nForce Watchdog problems

[Neil Joseph Schelly]

On Thursday 26 October 2006 05:26 pm, Ben Scott wrote:
>   You might tell us something about the hardware.  :-)  Such as:

I guess I was looking for more general experience to test the waters... 
anyway, the servers are brand new Penguin Computing Altus 1600 SATAs.

Here's a link to their configure page on their site, which includes a tech 
specs PDF link in the upper right.  That should have all the motherboard 
information. 
http://www.penguincomputing.com/index.php?option=com_content&task=view&id=335&Itemid=483

Ultimately, these are dual Opteron 2210 dual core processors on an nForce 
motherboard, so the forcedeth drivers are for the nForce chipset NICs, 
onboard.  

I've found a few helpful hints about using ethtool to disable tso, which 
offloads some network traffic checksumming functionality from the CPU to the 
NIC itself.  I've found suggestions that it's buffers overflow and the card 
just stops working.  I was particularly puzzled by the fact that a cold boot 
didn't even fix a machine that had worked fine for weeks, but only recently 
been introduced to servicing a load.

I think that buffer was overflowing and staying overflowed through reboots, 
since the NICs stay "on" to be able to listen for wake-on-lan functionality.  
By unplugging the power supplies, the NIC lights went out and the machines 
could connect to the network again.  So now I've setup a script to run 
ethtool -K eth0 tso off for each NIC when it gets initialized and I hope that 
will work.  I found at least a few things that suggest it would/migiht/did 
for someone else getting similar errors on similar hardware.

This ringing any bells for anyone here?  I have to launch this hardware soon 
and I should know with 24 hours if this fix works, since no server has 
maintained a load for that long yet without this happening.  I would just 
feel more comfortable I guess if there was more than a few search results in 
Google to bounce off of.

Thanks,
-N