[vortex] causes 3c905c jitter ???

Donald Becker becker@scyld.com
Tue Dec 18 23:17:02 2001


On Tue, 18 Dec 2001, Stefan Doehla wrote:

> At the end of my mail is something about my settings ...
> >>Can the 3com networkcard lock the PCI bus for some microseconds? Uses it 
> >>a fifo that's not checked before writing (due to performance gains)?

No, the no single read or write transaction to the chip should introduce
delays, unlike the command FIFO on video chips.

Almost all data is transferred with the NIC as a bus master.  While the
Tornado is capable of bursting an entire packet to or from the NIC in a
single v2.1 PCI burst transaction, the NIC will yield the bus when any
other master is requesting access.

> > What chip, and what driver version?
> > 
> > You can change various settings to limit the PCI burst size, but I don't
> > know what you are seeing that's a problem.

Use one of the tools to tune the PCI bus parameters.  While it's also
possible to limit the burst size using the NIC registers, this won't
likely improve your response latency.  The downside of limiting the
burst size is that you've just increased the bus overhead, creating a
higher overall latency.

> root@labpc7:/var/tmp# ./vortex-diag -aa
> vortex-diag.c:v2.05 5/15/2001 Donald Becker (becker@scyld.com)
>   http://www.scyld.com/diag/index.html
> Index #1: Found a 3c905C Tornado 100baseTx adapter at 0xa400.

A normal 905c, probably a "cx" chip.

> 00:0c.0 Ethernet controller: 3Com Corporation 3c905C-TX [Fast Etherlink] 
> (rev 74)
...
> 	Latency: 32 (2500ns min, 2500ns max), cache line size 08

These may be modified, but 32 is a standard setting.


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