On Apr 26, 2006, at 9:44 PM, Andy Ruhl wrote:

> On 4/26/06, Francis Koclas <netbsd@geekrants.org> wrote:
>> I have a couple questions related to these numbers;
>>
>> 1) What is causing the rather slow speeds in I and II? The Qube2's  
>> CPU
>> remains 100% idle while it forwards. In linux (debian 3.1 with kernel
>> 2.4.28 i think) it makes the Qube2 roughly 80% idle and the speeds  
>> get
>> bumped to a little bit more than 6 MB/s.
>>
>> 2) Is this a problem specific to the built-in 21143 chips?
>>
>> 3) Why is one way slower than the other? Even if I swap the
>> ifconfig.tlp0 and ifconfig.tlp1, and swap the network cables, the  
>> same
>> thing happens, so tlp0 is not any faster than tlp1 or whatever.  
>> The NICs
>> on the workstations are very capable and they're not the  
>> bottleneck here
>>   (Sun Gem in WorkstationA, 3c905C in WorkstationB). As you can  
>> see they
>> can pretty much saturate 100baseTX..
>>
>>
>> 4) Is there any cobalt specific tweaking to improve network  
>> performance?
>>
>> 5) What are good, known to work PCI NICs for the Qube2? I tried a  
>> 3Com
>> 3c905C and I only get a couple KB/s with it.. Also, anyone tried dual
>> ports cards? I would ask for quad ports too, but I doubt those would
>> actually fit in there.
>
> There's a known performance problem with the PCI bus on the Qube with
> NetBSD. Unfortunately, the fix is not known as far as I know. The
> problem either doesn't exist or isn't as bad with Linux.
>
> I don't know much more

I've always felt that the chipset needs to be re-initialized during  
boot, instead of relying on the rom initialization; the linux version  
for the qube2 does this re-initialization.  One symptom of this are  
the lights on the ethernet ports: on my qube, one of them doesn't  
have connected light turned on due to incorrect autosense detection.

One Linux, eeprom fixups are performed to fix this.

Alas, I have never had time to look into fixing this myself.

Byron

p.s.  my original message

<http://mail-index.netbsd.org/port-cobalt/2004/09/30/0000.html>