Re: 1000ms delay in networking stack or driver, new bug?

Henner Eisen (eis@baty.hanse.de)
20 Sep 1999 22:50:08 +0200


>>>>> "kuznet" == kuznet <kuznet@ms2.inr.ac.ru> writes:

kuznet> Hello!

>> It is very odd in that it isnt card dependant. It seems to be
>> more like a qdisc problem somehow affected by interfaces going
>> up and down and up again

kuznet> 1 sec is ping interval. Please, make "ping -i 3" to check
kuznet> this. If the time will be 3 sec, the guess is correct.

kuznet> And make tcpdump both on wire (on ethernet it should be
kuznet> third host) and on this host.

kuznet> Alexey

As such problem have been observed with isdn_ppp for quite I while
(and number of observations is increasing quickly as lots of isdn
users are currently switching to 2.2.x) we have already several
detailed reports. To summarize the characteristic observations:

- People have verified by means of tcpdump that the delay occurs at
the ping senderīs side: pings arrive at responders side with delay and
are replied immediately. And after beeing replied by the responder,
the echo is also reported by the sending ping command immediately.
- Delay is always a multiple of second (say N) (+ the small additional
delay which also occurs with a non-broken connection).
- There seems to be a 'law of preservation of delays' :-)
A single ping command shows delays of multiple of seconds (say N seconds).
If two ping commands are run concurrently, the delays (say n_i and m_i)
of each ping sum up to N (they fulfill the preservation law n_i+m_i = N :-)

There are also some other observations which are not so consistent:
Some people report that disabling tcp timestamps did help, others report
that it did not help. Some people also report that ifconfig'ing
netdevice down and up again makes the problem more likly to appear.
Others donīt report differences.

I myself used isdn_ppp continuously with 2.1.x since x around 30.
I donīt know exactly when I first observed the delay problem but as far as
I remeber it was roughly around the time when the packet scheduler
was merged.

Henner

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/