[PATCH 1/1] zram: better utilization of zram swap space

[PATCH 1/1] zram: better utilization of zram swap space

[Gopi Sai Teja]

If the length of the compressed page is greater than 75% of the PAGE_SIZE,
then the page is stored uncompressed in zram space. Zram space utilization
is improved if the threshold is 80%(5 compressed pages can be stored in
4 pages).

If the compressed length is greater than 3068 and less than 3261, pages
still can be stored in compressed form in zs_malloc class 3264.
Currently these compressed pages belong to 4096 zs malloc class.

Tested on ARM:
Before Patch:

class  size  obj_allocated   obj_used pages_used
  190  3072           3820       3819       2865
  202  3264             65         63         52
  254  4096          11791      11791      11791

 Total               65604      64616      29504

After Patch:

class  size  obj_allocated   obj_used pages_used
  190  3072           4244       4243       3183
  202  3264           1500       1499       1200
  254  4096           6306       6306       6306

 Total               65983      64896      26628

Signed-off-by: Gopi Sai Teja <gopi.st@samsung.com>
---
 drivers/block/zram/zram_drv.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/block/zram/zram_drv.h b/drivers/block/zram/zram_drv.h
index 5ece255..f7d9ddd 100644
--- a/drivers/block/zram/zram_drv.h
+++ b/drivers/block/zram/zram_drv.h
@@ -31,8 +31,12 @@ static const unsigned max_num_devices = 32;
 /*
  * Pages that compress to size greater than this are stored
  * uncompressed in memory.
+ * 16 is zsmalloc class interval and unsigned long is extra memory
+ * used by zsmalloc to store metadata.
  */
-static const size_t max_zpage_size = PAGE_SIZE / 4 * 3;
+
+static const size_t max_zpage_size = round_down((PAGE_SIZE / 5 * 4), 16)
+					- sizeof(unsigned long);
 
 /*
  * NOTE: max_zpage_size must be less than or equal to:
-- 
1.9.1

Re: [PATCH 1/1] zram: better utilization of zram swap space

[Sergey Senozhatsky]

On (12/07/17 13:52), Gopi Sai Teja wrote:
> If the length of the compressed page is greater than 75% of the PAGE_SIZE,
> then the page is stored uncompressed in zram space. Zram space utilization
> is improved if the threshold is 80%(5 compressed pages can be stored in
> 4 pages).
> 
> If the compressed length is greater than 3068 and less than 3261, pages
> still can be stored in compressed form in zs_malloc class 3264.
> Currently these compressed pages belong to 4096 zs malloc class.

so this makes sense. I had another idea awhile ago

lkml.kernel.org/r/1456061274-20059-2-git-send-email-sergey.senozhatsky@gmail.com

in short, 3261 is good, but not as good as it possibly can be. for the
time being, our huge-class watermark starts at 3264. but this can
change.


a side note, I think we have sort of wrong API. zsmalloc knows better which
object is huge. and who knows, may be we will change the number of huge
classes someday or huge-class watermark, etc. so having "hey zsmalloc, is
this object huge or not" API seems to be better than ZRAM's enforcement
"hey zsmalloc, this object is huge".

	-ss

Re: [PATCH 1/1] zram: better utilization of zram swap space

[Sergey Senozhatsky]

On (12/07/17 17:45), Sergey Senozhatsky wrote:
[..]
> On (12/07/17 13:52), Gopi Sai Teja wrote:
> > If the length of the compressed page is greater than 75% of the PAGE_SIZE,
> > then the page is stored uncompressed in zram space. Zram space utilization
> > is improved if the threshold is 80%(5 compressed pages can be stored in
> > 4 pages).
> > 
> > If the compressed length is greater than 3068 and less than 3261, pages
> > still can be stored in compressed form in zs_malloc class 3264.
> > Currently these compressed pages belong to 4096 zs malloc class.
> 
> so this makes sense. I had another idea awhile ago
> 
> lkml.kernel.org/r/1456061274-20059-2-git-send-email-sergey.senozhatsky@gmail.com
> 
> in short, 3261 is good, but not as good as it possibly can be. for the
> time being, our huge-class watermark starts at 3264. but this can
> change.
> 
> 
> a side note, I think we have sort of wrong API. zsmalloc knows better which
> object is huge. and who knows, may be we will change the number of huge
> classes someday or huge-class watermark, etc. so having "hey zsmalloc, is
> this object huge or not" API seems to be better than ZRAM's enforcement
> "hey zsmalloc, this object is huge".

and yes, I think I'd like to reduce the number of huge classes
right now we store objects [3264+, 4096] in huge clases. with
this patch
	lkml.kernel.org/r/1456061274-20059-4-git-send-email-sergey.senozhatsky@gmail.com

we have extra classes and smaller huge-class-range. so we store
objects [3840+, 4096] in huge classes. less huge classes - more
compressed objects; more compressed objects - lower memory usage.

	-ss

Re: [PATCH 1/1] zram: better utilization of zram swap space

[Minchan Kim]

Hi Gopi and Sergey,

On Thu, Dec 07, 2017 at 05:45:10PM +0900, Sergey Senozhatsky wrote:
> On (12/07/17 13:52), Gopi Sai Teja wrote:
> > If the length of the compressed page is greater than 75% of the PAGE_SIZE,
> > then the page is stored uncompressed in zram space. Zram space utilization
> > is improved if the threshold is 80%(5 compressed pages can be stored in
> > 4 pages).
> > 
> > If the compressed length is greater than 3068 and less than 3261, pages
> > still can be stored in compressed form in zs_malloc class 3264.
> > Currently these compressed pages belong to 4096 zs malloc class.
> 
> so this makes sense. I had another idea awhile ago
> 
> lkml.kernel.org/r/1456061274-20059-2-git-send-email-sergey.senozhatsky@gmail.com
> 
> in short, 3261 is good, but not as good as it possibly can be. for the
> time being, our huge-class watermark starts at 3264. but this can
> change.
> 
> 
> a side note, I think we have sort of wrong API. zsmalloc knows better which
> object is huge. and who knows, may be we will change the number of huge
> classes someday or huge-class watermark, etc. so having "hey zsmalloc, is
> this object huge or not" API seems to be better than ZRAM's enforcement
> "hey zsmalloc, this object is huge".
> 
> 	-ss

I agree. zram shouldn't be aware of allocator internal.
It would be better for zram to use *int zs_max_zpage_size(struct zs_pool *pool)*
to set up max_zpage_size.
Let's hide the allocator's detail to the exported function.

Thanks.

Re: [PATCH 1/1] zram: better utilization of zram swap space

[Sergey Senozhatsky]

On (12/11/17 09:16), Minchan Kim wrote:
[..]
> I agree. zram shouldn't be aware of allocator internal.
> It would be better for zram to use *int zs_max_zpage_size(struct zs_pool *pool)*
> to set up max_zpage_size.
> Let's hide the allocator's detail to the exported function.

Hi Minchan,

cool :)

I'll re-spin the patch [in a day or two]. and then *may be* we can
return to the huge-class watermark proposition. may be there can be
a knob or something of that kind.

	-ss