Date: Mon, 2 Oct 1995 13:25:34 +0100 (MET) From: Luigi Rizzo <luigi@labinfo.iet.unipi.it> To: hackers@freebsd.org Subject: Unreliable internet and RTT estimation code in BSD Message-ID: <199510021225.NAA23487@labinfo.iet.unipi.it>
next in thread | raw e-mail | index | archive | help
This is a followup to my previous posting on high packet losses
(10..60% and more) which I am experiencing between Europe and the US.
My first thought was that such a large loss rate should almost
completely defeat the algorithms used by TCP to determine the RTT,
to the point that ftp-ing a file larger than 100K takes a huge
amount of time if it succeeds at all.
To see what was going on, I have modified netstat on my FreeBSD
1.1 system to show the values of rtt/srtt/ rttvar from the tcp
protocol control block (a simple one-line addition to inet.c in
the BSD netstat sources: in protopr() I added the line
printf("rtt/srtt/rttvar %7d/%7d/%7d ",
tcpcb.t_rtt, tcpcb.t_srtt, tcpcb.t_rttvar);
I then started an ftp connection from overseas with hash marks on,
and, while the transfer proceeds (*very* slowly), netstat shows
the following numbers (these are not regular samples; I ran netstat
every few seconds by hand):
rtt srtt rttvar comment
------------------------
3 8 2 beginning of the connection. some 50KB go
5 8 2 through.
8 8 2
11 8 2
17 8 2 no data are flowing.
22 8 2
2 0 3 perhaps a packet or two have been sent.
17 0 3
0 0 3
5 0 3
86 0 3
0 0 3
...
0 0 5 the transfer seems to have blocked; but pinging
the remote host still gives about 50% success.
I get these numbers for a couple of minutes.
... here I was typing this message
25 15 3 and after, this is what i get.
...
0 15 3 again, no data are flowing.
0 0 6 and to my surprise, srtt goes back to 0!
...
0 23 6 and then up to 23...
...
0 0 24 and then back to 0 again!
There are two things which surprise me:
1) srtt goes to 0 so quickly, and stays there for some time, despite the
fact that the actual rtt is large;
2) rtt stays at 0 while no data are exchanged (and there should be
pending data/acks somewhere).
I am not very familiar with the way RTT is estimated and it implementation
in the BSD code (I guess I'll have to study it); however, the above
numbers look wrong to me.
So I was wondering about the existence of studies on the performance of
the RTT estimation code in the BSD implementation of TCP, especially in
presence of large packet losses.
Any pointers and comments are mostly appreciated.
Luigi
====================================================================
Luigi Rizzo Dip. di Ingegneria dell'Informazione
email: luigi@iet.unipi.it Universita' di Pisa
tel: +39-50-568533 via Diotisalvi 2, 56126 PISA (Italy)
fax: +39-50-568522 http://www.iet.unipi.it/~luigi/
====================================================================
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199510021225.NAA23487>