lkml.org 
[lkml]   [2016]   [Feb]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH v2 3/3] mm/zsmalloc: increase ZS_MAX_PAGES_PER_ZSPAGE
Hello Minchan,

sorry for very long reply.

On (02/24/16 01:05), Minchan Kim wrote:
[..]
> > And the thing is -- quite huge internal class fragmentation. These are the 'normal'
> > classes, not affected by ORDER modification in any way:
> >
> > class size almost_full almost_empty obj_allocated obj_used pages_used pages_per_zspage compact
> > 107 1744 1 23 196 76 84 3 51
> > 111 1808 0 0 63 63 28 4 0
> > 126 2048 0 160 568 408 284 1 80
> > 144 2336 52 620 8631 5747 4932 4 1648
> > 151 2448 123 406 10090 8736 6054 3 810
> > 168 2720 0 512 15738 14926 10492 2 540
> > 190 3072 0 2 136 130 102 3 3
> >
> >
> > so I've been thinking about using some sort of watermaks (well, zsmalloc is an allocator
> > after all, allocators love watermarks :-)). we can't defeat this fragmentation, we never
> > know in advance which of the pages will be modified or we the size class those pages will
> > land after compression. but we know stats for every class -- zs_can_compact(),
> > obj_allocated/obj_used, etc. so we can start class compaction if we detect that internal
> > fragmentation is too high (e.g. 30+% of class pages can be compacted).
>
> AFAIRC, we discussed about that when I introduced compaction.
> Namely, per-class compaction.
> I love it and just wanted to do after soft landing of compaction.
> So, it's good time to introduce it. ;-)

ah, yeah, indeed. I vaguely recall this. my first 'auto-compaction' submission
has had this "compact every class in zs_free()", which was a subject to 10+%
performance penalty on some of the tests. but with watermarks this will be less
dramatic, I think.

> >
> > on the other hand, we always can wait for the shrinker to come in and do the job for us,
> > but that can take some time.
>
> Sure, with the feature, we can remove shrinker itself, I think.
> >
> > what's your opinion on this?
>
> I will be very happy.

good, I'll take a look later, to avoid any conflicts with your re-work.

[..]
> > does it look to you good enough to be committed on its own (off the series)?
>
> I think it's good to have. Firstly, I thought we can get the information
> by existing stats with simple math on userspace but changed my mind
> because we could change the implementation sometime so such simple math
> might not be perfect in future and even, we can expose it easily so yes,
> let's do it.

thanks! submitted.

-ss

\
 
 \ /
  Last update: 2016-02-27 08:01    [W:0.063 / U:0.188 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site