Re: Auto-neg or not?

On Thu, 26 Sep 2002 [email protected] wrote:

> Both the uplink and the card were forced to 100/full and the results
> were less than pleasant.

Hmm, they shouldn't have. Forcing them to the same speed/duplex 
combination should work.

> Most of the worries come from our experience with Sun hardware, which
> definetly has problems with it's auto negotiating.

My fellow Sun admins told me the same thing...

> Thanks for this one, I've looked at it and I think we're going to try and 
> double it ( 64 ).  These machines are dual AMD 2100 CPUB boxes with 2 GB's 
> of ram.  They're high traffic web servers, basically serving everything 
> via squid.  If anyone has a better suggestion instead of 64, please let me 
> know what I should set it to.

If you have SMP boxes and use IO-APIC, the effect is not so big as during 
the interrupt work in the 3c59x driver the other interrupts are not 
disabled but eventually processed on the other CPU. The only negative 
impact would be on devices that share the same IRQ line with the NIC.

As expressed in the older messages, there is no rule for computing a 
"good" value for max_interrupt_work; any value that doesn't produce the 
error messages is "good"; there is also no rule for the value to be a 
power of 2 (as you suggested 64 :-)). It's also suggested to increase 
the size of the RX ring, but you have to modify the driver source and 
recompile it - you have plenty of RAM for some more (tens) of packets.

-- 
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: [email protected]

_______________________________________________
vortex mailing list
[email protected]
http://www.scyld.com/mailman/listinfo/vortex