[lkml]   [2004]   [Apr]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRE: 2.6.5-rc3-mm4 x86_64 sched domains patch
Can SLIT/SRAT be used here to define topology for the generic case?

SRAT is being used by i386 to build zonelists, but not for the scheduler -
any good reason why?


-----Original Message-----
[] On Behalf Of Nick Piggin
Sent: Thursday, April 08, 2004 16:42
To: Darren Hart
Cc: lkml;; Martin J Bligh; Rick Lindsley;; Ingo
Subject: Re: 2.6.5-rc3-mm4 x86_64 sched domains patch

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

>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.


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 14:02    [W:0.075 / U:0.980 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site