Re: 2.6.5-rc3-mm4 x86_64 sched domains patch

[Nick Piggin]

Darren Hart wrote:

> The current default implementations of arch_init_sched_domains
> constructs either a flat or two level topolology.  The two level
> topology is built if CONFIG_NUMA is set.  It seems that CONFIG_NUMA is
> not the appropriate flag to use for constructing a two level topology
> since some architectures which define CONFIG_NUMA would be better served
> with a flat topology.  x86_64 for example will construct a two level
> topology with one CPU per node, causing performance problems because
> balancing within nodes is pointless and balancing across nodes doesn't
> occur as often.

This is correct, although I don't know why there would be
performance problems. The rebalance in the degenerate node-local
domain should be basically unmeasurable. It would be nice to
get rid of it at some time. I have code to prune off degenerate
domains, which I will submit soonish.

The NUMA rebalance should occur more often than the old numasched
did, but perhaps with some recent Altix-centric changes to the
generic setup, this is no longer the case.

The STREAM performance problem is due mainly to the more
conservative nature of balancing, which is otherwise a good thing.
I think we can fix this in the short term by having x86_64 balance
between nodes more often. In the long term, we can merge Ingo's
balance on clone stuff, and the interested people can play with
that.

> This patch introduces a new CONFIG_SCHED_NUMA flag and uses it to decide
> between a flat or two level topology of sched_domains.  The patch is
> minimally invasive as it primarily modifies Kconfig files and sets the
> appropriate default (off for x86_64, on for everything that used to
> export CONFIG_NUMA) and should only change the sched_domains topology
> constructed on x86_64 systems.  I have verified this on a 4 node x86
> NUMAQ, but need someone to test x86_64.

I guess I can't see a big problem with this, other than more
complexity. In the long run, we should obviously have the arch
code set up optimal domains depending on the machine and config.

> This patch is intended as a quick fix for the x86_64 problem, and
> doesn't solve the problem of how to build generic sched domain
> topologies.  We can certainly conceive of various topologies for x86
> systems, so even arch specific topologies may not be sufficient.  Would
> sub-arch (ie NUMAQ) be the right way to handle different topologies, or
> will we be able to autodiscover the appropriate topology?  I will be
> looking into this more, but thought some might benefit from an immediate
> x86_64 fix.  I am very interested in hearing your ideas on this.

SGI want to do sub arch domains so they can do specific things
with their systems. I don't really care what the arch code does
with them, but it would be wise to only specialise it when there
is a genuine need. I'm glad you'll be looking into it, thanks.

Nick

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/