On Wed, 10 Feb 1999, Phil Karn wrote:
> Catching up on old email...
>
> > Kalman filtering would work if there exists in your signal a true
> >"tendency or behavior". Unfortunately, the RTT variations in response to
> >the variation of number of packets in flight are most of the time
> >uncorrelated. For example, increasing the number of packets in flight may
> >result in an increase of RTT as well as in a decrease of RTT. This is due,
>
> But there are some important special cases where it would work just
> fine. The canonical example, and one of interest to this group, is the
> slow bottleneck being used by one user (e.g., a dialup link). If TCP
> could detect by its own RTT measurements that increasing the cwind is
> just queueing up more and more packets at the bottleneck link, it
> could back off its cwind without having to deliberately drop packets
> at the bottleneck.
>
We extracted from the tcpdump traces collected by Vern Paxson the
coefficients of correlation between rtt and number of packets in flight.
We were able to analyse 14,218 connections over 809 Internet paths.
2296 connections were on slow links (~64Kbps). Indeed, there is a
strong correlation for slow links in comparison to fast links (T1+) : 25%
of the connections have a coefficient of correlation larger than 0.8.
However, we do have 30% of the connections with a coefficient of
correlation less than 0.2.
Worse than that, these coefficients are "dominated by outliers
(RTT spikes)" (Vern Paxson). For this reason, we also studied the
correlation between the signs of variations of rtt and number of packets
in
flight. Bad news !!!! Slow links do not exhibit globally stronger
correlation than fast links. Moreover, only 10% of the connections on slow
links exhibit coefficient of correlation larger than 0.6.
We have a paper in preparation describing these experiments. It
will be ready next week. If interested. please let me know.
Saad Biaz.
___________________________________________________________________
Saad Biaz (Abou Youssef & Rim) Lecturer
Texas A&M University Office : (409) 845-5007
Department of Computer Science Home : (409) 862-9135
College Station, TX, 77843-3112 Fax : (409) 847-8578
___________________________________________________________________
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 09:12:20 EST