Packet loss with traffic shaper and routing

[tpeixoto at widesoft.com.br]

Ferdinand Goldmann wrote:
> tpeixoto at widesoft.com.br wrote:
>> Hello.
>>
>> I did that and compiled the kernel.
>> Then I restarted the system and enabled sysctl kern.polling.enable=1
>>
>> It seems that it has no effect in the system. Maybe bge driver doesn't
>> like polling?
> 
> At least from a quick glance in the polling(4) manpage I cannot see that bge
> is among the supported devices.
> 

You're right. I read that too but I found something in Google 
(http://unix.derkeiler.com/Mailing-Lists/FreeBSD/net/2003-08/0241.html) 
and wanted to give it a shot.


> If you want to use polling, I suppose that you need to enable it via ifconfig,
> too:
> 
>      polling
>              If the driver has user-configurable polling(4) support, select
>              the polling mode on the interface.
> 

It seems to be default when you enable polling, then you can switch it 
off and on with -polling and polling via ifconfig.


> 
>> At this moment, I'm getting more than 50% interrupts and 20% packets lost.
>> I also disabled HT in BIOS and the interrupts are now passing 80% mark.
>> Don't know what else to do. Aren't these cards supposed to work at
>> 100Mbits or 1Gbit? They are failing with 12Mbits traffic on a 100Mbits
>> LAN. Something is wrong and I am having a hard time trying to identify the
>> problem.
>>
>> Thanks for the hints, anything else would be greatly appreciated.
> 
> Several wild guesses from my own experiences here:
> - SMP + networking in 5.x does not work too well, using em(4) I experienced
>   VERY poor performance (only ~5MB/s over a Gbit link)
> - Try upgrading to 6.x (as others have already suggested). I experienced all
>   kind of weird problems with 5.x, and although there is no proof that the
>   problems were actually related to 5.x, 6.x seems to work better.

We did. Now we're running 6.0-RELEASE.


> - What's the value of nmbclusters? Have you checked netstat -m? Do you see
>   memory requests for network memory denied?

AFAIK, nmbclusters aren't informed properly on SMP systems. Memory 
requests are always 0.


> - 50% interrupts on such a fast machine is quite high. I currently experience
>   about 30% interrupt load using two em(4) cards, shaping for about ~2000
>   clients on a 3.8GHz Xeon.
> 

Please, take a look in my previous post.
I guess the problem lies with IPFW and dummynet.
How do you shape your clients?

Here we have (for each client):

ipfw pipe 1 config bw 512Kbit/s
ipfw pipe 2 config bw 512Kbit/s
ipfw add pipe 1 ip from any to any mac any 00:11:22:33:44:55 in
ipfw add pipe 2 ip from any to any mac 00:11:22:33:44:55 any out


> Kind regards

Thank you.