[tulip-bug] transmit timed out + too much work during an inte rrupt problems

Donald Becker becker@scyld.com
Wed Aug 14 22:10:01 2002

On Wed, 14 Aug 2002, Gehrken Klaus wrote:

> Here is some additional infomration: output of cat /dev/proc/net  and output
> of tulip-diag -af on the living interfaces (the last three days we did not
> had the "hangup situation").
> A real strange thing is, the catalyst ports (all fixed at 100MB-FD) report
> no errors at all ?!

This type of problem generally isn't realated to packet errors, unless
there is a duplex mismatch.

> > You can increase the max_interrupt_work parameter, but this likely the
> driver protecting the system from interrupt thrashing.  During the
> no-interrupt periods the NIC runs short of receive buffers.
> Hm, its on the default of 25 (what is an event? carrier transition, rx
> packet ?)

It's events handled per interrupt.  There shouldn't be much of a problem
setting the value higher.

You might consider re-compiling the driver with a higher setting for
RX_RING_SIZE.  If you are doing just network traffic processing, a
setting of 256 OK.  In that case you can set mac_interrupt_work=1000.

> What are the risks increasing it?

Bad latency for other drivers, perhaps impacting disk bandwidth
(slightly) and real-time response for other devices.  But in this case
those are not issues.

Donald Becker				becker@scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Second Generation Beowulf Clusters
Annapolis MD 21403			410-990-9993