linux-kernel - Re: [RFC] Low overhead patches for the memory cgroup controller (v2)

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20090519131857.GC4858@balbir.in.ibm.com>
Date:	Tue, 19 May 2009 21:18:57 +0800
From:	Balbir Singh <balbir@...ux.vnet.ibm.com>
To:	KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
Cc:	"linux-mm@...ck.org" <linux-mm@...ck.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	"nishimura@....nes.nec.co.jp" <nishimura@....nes.nec.co.jp>,
	"lizf@...fujitsu.com" <lizf@...fujitsu.com>,
	"menage@...gle.com" <menage@...gle.com>,
	KOSAKI Motohiro <m-kosaki@...es.dti.ne.jp>
Subject: Re: [RFC] Low overhead patches for the memory cgroup controller
	(v2)

* KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com> [2009-05-19 01:01:00]:

> Balbir Singh wrote:
> > * KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com> [2009-05-18
> > 19:11:07]:
> >
> >> On Fri, 15 May 2009 23:46:39 +0530
> >> Balbir Singh <balbir@...ux.vnet.ibm.com> wrote:
> >>
> >> > * KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com> [2009-05-16
> >> 02:45:03]:
> >> >
> >> > > Balbir Singh wrote:
> >> > > > Feature: Remove the overhead associated with the root cgroup
> >> > > >
> >> > > > From: Balbir Singh <balbir@...ux.vnet.ibm.com>
> >> > > >
> >> > > > This patch changes the memory cgroup and removes the overhead
> >> associated
> >> > > > with LRU maintenance of all pages in the root cgroup. As a
> >> side-effect, we
> >> > > > can
> >> > > > no longer set a memory hard limit in the root cgroup.
> >> > > >
> >> > > > A new flag is used to track page_cgroup associated with the root
> >> cgroup
> >> > > > pages. A new flag to track whether the page has been accounted or
> >> not
> >> > > > has been added as well.
> >> > > >
> >> > > > Review comments higly appreciated
> >> > > >
> >> > > > Tests
> >> > > >
> >> > > > 1. Tested with allocate, touch and limit test case for a non-root
> >> cgroup
> >> > > > 2. For the root cgroup tested performance impact with reaim
> >> > > >
> >> > > >
> >> > > > 		+patch		mmtom-08-may-2009
> >> > > > AIM9		1362.93		1338.17
> >> > > > Dbase		17457.75	16021.58
> >> > > > New Dbase	18070.18	16518.54
> >> > > > Shared		9681.85		8882.11
> >> > > > Compute		16197.79	15226.13
> >> > > >
> >> > > Hmm, at first impression, I can't convice the numbers...
> >> > > Just avoiding list_add/del makes programs _10%_ faster ?
> >> > > Could you show changes in cpu cache-miss late if you can ?
> >> > > (And why Aim9 goes bad ?)
> >> >
> >> > OK... I'll try but I am away on travel for 3 weeks :( you can try and
> >> run
> >> > this as well
> >> >
> >> tested aim7 with some config.
> >>
> >> CPU: Xeon 3.1GHz/4Core x2 (8cpu)
> >> Memory: 32G
> >> HDD: Usual? Scsi disk (just 1 disk)
> >> (try_to_free_pages() etc...will never be called.)
> >>
> >> Multiuser config. #of tasks 1100 (near to peak on my host)
> >>
> >> 10runs.
> >> rc6mm1 score(Jobs/min)
> >> 44009.1 44844.5 44691.1 43981.9 44992.6
> >> 44544.9 44179.1 44283.0 44442.9 45033.8  average=44500
> >>
> >> +patch
> >> 44656.8 44270.8 44706.7 44106.1 44467.6
> >> 44585.3 44167.0 44756.7 44853.9 44249.4  average=44482
> >>
> >> Dbase config. #of tasks 25
> >> rc6mm1 score (jobs/min)
> >> 11022.7 11018.9 11037.9 11003.8 11087.5
> >> 11145.2 11133.6 11068.3 11091.3 11106.6 average=11071
> >>
> >> +patch
> >> 10888.0 10973.7 10913.9 11000.0 10984.9
> >> 10996.2 10969.9 10921.3 10921.3 11053.1 average=10962
> >>
> >> Hmm, 1% improvement ?
> >> (I think this is reasonable score of the effect of this patch)
> >>
> >
> > Thanks for the test, I have a 4 CPU system and I create 80 users,
> > larger config shows larger difference at my end.
> Sorry, above Dbase test was on 54 threads. I'll try 20*8=160 threads.
>

cool! Thanks
 
> > I think even 1% is
> > quite reasonable as you mentioned. If the patch looks fine, should we
> > ask for larger testing by Andrew?
> >
> Hmm, as you like. My interest is bugfix for swap leaking now.

I've seen that too.. I think that has been going on for long and I am
afraid it is hurting features like soft limit, but bug fixing is
important. Hopefully we'll have a good solution soon.

> Because this change adds big special case, we need much tests, anyway.
> And please show _environment_ where benchmarks run.
> BTW, I wonder whetere we can have more improvements in this special case...
> 
> >> Anyway, I'm afraid of difference between mine and your kernel config.
> >> plz enjoy your travel for now :)
> >
> > Sorry, I did not send you my .config, why do you think .config makes a
> > difference?
> I wanted to know what kind of DEBUG/TRACE config is on. and some others.
> 
> > I think loading AIM makes the difference and I also made
> > one other change to the aim tests. I run with "sync" linked to
> > /bin/true and use tmpfs for temporary partition and 20*numnber of cpus
> > for number of users.
> >
> Is it usual method at using AIM ? (Sorry, I'm not sure).
> It seems to break AIM7's purpose of "measuring typical workload"...
> 

No.. it is not.. but sync has a large overhead, so I use /bin/true. I
can try without it and report back.


> > If required, I can still send out my .config to you.
> >
> If you can, plz. (just for my interest ;)
>

Attached, please see 

-- 
	Balbir

View attachment "config-2.6.30-rc4-mm1" of type "text/plain" (54828 bytes)