over time)
is semi-random, if not fractal in nature, with results varying greatly
depending on operating system for transmitter and receiver.

bye,Kai



This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:12:09 EDT