On Tue, 13 Nov 2007, J=F6rn Engel wrote:

quoted text
> On Mon, 12 November 2007 20:41:10 -0800, Christoph Lameter wrote:
> > On Mon, 12 Nov 2007, Ray Lee wrote:
> >=20
> > > Discontig obviously needs to die. However, FlatMem is consistently
> > > faster, averaging about 2.1% better overall for your numbers above. I=
s
quoted text
> > > the page allocator not, erm, a fast path, where that matters?
> > >=20
> > > Order=09Flat=09Sparse=09% diff
> > > 0=09639=09641=090.3
> >=20
> > IMHO Order 0 currently matters most and the difference is negligible=20
> > there.
>=20
> Is it?  I am a bit concerned about the non-monotonic distribution.
> Difference starts a near-0, grows to 4.4, drops to near-0, grows to 4.9,
> drops to near-0.

The problem also is that the comparison here is between a SMP config for=20
flatmem vs a NUMA config for sparsemem. There is additional overhead in=20
the NUMA config.=20

The effect may also be due to the system being able to place=20
some pages in the same 2MB section as the memmap with flatmem. However,=20
that is only feasable immeidately after bootup. In regular operations this=
=20
should vanish.

Could you run your own test to verify?

quoted text
> Is there an explanation for this behaviour?  More to the point, could
> repeated runs also return 4% difference for order-0?

I hope I have given some above. The number of the page allocator suggests=
=20
that we have far too much fat in the allocation paths. IMHO reasonable=20
numbers for an order-0 alloc should be ~100 cycles.