On Wed, Oct 01, 2014 at 11:14:59AM +0200, Nils Beyer wrote:
> Hi,
>
> Yonghyeon PYUN wrote:
> > Default interrupt moderation policy is targeted to reduce latency
> > so it will generate up to 10k interrupts/sec under high network
> > load. If you want to reduce number of interrupts/sec, tune
> > interrupt moderation sysctl variables mentioned in alc(4).
>
> Tried several values here:
>
> dev.alc.0.int_rx_mod={1000,10000,100000}
> dev.alc.0.int_tx_mod={1000,10000,100000}
>
> but didn't notice any changes neither in CPU usage nor throughput during the
> "iperf" test; "kernel{alc0 taskq}" stays at 70-75%. I've downed/upped the
> interface "alc0" after every change.
>
You may see difference when H/W handles tiny grams(i.e. 64 bytes
UDP frames). For bulk TCP/UDP transfers, alc(4) can easily
saturate the link.
> A simple iSCSI test using the native CTL interface works really well. A "fio"
> test results in 100MB/s read and write. Double-checking using "netstat -I"
> confirms gigabit-line speeds at around 120MB/s.
> CPU usage at "kernel{alc0 taskq}" is as high as in the "iperf" test. So I
> think
> that's a limitation of the AR8161 chip.
>
>
> > Updated the diff to address link establishment issue.
> > http://people.freebsd.org/~yongari/alc/pci.quirk.diff
> > http://people.freebsd.org/~yongari/alc/alc.diff.20141001
>
> Confirmed; with the anti-hibernation patch, link estalishment is now working
> flawlessly.
>
> Thank you very much for your work...
Thanks for your testing. Patch updated again to fix wrong lock
assertion.
http://people.freebsd.org/~yongari/alc/alc.diff.20141002
_______________________________________________
freebsd-net@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-net
To unsubscribe, send any mail to "freebsd-net-unsubscr...@freebsd.org"
