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
