On (06/04/15 12:15), Sergey Senozhatsky wrote:
> I'm still thinking how good it should be.
>
> for automatic compaction we don't want to uselessly move objects between
> pages and I tend to think that it's better to compact less, than to waste
> more cpu cycless.
>
>
> on the other hand, this policy will miss cases like:
>
> -- free objects in class: 5 (free-objs class capacity)
> -- page1: inuse 2
> -- page2: inuse 2
> -- page3: inuse 3
> -- page4: inuse 2
>
> so total "insuse" is greater than free-objs class capacity. but, it's
> surely possible to compact this class. partial inuse summ <= free-objs class
> capacity (a partial summ is a ->inuse summ of any two of class pages:
> page1 + page2, page2 + page3, etc.).
>
> otoh, these partial sums will badly affect performance. may be for automatic
> compaction (the one that happens w/o user interaction) we can do zs_can_compact()
> and for manual compaction (the one that has been triggered by a user) we can
> old "full-scan".
>
> anyway, zs_can_compact() looks like something that we can optimize
> independently later.
>
so what I'm thinking of right now, is:
-- first do "if we have enough free objects to free at least one page"
check. compact if true.
-- if false, then we can do on a per-page basis
"if page->inuse <= class free-objs capacity" then compact it,
else select next almost_empty page.
here would be helpful to have pages ordered by ->inuse. but this
is far to expensive.
I have a patch that I will post later that introduces weak/partial
page ordering within fullness_list (really inexpensive: just one int
compare to add a page with a higher ->inuse to list head instead of
list tail).
-ss
--
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/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>