lkml.org 
[lkml]   [2016]   [Mar]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH v3 1/5] mm/zsmalloc: introduce class auto-compaction
On Tue, Mar 15, 2016 at 10:33:03AM +0900, Sergey Senozhatsky wrote:
> On (03/15/16 09:46), Minchan Kim wrote:
> [..]
> > > yes,
> > >
> > > we do less work this way - scan and compact only one class, instead
> > > of locking and compacting all of them; which sounds reasonable.
> >
> > Hmm,, It consumes more memory(i.e., sizeof(work_struct) + sizeof(void *)
> > + sizeof(bool) * NR_CLASS) as well as kicking many work up to NR_CLASS.
>
> yes, it does. not really happy with it either.
>
> > I didn't test your patch but I guess I can make worst case scenario.
> >
> > * make every class fragmented under 40%
> > * On the 40% boundary, repeated alloc/free of every class so every free
> > can schedule work if it was not scheduled.
> > * Although class fragment is too high, it's not a problem if the class
> > consumes small amount of memory.
>
> hm, in this scenario both solutions are less than perfect. we jump
> X times over 40% margin, we have X*NR_CLASS compaction scans in the
> end. the difference is that we queue less works, yes, but we don't
> have to use workqueue in the first place; compaction can be done
> asynchronously by a pool's dedicated kthread. so we will just
> wake_up() the process.

Hmm, kthread is over-engineered to me. If we want to create new kthread
in the system, I guess we should persuade many people to merge in.
Surely, we should have why it couldn't be done by others(e.g., workqueue).

I think your workqueue approach is good to me.
Only problem I can see with it is we cannot start compaction when
we want instantly so my conclusion is we need both direct and
background compaction.

For shrinker and user-space trigger knob, we could compact in that context
while we could queue background job to compact in zs_free.

>
> > I guess it can make degradation if I try to test on zsmalloc
> > microbenchmark.
> >
> > As well, although I don't know workqueue internal well, thesedays,
> > I saw a few of mails related to workqueue(maybe, vmstat) and it had
> > some trouble if system memory pressure is heavy IIRC.
>
> yes, you are right. wq provides WQ_MEM_RECLAIM bit for this
> case -- a special kthread that it will wake up to process works.
>
> > My approach is as follows, for exmaple.
> >
> > Let's make a global ratio. Let's say it's 4M.
>
> ok. should it depend on pool size? min(20% of pool_size, XXMB)?

Maybe, that could be a knob but need to think more what should be
default. In this moment, clear thing is that we should prevent
frequent ping-pong background compaction as repeated alloc/free
with dancing on threshold boundary.

>
> > If zs_free(or something) realizes current fragment is over 4M,
> > kick compacion backgroud job.
>
> yes, zs_free() is the only place that introduces fragmentation.
>
> > The job scans from highest to lower class and compact zspages
> > in each size_class until it meets high watermark(e.g, 4M + 4M /2 =
> > 6M fragment ratio).
>
> ok.
>
> > And in the middle of background compaction, if we find it's too
> > many scan(e.g., 256 zspages or somethings), just bail out the
> > job for the latency and reschedule it for next time. At the next
> > time, we can continue from the last size class.
>
> ok. I'd probably prefer more simple rules here:
> -- bail out because it has compacted XXMB
> so the fragmentation ratio is *expected* to be below the watermark

Need high watermark to stop compaction.
It will prevent frequent background compaction triggering.


> -- nothing to scan anymore
> compaction is executed concurrently with zs_free()/zs_malloc()
> calls, it's harder to control/guarantee some global state.
>
> overall, no real objections. this approach can work, I think. need
> to test it.

Thanks, Sergey!

\
 
 \ /
  Last update: 2016-03-15 08:01    [W:0.078 / U:0.288 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site