On Tue, Sep 06, 2011 at 04:39:03PM +0100, Mel Gorman wrote:
> On Wed, Sep 07, 2011 at 12:11:40AM +0900, Minchan Kim wrote:
> > > > <SNIP>
> > > > If we consider that, we have to fix other reset_reclaim_mode cases as
> > > > well as mlocked pages.
> > > > Or
> > > > fix isolataion logic for the lumpy? (When we find the page isn't able
> > > > to isolate, rollback the pages in the lumpy block to the LRU)
> > > > Or
> > > > Nothing and wait to remove lumpy completely.
> > > >
> > > > What do you think about it?
> > >
> > > The rollback may be overkill and we already abort clustering the
> > > isolation when one of the pages fails.
> >
> > I think abort isn't enough
> > Because we know the chace to make a bigger page is gone when we isolate page.
> > But we still try to reclaim pages to make bigger space in a vain.
> > It causes unnecessary unmap operation by try_to_unmap which is costly operation
> > , evict some working set pages and make reclaim latency long.
> >
> > As a matter of fact, I though as follows patch to solve this problem(Totally, untested)
> >
>
> I confess I haven't read this patch carefully or given it much
> thought. I agree with you in principal that it would be preferred if
> lumpy reclaim disrupted the LRU lists as little as possible but I'm
> wary about making lumpy reclaim more complex when it is preferred that
> compaction is used and we expect lumpy reclaim to go away eventually.
Agreed.
But I think the concept of the patch could be applied to compaction for high order pages.
If we know some block has a pinned page when we do compaction for high order pages,
migration of the pages isolated in the block is pointless.
>
> > > <SNIP{>
> > >
> > > I would go with the last option. Lumpy reclaim is on its way out and
> > > already disabled for a rather common configuration, so I would defer
> > > non-obvious fixes like these until actual bug reports show up.
> >
> > It's hard to report above problem as it might not make big difference on normal worklaod.
>
> I doubt it makes a noticable difference as lumpy reclaim disrupts
> the system quite heavily.
Yes. I don't know such workload but I think it apparently could make relcaim latency
long with be not able to make bigger page.
>
> > But I agree last option, too. Then, when does we suppose to remove lumpy?
> > Mel, Could you have a any plan?
> >
>
> I think it should be removed after all the major distributions release
> with a kernel with compaction enabled. At that point, we'll know
> that lumpy reclaim is not being depended upon.
It does make sense.
>
> --
> Mel Gorman
> SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>