Thus spake Teun Nijssen (teun.nijssen@xxxxxx):
> Currently the TORy1 to TORy3 instances are slowly speeding up. The machine
> runs Linux version 2.6.26-2-amd64 (Debian 2.6.26-24lenny1), currently without
> any of the high performance tuning that was discussed lately. Even iptables
> does its usual work.
Please keep us posted on your total capacity (using a network monitor
like munin, mrtg, or the console-based nload). I find that without killing
irqbalance and manually setting the smp_affinity, I can't push more
than 400Mbit without a CPU hitting 100% utilization (run top then hit
'1' - look at the idle column). Moritz has found the same.
Though maybe you got lucky and your ethernet card + kernel version
combo are already properly balancing IRQs out of the box. Do a 'watch
"cat /proc/interrupts | grep eth"' to keep an eye on your ethernet
interrupt distribution to see.
Also keep your eyes on 'dmesg | grep conntrack' messages. Once you
start to get above 65k connections ('netstat -na | wc'), you should
start getting those too, and your individual tcp connection throughput
will begin to suffer.
The options we've been discussing aren't needed until you actually
start pushing more than a few hundred Mbit, but I would be very
surprised if you were able to do a full gbit without them.
--
Mike Perry
Mad Computer Scientist
fscked.org evil labs
Attachment:
pgpKLbznh41Rm.pgp
Description: PGP signature