Date:	Mon, 27 May 2013 17:12:56 +0900
From:	Minchan Kim <minchan@...nel.org>
To:	Andrew Morton <akpm@...ux-foundation.org>
Cc:	linux-mm@...ck.org, linux-kernel@...r.kernel.org,
	Rik van Riel <riel@...hat.com>,
	Michael Kerrisk <mtk.manpages@...il.com>,
	Dave Hansen <dave.hansen@...el.com>,
	Namhyung Kim <namhyung@...nel.org>,
	Minkyung Kim <minkyung88@....com>
Subject: Re: [PATCH v5 0/7] Per process reclaim

Hello Andrew,

Sorry for the late response.
It was bad timing because I was off during last week.

On Tue, May 21, 2013 at 04:16:56PM -0700, Andrew Morton wrote:
> On Thu,  9 May 2013 16:21:22 +0900 Minchan Kim <minchan@...nel.org> wrote:
> 
> > These day, there are many platforms avaiable in the embedded market
> > and they are smarter than kernel which has very limited information
> > about working set so they want to involve memory management more heavily
> > like android's lowmemory killer and ashmem or recent many lowmemory
> > notifier(there was several trial for various company NOKIA, SAMSUNG,
> > Linaro, Google ChromeOS, Redhat).
> > 
> > One of the simple imagine scenario about userspace's intelligence is that
> > platform can manage tasks as forground and backgroud so it would be
> > better to reclaim background's task pages for end-user's *responsibility*
> > although it has frequent referenced pages.
> > 
> > The patch[1] prepares that force_reclaim in shrink_page_list can
> > handle anonymous pages as well as file-backed pages.
> > 
> > The patch[2] adds new knob "reclaim under proc/<pid>/" so task manager
> > can reclaim any target process anytime, anywhere. It could give another
> > method to platform for using memory efficiently.
> > 
> > It can avoid process killing for getting free memory, which was really
> > terrible experience because I lost my best score of game I had ever
> > after I switch the phone call while I enjoyed the game.
> > 
> > Reclaim file-backed pages only.
> > 	echo file > /proc/PID/reclaim
> > Reclaim anonymous pages only.
> > 	echo anon > /proc/PID/reclaim
> > Reclaim all pages
> > 	echo all > /proc/PID/reclaim
> 
> Oh boy.  I think I do agree with the overall intent, but there are so
> many ways of doing this.
> 
> - Do we reclaim the pages altogether, or should we just give them one
>   round of aging?  If the latter then you'd need to run "echo anon >
>   /proc/PID/reclaim" four times to firmly whack the pages, but that's
>   more flexible.

The concern of idea you suggested is that it's tied up to kernel's
reclaim implementation(active/inactive/referenced) so if we changes
the logic, user's behavior would be broken easily.

> 
> - Why do it via the pid at all?  Would it be better to instead do
>   this to a memcg and require that the admin put these processes into
>   memcgs?  In fact existing memcg controls could get us at least
>   partway to this feature.

Hmm, I don't have an idea to handle this with memcg.

1) Is there any way to reclaim anon pages as top prioirty
   rather than file pages? Many embedded system uses zram as swap device
   so somecase, anon reclaim could be cheaper rather than file-backd pages
   reclaim on really slow storage.
   Of course, we can control it with "swappiness" but at most, it makes
   balancing ratio 1:1.
   Should we expand the value to 200? I have no idea.
   What I want is VM smarter that it can understand swap storage's speed
   so that it can balance anon/file more smart. But it's not a simple
   problem, I think.
   
2) Another usecase,
   We used per process reclaim to make smart snapshot image which is
   like hibernation image. As we know well, embedded system's boot time
   is very critical to lunch the product so a few of system have used
   snapshot booting. Snapshot booting's core technology is to minimise
   snapshot image size but it includes working set pages which will
   be needed soon after resuming so we can profile working set pages
   ,discard unnecessary pages via per-process reclaim, then make snapshot
   images.

 
> 
> - I don't understand the need for "Enhance per process reclaim to
>   consider shared pages".  If "echo file > /proc/PID/reclaim" causes
>   PID's mm's file-backed pte's to be unmapped (which seems to be the
>   correct effect) then we get this automatically: unshared file pages
>   will be freed and shared file pages will remain in core until the
>   other sharing process's also unmap them.

In current implemenation, even shared page is detached from all of
processes and is freed.

> 
> 
> Overall, I'm unsure whether/how to proceed with this.  I'd like to hear
> from a lot of the potential users, and hear them say "yes, we can use
> this".

We already used it for creating snapshot image and have a plan to use it
for reducing system latency and avoiding frequent OOM killing.

> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@...ck.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@...ck.org"> email@...ck.org </a>

-- 
Kind regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists