On Thu, 9 Sep 2004, François Baligant wrote:
> Hi,
>
> We have a tg3 NIC under heavy load (30megabit/s, around 15k to 20k
> packet/sec)
>
> Under 2.6.7, the NIC behaves quite correctly (except errors see below)
> Under 2.6.8, the NIC's link will go down after some time and not come up
> again (rx_discards grow more quickly than in 2.6.7)
> Under 2.6.9-rc1, the NIC will go down from time to time also but recover
> after 5 seconds and link goes up again (dito)
>
> Should I worry about the errors in ethtool, like rx_discards ? What does
> it mean ? Can i tune something to avoid them ? (netdev backlog?)
I believe that when rx_discards increments, it means that the card
dropped the packet due to a resource constraint.
I've seen similar behavior (but at much higher data rates.) In my
case, enabling h/w flow control fixed things. If it's possible to use
h/w flow control in your configuration, then you might give that a try.
>
> Is there any way to lower the amount of interrupt needed to handle this
> load? (it's eating quite a lot of CPU)
> ....
That is a _feature_ of NAPI ;-)
--
Arthur
|