Re: [RFC 0/4] ZRAM: make it just store the high compression rate page

[Hui Zhu]

On Mon, Sep 5, 2016 at 10:18 AM, Minchan Kim <minchan@xxxxxxxxxx> wrote:
> On Thu, Aug 25, 2016 at 04:25:30PM +0800, Hui Zhu wrote:
>> On Thu, Aug 25, 2016 at 2:09 PM, Sergey Senozhatsky
>> <sergey.senozhatsky.work@xxxxxxxxx> wrote:
>> > Hello,
>> >
>> > On (08/22/16 16:25), Hui Zhu wrote:
>> >>
>> >> Current ZRAM just can store all pages even if the compression rate
>> >> of a page is really low.  So the compression rate of ZRAM is out of
>> >> control when it is running.
>> >> In my part, I did some test and record with ZRAM.  The compression rate
>> >> is about 40%.
>> >>
>> >> This series of patches make ZRAM can just store the page that the
>> >> compressed size is smaller than a value.
>> >> With these patches, I set the value to 2048 and did the same test with
>> >> before.  The compression rate is about 20%.  The times of lowmemorykiller
>> >> also decreased.
>> >
>> > I haven't looked at the patches in details yet. can you educate me a bit?
>> > is your test stable? why the number of lowmemorykill-s has decreased?
>> > ... or am reading "The times of lowmemorykiller also decreased" wrong?
>> >
>> > suppose you have X pages that result in bad compression size (from zram
>> > point of view). zram stores such pages uncompressed, IOW we have no memory
>> > savings - swapped out page lands in zsmalloc PAGE_SIZE class. now you
>> > don't try to store those pages in zsmalloc, but keep them as unevictable.
>> > so the page still occupies PAGE_SIZE; no memory saving again. why did it
>> > improve LMK?
>>
>> No, zram will not save this page uncompressed with these patches.  It
>> will set it as non-swap and kick back to shrink_page_list.
>> Shrink_page_list will remove this page from swapcache and kick it to
>> unevictable list.
>> Then this page will not be swaped before it get write.
>> That is why most of code are around vmscan.c.
>
> If I understand Sergey's point right, he means there is no gain
> to save memory between before and after.
>
> With your approach, you can prevent unnecessary pageout(i.e.,
> uncompressible page swap out) but it doesn't mean you save the
> memory compared to old so why does your patch decrease the number of
> lowmemory killing?
>
> A thing I can imagine is without this feature, zram could be full of
> uncompressible pages so good-compressible page cannot be swapped out.
> Hui, is this scenario right for your case?
>

That is one reason.  But it is not the principal one.

Another reason is when swap is running to put page to zram, what the
system wants is to get memory.
Then the deal is system spends cpu time and memory to get memory. If
the zram just access the high compression rate pages, system can get
more memory with the same amount of memory. It will pull system from
low memory status earlier. (Maybe more cpu time, because the
compression rate checks. But maybe less, because fewer pages need to
digress. That is the interesting part. :)
I think that is why lmk times decrease.

And yes, all of this depends on the number of high compression rate
pages. So you cannot just set a non_swap limit to the system and get
everything. You need to do a lot of test around it to make sure the
non_swap limit is good for your system.

And I think use AOP_WRITEPAGE_ACTIVATE without kicking page to a
special list will make cpu too busy sometimes.
I did some tests before I kick page to a special list. The shrink task
will be moved around, around and around because low compression rate
pages just moved from one list to another a lot of times, again, again
and again.
And all this low compression rate pages always stay together.

Thanks,
Hui


> Thanks.