[Web10g-user] difference in curCWND and SlowStart
mojgan ghasemi
mojgan.ghasemi at gmail.com
Sat Jul 12 12:20:45 EDT 2014
Hi,
Thank you for your response. You were right! the scaling was on, however
even after I turned it off the CWND in the two approaches differ, but the
values are much closer now (error rate is about 3% between my estimation
and the value reported by Web10G)
Thanks for your help!
Mojgan
On Fri, Jul 11, 2014 at 11:59 AM, <Valdis.Kletnieks at vt.edu> wrote:
> On Fri, 11 Jul 2014 11:50:39 -0400, mojgan ghasemi said:
>
> > I have a connection that according to Web10G, never exists the SS, and
> its
> > SlowStart variable is 235, meaning that the CWND has been increased a
> total
> > of 235 times.
> >
> > However, when I look at the packet traces of the connection using
> tcpdump,
> > I see more than 15K new ACKs being delieverd for this connection (even
> with
> > byte counting approach), which according to many RFCs such as 2581, each
> of
> > them must make the CWND to increase by one MSS. So why only increase CWND
> > 235 times?
>
> Is TCP Window Scaling in effect? If so, you could be seeing "steps" in
> the window. For example, say you start with a (hypothetical) window of
> 12, and a scaling of 3, and packets coming in that each increase the window
> by 3.
>
> packet internal window on the wire
> 0 12 1
> 1 15 1
> 2 18 2
> 3 21 2
> 4 24 3
> 5 27 3
> 6 30 3
> 7 33 4
>
> You get the idea. 15K ACKs and 235 updates works out to 63 acks per
> update - making me take a wild guess that you have TCP scaling of 6
> in effect?
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.psc.edu/pipermail/web10g-user/attachments/20140712/c51adea4/attachment.html
More information about the Web10g-user
mailing list