On Tue, Nov 09, 2010 at 11:22:40PM +0100, Andrea Arcangeli wrote:
quoted text
> Hi Mel,
> 
> On Tue, Nov 09, 2010 at 09:38:55PM +0000, Mel Gorman wrote:
> > Specifically, I measured that lumpy in combination with compaction is
> > more reliable and lower latency but that's not the same as deleting it.
> 
> Thanks for the clarification. Well no doubt that using both could only
> increase the success rate. So the thing with hugetlbfs you may want to
> run both, but with THP we want to stop at compaction.

Agreed. Any performance increase from THP is not likely to offset the
cost of lumpy reclaim.

quoted text
> So this would
> then require a __GFP_LUMPY if we want hugetlbfs to fallback on lumpy
> whenever compaction isn't successful. We can't just nuke and ignore
> young bits in pte if compaction fails. Trying later in khugepaged once
> every 10 seconds is a lot better.
> 

Again agreed, I have no problem with lumpy reclaim being pushed aside.
I'm just less keen on it being disabled altogether. I have high hopes
for the series I'm working on that it can be extended slightly to suit
the needs of THP.

quoted text
> > That said, lumpy does hurt the system a lot.  I'm prototyping a series at the
> > moment that pushes lumpy reclaim to the side and for the majority of cases
> > replaces it with "lumpy compaction". I'd hoping this will be sufficient for
> > THP and alleviate the need to delete it entirely - at least until we are 100%
> > sure that compaction can replace it in all cases.
> > 
> > Unfortunately, in the process of testing it today I also found out that
> > 2.6.37-rc1 had regressed severely in terms of huge page allocations so I'm
> > side-tracked trying to chase that down. My initial theories for the regression
> > have shown up nothing so I'm currently preparing to do a bisection. This
> > will take a long time though because the test is very slow :(
> 
> On my side (unrelated) I also found 37-rc1 broke my mic by changing
> soundcard type (luckily csipsimple and skype on my cellphone are now
> working better than laptop for making voip calls so it was easy to
> workaround) and my backlight goes blank forever after a "xset dpms
> force standby" (so I'm stuck in presentation mode to workaround it,
> suspend to ram was successful to avoid having to reboot too as the
> bios restarts the backlight during boot).
> 

I do not believe they are related. Fortunately, I did not have to do a
full bisect but I know roughly what area the problem must be in. The
problem commit looks like d065bd81. I'm running further tests with it
reverted to see if it's true but it'll take a few hours to complete.

quoted text
> > I can still post the series as an RFC if you like to show what direction
> > I'm thinking of but at the moment, I'm unable to test it until I pin the
> > regression down.
> 
> Sure feel free to post it, if it's already worth testing it, I can
> keep at the end of the patchset considering it's new code while what I
> posted had lots of testing.
> 

As I hopefully have pinned down the problem commit, I'm going to hold
off for another day to see can I get real data.

quoted text
> With THP we have khugepaged in the background, nothing is mandatory at
> allocation time. I don't want a super aggressive thing at allocation
> time, and lumpy by ignoring all young bits is too aggressive and
> generates swap storms for every single allocation. We need to fail
> order 9 allocation quick even if compaction fails (like if more than
> 90% of the ram is asked in hugepages so having to use ram in the
> unmovable page blocks selected by anti-frag) to avoid hanging the
> system during allocations. Looking my stats things seem to be working
> ok with compaction in 37-rc1, so maybe it's just the lumpy changes
> that introduced your regression?
> 

Nah, the first thing I did was eliminate being "my fault" :). It would
have surprised me because the patches in isolation worked fine. It
thought the inode changes might have had something to do with it so I
was chasing blind alleys for a while. Hopefully d065bd81 will prove to
be the real problem.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/