svn commit: r253002 - head

[John Baldwin]

On Monday, July 08, 2013 7:24:04 pm Garrett Cooper wrote:
> On Mon, Jul 8, 2013 at 2:13 PM, John Baldwin <jhb at freebsd.org> wrote:
> > On Monday, July 08, 2013 2:23:31 am Garrett Cooper wrote:
> >> On Sun, Jul 7, 2013 at 7:25 PM, Garrett Cooper <yaneurabeya at gmail.com>
> > wrote:
> >> > On Jul 7, 2013, at 2:15 PM, Alfred Perlstein <alfred at freebsd.org> wrote:
> >> >
> >> >> On 7/7/13 2:01 PM, Garrett Cooper wrote:
> >> >>> Why the magic number 12?
> >> >>
> >> >> Numbers higher seem to result in worse performance as reported by some
> > members of my team.
> >> >
> >> > The suggestion is good in spirit, but this doesn't justify the reasoning
> > for this recommendation for all cases.
> >> >
> >> > Please revert this change and add a doc page or notes to the dev handbook
> > discussing what the empirical process and results were for determining this
> > value so people can come up with their own values that work best with their
> > hardware and software config. This recommendation is prone to bitrot like some
> > of the recommendations in tuning(7).
> >> >
> >> > Misinformation is sometimes more harmful than no information.
> >>
> >> I spoke with Alfred over the phone and did some more careful thought
> >> about this and I'm rescinding this request.
> >>
> >> Alfred did a good job at documenting how JFLAG works (it was
> >> previously undocumented). My concern over -j12 was performance
> >> related, and after giving things more careful thought it actually
> >> makes sense why -j12 was chosen because Westmere and newer processors
> >> have issues with NUMA and cache locality between multiple processor
> >> packages as we've seen non-empirically and empirically at Isilon with
> >> FreeBSD 7 and 10 (it's a known issue that jeffr@ and jhb@ are aware
> >> of).
> >>
> >> I'll come up with a concise patch that does what Alfred was trying to
> >> achieve and have Alfred review it.
> >>
> >> Thanks (and thank you Alfred for the contribution!!!)!
> >
> > Westmere is fine, it's post-Westmere that is more troublesome.
> 
> Even the 6-core Westmeres (I'm being completely dumb here as you and
> Jeff know a lot more about the NUMA issue than I do as I just caught
> the tail end of the conversation at BSDCan)?

Yes.  NUMA can matter some on Nehalem and Westmere, but for workloads
that involve I/O, the SB/IB generation where the I/O hub moves onto
the processor itself is much more finicky.

-- 
John Baldwin