Re: [PATCH 6/6] mm: have order > 0 compaction start near a pageblockwith free pages

[Mel Gorman]

On Tue, Aug 07, 2012 at 09:20:08AM -0600, Jim Schutt wrote:
> On 08/07/2012 08:52 AM, Mel Gorman wrote:
> >On Tue, Aug 07, 2012 at 10:45:25AM -0400, Rik van Riel wrote:
> >>On 08/07/2012 08:31 AM, Mel Gorman wrote:
> >>>commit [7db8889a: mm: have order>   0 compaction start off where it left]
> >>>introduced a caching mechanism to reduce the amount work the free page
> >>>scanner does in compaction. However, it has a problem. Consider two process
> >>>simultaneously scanning free pages
> >>>
> >>>				    			C
> >>>Process A		M     S     			F
> >>>		|---------------------------------------|
> >>>Process B		M 	FS
> >>
> >>Argh. Good spotting.
> >>
> >>>This is not optimal and it can still race but the compact_cached_free_pfn
> >>>will be pointing to or very near a pageblock with free pages.
> >>
> >>Agreed on the "not optimal", but I also cannot think of a better
> >>idea right now. Getting this fixed for 3.6 is important, we can
> >>think of future optimizations in San Diego.
> >>
> >
> >Sounds like a plan.
> >
> >>>Signed-off-by: Mel Gorman<mgorman@xxxxxxx>
> >>
> >>Reviewed-by: Rik van Riel<riel@xxxxxxxxxx>
> >>
> >
> >Thanks very much.
> >
> >Jim, what are the chances of getting this series tested with your large
> >data workload? As it's on top of 3.5, it should be less scary than
> >testing 3.6-rc1 but if you are comfortable testing 3.6-rc1 then please
> >test with just this patch on top.
> >
> 
> As it turns out I'm already testing 3.6-rc1, as I'm on
> the trail of a Ceph client messaging bug.  I think I've
> about got that figured out, and am working on a patch, but
> I need it fixed in order to generate enough load to trigger
> the problem that your patch addresses.
> 

Grand, good luck with the Ceph bug.

> Which is a long-winded way of saying:  no problem, I'll
> roll this into my current testing, but I'll need another
> day or two before I'm likely to be able to generate a
> high enough load to test effectively.  OK?
> 

That is perfectly reasonable, thanks.

> Also FWIW, it occurs to me that you might be interested
> to know that my load also involves lots of network load
> where I'm using jumbo frames.  I suspect that puts even
> more stress on higher page order allocations, right?
> 

It might. It depends on whether the underlying driver needs contiguous
pages to handle jumbo frame, if it can do scatter/gather IO or some
combination like trying for a contiguous page but using scatter/gather as
a fallback. Certainly it is interesting and I will keep it in mind.

-- 
Mel Gorman
SUSE Labs
--
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/