Multiple netgraph threads

Alexander Motin mav at FreeBSD.org
Wed Dec 3 00:17:47 PST 2008


Hi.

Vladimir V. Kobal wrote:
> I'm interested in the information of further development of Subject. I have
> a bottleneck: swi_net limited by one CPU core time. There was a thread
> http://lists.freebsd.org/pipermail/freebsd-net/2008-March/017447.html .
> Patch http://people.freebsd.org/~mav/netgraph.threads.patch is not available
> :(. Could somebody mail it to me. Are there any other similar patches?

I have uploaded that patch back. Not sure is it correct at this moment,
there was some changes, but that time it worked fine.

I have measured some benefit on my tests with Core2Quad using NetPerf
cluster. But even having all 4 cores busy, I have got just only about
20% performance benefit on that test setup (routing between three
Gigabit PPTP links). I suppose it was result of constant cache trashing,
when the same netgraph nodes write-accessed by different CPUs. But may
be on more diverse workload, or with some heavy compression/encryption
used benefit can be bigger.

-- 
Alexander Motin


More information about the freebsd-net mailing list