[Soekris] high CPU usage for network interrupts on 4801

[Jan Hoevers]

Trevor Talbot wrote on 16-9-2007 22:26:
>> I've tried downloading a large file to the NTP machine at a much 
>> higher speed (6800 kbit/sec), which takes about 75% CPU for interrupt 
>> processing. I guess the small size of NTP packets is inefficient.
> Yes, figures like 40Mbps are a "best case" scenario with full-size 
> packets.  With small packets, the limiting factor is packets per 
> second, as each one incurs processing.
Ok, I understand. No 40 Mbps of NTP traffic can be expected. My upstream 
connection wouldn't handle that anyway.
But I assume that downloading a 100 Mb file by FTP involves full size 
packets.

>> However, this 75% is stil much higher than I would expect.
> That seems a bit high to me too.  What are the other 5 ethernet ports 
> doing?
Four of them doing nothing (unplugged for the occasion). The last one 
carrying two ssh sessions to monitor the top displays. Traffic nothing 
really compared to the NTP packets flowing.

>> I've tried moving the NTP machine from an ethernet port on the 1641 to 
>> one on the 4801 board in an attempt to avoid the PCI bus, but it makes 
>> little or no difference. (Maybe I'm not avoiding the PCI bus this way, 
>> I'm not sure.)
> You're not, it all goes over the PCI bus either way.
Ok, that's clear too.

> I'll let others cover FreeBSD specifically, since I haven't done any 
> testing with recent versions.  One thing that would help is getting a 
> handle on what pps rates you're actually seeing, though.  I don't know 
> if pfsense measures it, but you can get an idea with:
>      netstat -I sis0 -w 1
I believe something like 1500 pps, plus 1500 replies going the other 
way, but it's a guess. I hope to get back to you with some measurement, 
but I'll have to wait for the next burst.

Thanks for answering,
Jan