From: Minchan Kim <minchan@kernel.org>
To: Johannes Weiner <hannes@cmpxchg.org>
Cc: Zheng Liu <gnehzuil.liu@gmail.com>, Michal Hocko <mhocko@suse.cz>,
linux-mm@kvack.org, linux-kernel@vger.kernel.org,
Andrew Morton <akpm@linux-foundation.org>,
Mel Gorman <mel@csn.ul.ie>,
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>,
Rik van Riel <riel@redhat.com>
Subject: Re: [PATCH] mm: consider all swapped back pages in used-once logic
Date: Mon, 25 Jun 2012 17:25:56 +0900 [thread overview]
Message-ID: <4FE82094.8090002@kernel.org> (raw)
In-Reply-To: <20120625080832.GX27816@cmpxchg.org>
On 06/25/2012 05:08 PM, Johannes Weiner wrote:
> On Mon, Jun 25, 2012 at 08:53:11AM +0900, Minchan Kim wrote:
>> Hi Hannes,
>>
>> On 06/23/2012 08:04 PM, Johannes Weiner wrote:
>>
>>> On Mon, May 21, 2012 at 07:07:00PM +0800, Zheng Liu wrote:
>>>> On Mon, May 21, 2012 at 11:37:05AM +0200, Johannes Weiner wrote:
>>>> [snip]
>>>>>>> Is it because the read()/write() IO is high throughput and pushes
>>>>>>> pages through the LRU lists faster than the mmap pages are referenced?
>>>>>>
>>>>>> Yes, in this application, one query needs to access mapped file page
>>>>>> twice and file page cache twice. Namely, one query needs to do 4 disk
>>>>>> I/Os. We have used fadvise(2) to reduce file page cache accessing to
>>>>>> only once. For mapped file page, in fact them are accessed only once
>>>>>> because in one query the same data is accessed twice. Thus, one query
>>>>>> causes 2 disk I/Os now. The size of read/write is quite larger than
>>>>>> mmap/munmap. So, as you see, if we can keep mmap/munmap file in memory
>>>>>> as much as possible, we will gain the better performance.
>>>>>
>>>>> You access the same unmapped cache twice, i.e. repeated reads or
>>>>> writes against the same file offset?
>>>>
>>>> No. We access the same mapped file twice.
>>>>
>>>>>
>>>>> How do you use fadvise?
>>>>
>>>> We access the header and content of the file respectively using read/write.
>>>> The header and content are sequentially. So we use fadivse(2) with
>>>> FADV_WILLNEED flag to do a readahead.
>>>>
>>>>>> In addition, another factor also has some impacts for this application.
>>>>>> In inactive_file_is_low_global(), it is different between 2.6.18 and
>>>>>> upstream kernel. IMHO, it causes that mapped file pages in active list
>>>>>> are moved into inactive list frequently.
>>>>>>
>>>>>> Currently, we add a parameter in inactive_file_is_low_global() to adjust
>>>>>> this ratio. Meanwhile we activate every mapped file pages for the first
>>>>>> time. Then the performance gets better, but it still doesn't reach the
>>>>>> performance of 2.6.18.
>>>>>
>>>>> 2.6.18 didn't have the active list protection at all and always
>>>>> forcibly deactivated pages during reclaim. Have you tried fully
>>>>> reverting to this by making inactive_file_is_low_global() return true
>>>>> unconditionally?
>>>>
>>>> No, I don't try it. AFAIK, 2.6.18 didn't protect the active list. But
>>>> it doesn't always forcibly deactivate the pages. I remember that in
>>>> 2.6.18 kernel we calculate 'mapped_ratio' in shrink_active_list(), and
>>>> then we get 'swap_tendency' according to 'mapped_ratio', 'distress', and
>>>> 'sc->swappiness'. If 'swap_tendency' is not greater than 100. It
>>>> doesn't reclaim mapped file pages. By this equation, if the sum of the
>>>> anonymous pages and mapped file pages is not greater than the 50% of
>>>> total pages, we don't deactivate these pages. Am I missing something?
>>>
>>> I think we need to go back to protecting mapped pages based on how
>>> much of reclaimable memory they make up, one way or another.
>>
>>
>> I partly agreed it with POV regression.
>> But I would like to understand rationale of "Why we should handle specially mmapped page".
>> In case of code pages(VM_EXEC), we already have handled it specially and
>> I understand why we did. At least, my opinion was that our LRU algorithm doesn't consider
>> _frequency_ fully while it does _recency_ well. I thought code page would be high frequency of access
>> compared to other pages.
>> But in case of mapped data pages, why we should handle specially?
>> I guess mapped data pages would have higher access chance than unmapped page because
>> unmapped page doesn't have any owner(it's just for caching for reducing I/O) while mapped page
>> has a owner above.
>>
>> Doesn't it make sense?
>
> I agree that the reason behind protecting VM_EXEC pages was that our
> frequency information for mapped pages is at LRU cycle granularity.
>
> But I don't see why you think this problem wouldn't apply to all
> mapped pages in general.
Code page is very likely to share by other processes so I think it's very special
than normal mmaped page. So I would like to raise bonus on code page than normal mmaped pages.
So I would like to make following as if we can.
Reclaim preference :
unmapped page >> mapped page > VM_EXEC mapped page
But at least, we can't solve Zheng's regression with current VM_EXEC protection logic
because it seems he already used VM_EXEC tric :(
I hope Erecalaimbe LRU list can solve it.
--
Kind regards,
Minchan Kim
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
next prev parent reply other threads:[~2012-06-25 8:25 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2012-05-17 9:13 [PATCH] mm: consider all swapped back pages in used-once logic Michal Hocko
2012-05-17 9:24 ` Andrew Morton
2012-05-17 12:10 ` Michal Hocko
2012-05-17 20:23 ` Andrew Morton
2012-05-18 6:50 ` Michal Hocko
2012-05-17 13:14 ` Rik van Riel
2012-05-17 19:54 ` Johannes Weiner
2012-05-21 2:51 ` Zheng Liu
2012-05-21 7:36 ` Johannes Weiner
2012-05-21 8:59 ` Zheng Liu
2012-05-21 9:37 ` Johannes Weiner
2012-05-21 11:07 ` Zheng Liu
2012-06-23 11:04 ` Johannes Weiner
2012-06-23 15:22 ` Rik van Riel
2012-06-24 23:53 ` Minchan Kim
2012-06-25 8:08 ` Johannes Weiner
2012-06-25 8:25 ` Minchan Kim [this message]
2012-06-26 13:51 ` Johannes Weiner
2012-06-26 23:47 ` Minchan Kim
2012-05-18 0:40 ` Minchan Kim
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=4FE82094.8090002@kernel.org \
--to=minchan@kernel.org \
--cc=akpm@linux-foundation.org \
--cc=gnehzuil.liu@gmail.com \
--cc=hannes@cmpxchg.org \
--cc=kamezawa.hiroyu@jp.fujitsu.com \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=mel@csn.ul.ie \
--cc=mhocko@suse.cz \
--cc=riel@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).