Re: [RFC][PATCHv2 2/8] zsmalloc: partial page ordering within a fullness_list

[Sergey Senozhatsky]

Hello,

On (06/16/15 22:19), Minchan Kim wrote:
> Hello Sergey,
> 
> On Fri, Jun 05, 2015 at 09:03:52PM +0900, Sergey Senozhatsky wrote:
> > We want to see more ZS_FULL pages and less ZS_ALMOST_{FULL, EMPTY}
> > pages. Put a page with higher ->inuse count first within its
> > ->fullness_list, which will give us better chances to fill up this
> > page with new objects (find_get_zspage() return ->fullness_list head
> > for new object allocation), so some zspages will become
> > ZS_ALMOST_FULL/ZS_FULL quicker.
> > 
> > It performs a trivial and cheap ->inuse compare which does not slow
> > down zsmalloc, and in the worst case it keeps the list pages not in
> > any particular order, just like we do it now.
> 
> Fair enough.
> 
> I think it would be better with small cost and it matches current
> zsmalloc design which allocates a object from ALMOST_FULL zspage
> first to reduce memory footprint.
> 
> Although we uses page->lru to link zspages, it's not lru order
> so I like your idea.
> 
> One think I want to say is it doesn't need to be part of this patchset.
> I hope you gives some data to prove gain and includes it in changelog
> and resubmit, please.

OK, I'll do some testing and come back with some numbers.

Well, none of these patches will get into 4.2, so splitting
the patchset will make things a bit less easy, but I guess I
can do this.

	-ss

> > 
> > A more expensive solution could sort fullness_list by ->inuse count.
> > 
> > Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@xxxxxxxxx>
> > ---
> >  mm/zsmalloc.c | 12 ++++++++++--
> >  1 file changed, 10 insertions(+), 2 deletions(-)
> > 
> > diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
> > index ce3310c..cd37bda 100644
> > --- a/mm/zsmalloc.c
> > +++ b/mm/zsmalloc.c
> > @@ -658,8 +658,16 @@ static void insert_zspage(struct page *page, struct size_class *class,
> >  		return;
> >  
> >  	head = &class->fullness_list[fullness];
> > -	if (*head)
> > -		list_add_tail(&page->lru, &(*head)->lru);
> > +	if (*head) {
> > +		/*
> > +		 * We want to see more ZS_FULL pages and less almost
> > +		 * empty/full. Put pages with higher ->inuse first.
> > +		 */
> > +		if (page->inuse < (*head)->inuse)
> > +			list_add_tail(&page->lru, &(*head)->lru);
> > +		else
> > +			list_add(&page->lru, &(*head)->lru);
> > +	}
> >  
> >  	*head = page;
> >  	zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
> > -- 
> > 2.4.2.387.gf86f31a
> > 
> 
> -- 
> Kind regards,
> Minchan Kim
> 
--
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/