Re: [PATCH 10/13] x86: mm: Enable deferred struct page initialisation on x86-64

[Waiman Long]

On 04/24/2015 11:20 AM, Mel Gorman wrote:
> On Fri, Apr 24, 2015 at 10:35:49AM -0400, Waiman Long wrote:
> > On 04/23/2015 05:23 AM, Mel Gorman wrote:
> > > On Wed, Apr 22, 2015 at 04:45:00PM -0700, Andrew Morton wrote:
> > > > On Wed, 22 Apr 2015 18:07:50 +0100 Mel Gorman<mgorman@xxxxxxx>   wrote:
> > > >
> > > > > --- a/arch/x86/Kconfig
> > > > > +++ b/arch/x86/Kconfig
> > > > > @@ -32,6 +32,7 @@ config X86
> > > > >   	select HAVE_UNSTABLE_SCHED_CLOCK
> > > > >   	select ARCH_SUPPORTS_NUMA_BALANCING if X86_64
> > > > >   	select ARCH_SUPPORTS_INT128 if X86_64
> > > > > +	select ARCH_SUPPORTS_DEFERRED_STRUCT_PAGE_INIT if X86_64&&   NUMA
> > > > Put this in the "config X86_64" section and skip the "X86_64&&"?
> > > Done.
> > >
> > > > Can we omit the whole defer_meminit= thing and permanently enable the
> > > > feature?  That's simpler, provides better test coverage and is, we
> > > > hope, faster.
> > > Yes. The intent was to have a workaround if there were any failures like
> > > Waiman's vmalloc failures in an earlier version but they are bugs that
> > > should be fixed.
> > >
> > > > And can this be used on non-NUMA?  Presumably that won't speed things
> > > > up any if we're bandwidth limited but again it's simpler and provides
> > > > better coverage.
> > > Nothing prevents it. There is less opportunity for parallelism but
> > > improving coverage is desirable.
> > Memory access latency can be more than double for local vs. remote
> > node memory. Bandwidth can also be much lower depending on what kind
> > of interconnect is between the 2 nodes. So it is better to do it in
> > a NUMA-aware way.
> I do not believe that is what he was asking. He was asking if we could
> defer memory initialisation even when there is only one node. It does not
> gain much in terms of boot times but it improves testing coverage.

Thanks for the clarification.

> > Within a NUMA node, however, we can split the
> > memory initialization to 2 or more local CPUs if the memory size is
> > big enough.
> I considered it but discarded the idea. It'd be more complex to setup and
> the two CPUs could simply end up contending on the same memory bus as
> well as contending on zone->lock.

I don't think we need that now. However, we may have to consider this 
when one day even a single node can have TBs of memory unless we move to 
a page size larger than 4k.

Cheers,
Longman
--
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/