TCP/IP hang

Maximilian Rixius maxi@parsytec.de
Tue Jul 28 09:02:58 1998


Hi,

I've got a problem with two intel DK440LX boards (i82557 chips) running
under stress tests. With the v0.99B driver the network connection will
not hang up completely (as it did some versions before), but after an
transmit timeout the network performance test's TCP/IP connection seems
to be unrecoverably stuck with netstat reporting different states on
both ends. This results in

- test stuck with bw_tcp of the lmbench suite or
- test program dies with netperf (client side only).

This happens after several hours (e.g. eight) of continued testing.
Afterwards both previously established (but at the time unused)
connections as well as establishing new connections is no problem.

Could this result from a driver problem or is it more likely to be a
TCP/IP problem?

The enviroment: 2x DK440LX with dual PentiumII 333MHz, 128/256MB,
Linux 2.0.33 with eepro100 v0.99B (testing v0.99C right now).
Additionally to ethernet performance testing, hard disk tests are run at
the same time. One of the nodes bootes with root as NFS.

Spuriously there are "eth0:SMP simulaneus entry of interrupt handler" in
the logfile, but this does not seem to produce trouble.

Hope somebody can help,

  Maxi

--------------------------------------------------------------
Maximilian Rixius                             maxi@parsytec.de
parsytec Gmbh                   private:  maxl@pinguin.oche.de

          I'm the fellow throwing peas into the air
--------------------------------------------------------------