[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Xen-devel] [PATCH 08 of 10 v2] libxl: enable automatic placement of guests on NUMA nodes
On Tue, 2012-06-26 at 17:25 +0100, Dario Faggioli wrote:
> On Fri, 2012-06-22 at 11:14 +0100, Ian Campbell wrote:
> > > > > + /* Bring the best candidate in front of the list -->
> > > > > candidates */
> > > > > + if (nr_candidates > 1)
> > > > > + libxl__sort_numa_candidates(candidates, nr_candidates,
> > > > > numa_cmpf);
> > > >
> > > > Is the start up cost of libxl__sort_numa_candidates significant enough
> > > > to make that if worthwhile?
> > > >
> > > I'm not sure what you mean here, I'm afraid ... :-(
> > Is the cost of sorting the 1 element list so high it is worth avoiding.
> > Since the sort itself would be trivial the startup costs are what would
> > dominate.
> So, the answer is that I really don't know, as I'm using a library
> function for doing the actual work. I really do not expect for them to
> be that high, and I hence can remove the if() if you think it looks
> ugly. :-)
Not really ugly, it just seemed like an odd optimisation to make.
> > > > I'm not sure about this, if numa_place_domain fails for any reason would
> > > > we be not better off logging and continuing without placement? We
> > > > already do that explicitly in a couple of cases.
> > > >
> > > Well, actually, if it "fails" in the sense it finds no candidates or
> > > does not manage in placing the domain, it returns 0, so we do right what
> > > you're suggesting. I'm sure this is stated in some comment but I guess I
> > > can repeat it here. If !rc, it means we stepped into some ERROR_FAIL or
> > > something, which I think has to be handled like this, hasn't it?
> > That makes sense.
> > > Perhaps I can also add some logging about "successfully failing", i.e.,
> > > not getting into any error but not being able to place the domain. If
> > > yes, that will happen _inside_ numa_place_domain() rather than here.
> > Logging in that case seems wise in any case since it will have
> > performance implications I think.
> Ok, I'll log something out.
> Thanks and Regards,
Xen-devel mailing list