On 2007/10/19 12:26, Russell Fulton wrote: > It has become very clear now that one of the significant bottlenecks is > interrupt handling in the kernel.
I am pretty sure that PF's cpu use is counted in interrupt%. Are you aware of the ruleset optimizer in pfctl? It reorders the ruleset to take advantage of skip steps while keeping the same meaning. Before 4.2, it had to be enabled manually (with the -o flag to pfctl); now it's turned on by default and can be enabled/disabled from pf.conf. This is apart from the various other improvements to PF and the network stack (which still help when the ruleset is optimal), these have had very positive feedback: http://marc.info/?l=openbsd-misc&m=119257298926670&w=2 Some benchmarks on Sun X4100 have also been posted recently, seeing c.500kpps without packet loss: http://www.layer17.net/openbsd-router-intro.html http://www.layer17.net/openbsd-test-rfc2544-throughput-latency.html > We are using IBM X series hardware that have Intel NICs (em) on the > motherboard ( and 64bit PCIX bus). We also have some NIC cards that use > Broadcom chipsets (bge). em(4) sk(4) and most bge(4) have been fine for me, but I've had problems with dropped frames on "Broadcom BCM5704C" rev 0x10, BCM5704 B0 (0x2100)
