Re: Networking slowdowns

Christoph Lameter (
Sun, 8 Sep 1996 06:44:20 -0700 (PDT)

More details:
- After the link is up a while the stoppages increase
- You can boost the effect by having another linux box
attached via ethernet to the box you are having the ppp
connection with. Work on that other linux box.

I had stoppages that lasted minutes.

On Sun, 8 Sep 1996, Eric Schenk wrote:

> Christoph Lameter <> writes:
> >I think I found a way to demonstrate networking slowdown.
> >
> >Set the MTU/MRU of a PPP Link to 3000. Then telnet into your
> >machine and cat a file of 10-20K.
> >The session will stall while you still can telnet to the machine in a second
> >session working just fine.
> >
> >Hope I did not shoot myself in the foot here...
> Ok. I tried this here with a test rig. Both ends running 2.0.15.
> I see odd "stalling" patterns in the telnet test, in that output
> is bursty. This is not entirely unexpected, since a 3000 byte packet
> takes a second or so to pass through the link, but much less than
> a second to display on the telnet screen.
> I do not see a complete "stall", that is, at no time have I been able
> to make the traffic stop completely. I have, however, observed pauses
> that last a few (4-5) seconds.
> Looking at the tcpdump I also see that the system is very rarely
> sending a full sized packet, which is quite odd. This means that telnet
> is not feeding characters to the PPP link fast enough to prevent
> a Nagle timeout, and occasionaly appears to have to characters ready
> to send for several seconds.
> As a test of this theory I ran FTP sessions across the same link.
> All FTP sessions ran a full speed, no stalling, with a sequence of
> full sized packets only.
> This makes me suspect something in either telnet, or the pty layer.
> Further evidence of this: when running a tcpdump in parallel with a telnet
> I observe that packets finish crossing the link several seconds before
> they get displayed by telnet.
> I'd be interested to know if others see similar results with this test..
> If so, then perhaps we can narrow the range of inquery into this problem.