Re: [PATCH] Bias the location of pages freed for min_free_kbytes inthe same MAX_ORDER_NR_PAGES blocks

[Mel Gorman]

On Sun, 18 Mar 2007, Andrew Morton wrote:

> On Sun, 18 Mar 2007 11:35:36 +0000 (GMT) Mel Gorman <mel@xxxxxxxxx> wrote:
>
> > > But let me leap ahead of myself.
> > >
> > > > CONFIG_PAGE_GROUP_BY_MOBILITY
> > >
> > > Why does this config item exist?  It's not good to have some mysterious
> > > knob which affects mm behaviour at compile time.  We need to make up our
> > > minds and stick with it.
> >
> > The configuration item exists because there were concerns over the memory
> > footprint and cache line footprint. It was introduced to address that
> > concern and also so that it would be possible to compare the performance
> > behavior of anti-fragmentation. Your comment rang a bell though so I
> > searched the archives to see this comment from Andi Kleen;
> >
> > ===
> > If anything this should be a boot time option or perhaps sysctl, not a
> > config. In general CONFIGs that change runtime behaviour are evil - just
> > makes changing the option more painful, causes problems for distribution
> > users, doesn't make much sense, etc.etc.
> >
> > Also #ifdef as a documentation device is a really really scary concept.
> > Yuck.
> > ===
> >
> > A sysctl would avoid any cache line footprint but not the memory overhead
> > because the freelists in struct zone as those freelists would still exist.
> > I could make the option depend on CONFIG_EMBEDDED for the zone overhead.
> > Would that make sense or would it be preferable to ditch the option
> > altogether?
> >
> > I'll start looking at doing a sysctl so it can be disabled at runtime if
> > necessary. I strongly suspect that it cannot be enabled again once
> > disabled but I don't see that as a problem as such.
>
> How much additional memory consumption are we expecting here?

Short answer, about 1.5KB on a 1GB system of which 1.3KB is statically 
defined in the 3 struct zones on a 1 node x86 system.

Longer answer that I hopefully have not made any mistakes in - There is 
the zone overhead which is statically sized and a runtime overhead which 
depends on the amount of memory in the system. The additional zone 
overhead is the overhead for additional freelists (larger struct 
free_area) and is as follows;

(MIGRATE_TYPES-1) * sizeof(list_head) * (MAX_ORDER-1)

so, on 32 bit in general, thats

4 * 8 * 10 = 320 bytes per zone (would be 240 bytes if MIGRATE_RESERVE is
				sufficient for higher order allocations
				instead of MIGRATE_HIGHALLOC)

on x86 with DMA, Normal and HighMem, thats 1280 bytes. On a NUMA system, 
it's 1280 bytes per node. On 64 bit, it would be double because of the 
larger pointer size. At worst, I guess you are looking at 3KB per node.

The size of the bitmap for the flags depends on the amount of memory. On 
FLATMEM and DISCONTIG, it's 3 bits per MAX_ORDER_NR_PAGES spanned by the 
zones (note spanned, not present). On a 1GiB system without holes on an 
x86 with MAX_ORDER of 11, that would be 96 bytes. The calculation is more 
complex for sparsemem but will be at least 4 bytes per active section for 
a pointer and another 4 bytes for most bitmaps. I have a patch that uses 
the pointer itself when the bitmap is small enough to be contained in 4 
bytes (which is usually is) so it could be just 4 bytes per active section 
(4 bytes per 16MB on powerpc for example).

> Whether it's runtime or compile-time, the optionality is not good.

The main situation where I think it would be desirable to disable 
anti-fragmentation is for zones that are smaller than 
MIGRATE_TYPES*MAX_ORDER_NR_PAGES and this is for performance reasons to 
avoid regularly entering __rmqueue_fallback(). However, that situation 
could be automatically detected and handled by having 
allocflags_to_migratetype() always return MIGRATE_UNMOVABLE for small 
zones and similar for get_pageblock_migratetype(). There would be no need 
for a sysctl.

If a sysctl was available for cases where profiles showed a problem with 
the bitmap for example, it would work by having 
allocflags_to_migratetype() and get_pageblock_migratetype() always return 
MIGRATE_UNMOVABLE so it would not be totally different code paths that are 
being executed in the enabled/disabled cases.

--
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
-
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/