Re: [PATCH] mm: memmap_init_zone() performance improvement

[Mike Yoknis <mike.yoknis@hp.com>]

On Sat, 2012-10-20 at 09:29 +0100, Mel Gorman wrote:
> On Fri, Oct 19, 2012 at 01:53:18PM -0600, Mike Yoknis wrote:
> > On Tue, 2012-10-09 at 08:56 -0600, Mike Yoknis wrote:
> > > On Mon, 2012-10-08 at 16:16 +0100, Mel Gorman wrote:
> > > > On Wed, Oct 03, 2012 at 08:56:14AM -0600, Mike Yoknis wrote:
> > > > > memmap_init_zone() loops through every Page Frame Number (pfn),
> > > > > including pfn values that are within the gaps between existing
> > > > > memory sections.  The unneeded looping will become a boot
> > > > > performance issue when machines configure larger memory ranges
> > > > > that will contain larger and more numerous gaps.
> > > > > 
> > > > > The code will skip across invalid sections to reduce the
> > > > > number of loops executed.
> > > > > 
> > > > > Signed-off-by: Mike Yoknis <mike.yoknis@hp.com>
> > > > 
> > > > I do not see the need for
> > > > the additional complexity unless you can show it makes a big difference
> > > > to boot times.
> > > > 
> > > 
> > > Mel,
> > > 
> > > Let me pass along the numbers I have.  We have what we call an
> > > "architectural simulator".  It is a computer program that pretends that
> > > it is a computer system.  We use it to test the firmware before real
> > > hardware is available.  We have booted Linux on our simulator.  As you
> > > would expect it takes longer to boot on the simulator than it does on
> > > real hardware.
> > > 
> > > With my patch - boot time 41 minutes
> > > Without patch - boot time 94 minutes
> > > 
> > > These numbers do not scale linearly to real hardware.  But indicate to
> > > me a place where Linux can be improved.
> > > 
> > > Mike Yoknis
> > > 
> > Mel,
> > I finally got access to prototype hardware.  
> > It is a relatively small machine with only 64GB of RAM.
> >  
> > I put in a time measurement by reading the TSC register.
> > I booted both with and without my patch -
> >  
> > Without patch -
> > [    0.000000]   Normal zone: 13400064 pages, LIFO batch:31
> > [    0.000000] memmap_init_zone() enter 1404184834218
> > [    0.000000] memmap_init_zone() exit  1411174884438  diff = 6990050220
> >  
> > With patch -
> > [    0.000000]   Normal zone: 13400064 pages, LIFO batch:31
> > [    0.000000] memmap_init_zone() enter 1555530050778
> > [    0.000000] memmap_init_zone() exit  1559379204643  diff = 3849153865
> >  
> > This shows that without the patch the routine spends 45% 
> > of its time spinning unnecessarily.
> >  
> 
> I'm travelling at the moment so apologies that I have not followed up on
> this. My problem is still the same with the patch - it changes more
> headers than is necessary and it is sparsemem specific. At minimum, try
> the suggestion of 
> 
> if (!early_pfn_valid(pfn)) {
>       pfn = ALIGN(pfn + MAX_ORDER_NR_PAGES, MAX_ORDER_NR_PAGES) - 1;
>       continue;
> }
> 
> and see how much it gains you as it should work on all memory models. If
> it turns out that you really need to skip whole sections then the strice
> could MAX_ORDER_NR_PAGES on all memory models except sparsemem where the
> stride would be PAGES_PER_SECTION
> 
Mel,
I tried your suggestion.  I re-ran all 3 methods on our latest firmware.

The following are TSC difference numbers (*10^6) to execute
memmap_init_zone() -

No patch   - 7010
Mel's patch- 3918
My patch   - 3847

The incremental improvement of my method is not significant vs. yours.

If you believe your suggested change is worthwhile I will create a v2
patch.
Mike Y


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>