Francois Romieu wrote:
Ben Greear <greearb@xxxxxxxxxxxxxxx> :
[...]
I'm running tests at the ethernet frame (skb) level, so I would not
expect the IP offload to help much.
Do you have any performance numbers for your 8169 cards for me to
compare against?
Gross numbers:
- with a receiving amd64, nttcp should saturate the GigaE link;
- halved bandwidth for Tx on the same setup. Untested since changes in -mm;
- 10% cpu on a 100bt link;
- TSO saves 5~7% of the cpu when it can be used.
Cool, I am glad to know the chipset can handle it at least...
I'll get this in a bit, as I have locked up the keyboard again. This time,
I was still using the non-NAPI code which had run over-night at around
5Mbps...
I tried a different network setup, and noticed that the 8169 was only
negotiating 10bt/HD. That would explain the poor performance numbers.
Call me dense but... Are the 5Mbps below in the same unit as the 30Mbps
in your previous message (are they bytes or bits) ? If yes something is
badly broken.
Yes, but I think now the problem is that it was stuck at 10bt/HD for
whatever
reason. There seems to be a way to lock the keyboard by trying to make the
NIC change it's negotiate settings. I am going to try to track down exactly
how this can be reproduced with minimum number of steps...
Thanks,
Ben
--
Ben Greear <greearb@xxxxxxxxxxxxxxx>
Candela Technologies Inc http://www.candelatech.com
|