* Ronnie Garcia <[EMAIL PROTECTED]> [2007-01-22 21:10]: > Ronnie Garcia a icrit : > >I recently switched one of our firewalls from Linux to oBSD 4.0. > >Its handling approx 8-9 kpps (in+out) on both interfaces. It has a > >D-Link DFE-570TX quad ports NIC (dc driver), two ports are used. > >On Linux, the CPU was loaded at approx 20% when, and on oBSD, its > >actually loaded at ~30%. No big deal, but on Linux we had queueing > >(shaping) with TC/HTB, whereas ALTQ is not (yet) enabled on oBSD. > > > >The CPU usage is almost only "interrupt", as you can see on this top > >output : > > [The rest of the message is left bellow for the record.] > > I can now tell that i have the exact same behaviour with bsd.mp. > > I'm graphing a lot of kernel/pf variables with cacti, and i'm clearly > seeing the box maxing at 15k interrupts/s.
that is not necessarily a problem. > I'm raising 15k interrupts/s when the box is routing approx 13k pps and > then the CPU is at 50-55%. at 13k pps you definately want good nics which have proper interrupt mitigation. most gigE NICs fall into that category; sk, msk and em fall definately into that category. > When i disable pf (pfctl -d), the CPU downs to ~40% but the interrupts > rate does not decrease. This means that the high interrupts rate is due > to network activity, and not to pf. the whole fwding and pf run in (partially soft-) interrupt context and are counted as interrupt time. > I might try with an Intel Pro/1000MT quad instead of the D-Link > DFE-570TX quad to see if my problem is the NIC or the PCI bus/chipset. that will do. -- Henning Brauer, [EMAIL PROTECTED], [EMAIL PROTECTED] BS Web Services, http://bsws.de Full-Service ISP - Secure Hosting, Mail and DNS Services Dedicated Servers, Rootservers, Application Hosting - Hamburg & Amsterdam
