if_bridge + polling get lower througphts

[Andrew Thompson]

On Wed, Mar 15, 2006 at 06:16:29PM +0800, Chih-Chang Hsieh wrote:
> We have a FreeBSD 6.1-PRELEASE box.
> 
> It runs with 2 em NICs and uses if_bridge + IPFW + ipfilter + pf.
> 
> This box usually gets a very high interrupt rate >90%.
> 
> By using netstat -I em0 1, we see:
> 
> input (em0) output
> packets errs bytes packets errs bytes colls
> 15164 0 9549410 16842 0 12132262 0
> 15331 0 9220723 17466 0 12354486 0
> 16486 0 10089165 17926 0 13256014 0
> 14936 10 8363433 17090 0 13367271 0
> 14712 0 8449705 16701 0 13443079 0
> 
> 
> When we enable the polling on em0 and em1 by ifconfig, netstat -I em0 shows:
> 
> input (em0) output
> packets errs bytes packets errs bytes colls
> 10910 0 6629176 11467 0 7681346 0
> 10951 0 6916042 11397 0 7954205 0
> 11281 0 7437813 10804 0 7390386 0
> 10711 0 6308033 11605 0 8400405 0
> 10807 0 6270716 10973 0 7700748 0
> 
> the interrupt rate decreases (<30%) but we get lower throuphts.
> 
> net.link.bridge.ipfw: 1
> net.link.bridge.pfil_member: 1
> net.link.bridge.pfil_bridge: 1
> net.link.bridge.pfil_onlyip: 0
> 
> How could we tune the box to reach higher throughputs

In the above setup you are running four packet filters, ipfilter, pf and
IPFW twice (ipfw=1 is an extra layer2 run). More specifically the packet filters
are actually invoked _10_ times per packet, in_if -> bridge_if -> out_if
per pfil_hook, plus the extra ipfw layer2.

This explains your high cpu usage, you may want to reorganise the
filtering to optimise it.


Andrew