[talk] *BSD multi-core scalability and NUMA

[Pete Wright]

hey there folks - i've been working on building out some
low-latency/high-througput systems running one of those penguin
distributions lately and it has me thinking about the current state of
multi-core and multi-socket processing on *BSD.  I've found this post
that is pretty recent regarding NUMA(0) support on FreeBSD:

http://freebsd.1045724.n5.nabble.com/Working-on-NUMA-support-td5874392.html

This is actually pretty exciting work from my POV, especially with the
default use of jemalloc on FreeBSD I may be able to make a good case to
move some of my critical systems off of centos finally.  I've actually
seen a handful of closed-source systems we run use jemalloc on centos
since it works quite nice with NUMA.

I was wondering how people manage scaling out multi-socket systems
currently on the *BSD's.  I mean beyond just having multiple CPU's
availble for the scheduler - for example ensuring your NIC driver is
handling interrupts on a single CPU package for example, or having
processes and/or threads run on the same CPU package for potential cache
affinity.


Cheers,
-pete


0: http://en.wikipedia.org/wiki/Non-uniform_memory_access

-- 
Pete Wright
pete at nomadlogic.org
twitter => @nomadlogicLA