Re: [PATCH v8 9/9] vmcore: support mmap() on /proc/vmcore

From: Maxim Uvarov <muvarov@gmail.com>
To: Andrew Morton <akpm@linux-foundation.org>
Cc: HATAYAMA Daisuke <d.hatayama@jp.fujitsu.com>,
	Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>,
	riel@redhat.com, kexec@lists.infradead.org, hughd@google.com,
	linux-kernel@vger.kernel.org, lisa.mitchell@hp.com,
	vgoyal@redhat.com, linux-mm@kvack.org,
	zhangyanfei@cn.fujitsu.com, ebiederm@xmission.com,
	kosaki.motohiro@jp.fujitsu.com, walken@google.com, cpw@sgi.com,
	jingbai.ma@hp.com
Subject: Re: [PATCH v8 9/9] vmcore: support mmap() on /proc/vmcore
Date: Tue, 2 Jul 2013 11:00:37 +0400	[thread overview]
Message-ID: <CAJGZr0+vcPmG7e54caYm+FM3g4NibZgut8Gf8n73A3HE3HKw3g@mail.gmail.com> (raw)
In-Reply-To: <20130701125345.c4a383c7b8345f9c5ae54023@linux-foundation.org>

[-- Attachment #1: Type: text/plain, Size: 2778 bytes --]

2013/7/1 Andrew Morton <akpm@linux-foundation.org>

> On Mon, 1 Jul 2013 18:34:43 +0400 Maxim Uvarov <muvarov@gmail.com> wrote:
>
> > 2013/7/1 HATAYAMA Daisuke <d.hatayama@jp.fujitsu.com>
> >
> > > (2013/06/29 1:40), Maxim Uvarov wrote:
> > >
> > >> Did test on 1TB machine. Total vmcore capture and save took 143
> minutes
> > >> while vmcore size increased from 9Gb to 59Gb.
> > >>
> > >> Will do some debug for that.
> > >>
> > >> Maxim.
> > >>
> > >
> > > Please show me your kdump configuration file and tell me what you did
> in
> > > the test and how you confirmed the result.
> > >
> > >
> > Hello Hatayama,
> >
> > I re-run tests in dev env. I took your latest kernel patchset from
> > patchwork for vmcore + devel branch of makedumpfile + fix to open and
> write
> > to /dev/null. Run this test on 1Tb memory machine with memory used by
> some
> > user space processes. crashkernel=384M.
> >
> > Please see my results for makedumpfile process work:
> > [gzip compression]
> > -c -d31 /dev/null
> > real 37.8 m
> > user 29.51 m
> > sys 7.12 m
> >
> > [no compression]
> > -d31 /dev/null
> > real 27 m
> > user 23 m
> > sys   4 m
> >
> > [no compression, disable cyclic mode]
> > -d31 --non-cyclic /dev/null
> > real 26.25 m
> > user 23 m
> > sys 3.13 m
> >
> > [gzip compression]
> > -c -d31 /dev/null
> > % time     seconds  usecs/call     calls    errors syscall
> > ------ ----------- ----------- --------- --------- ----------------
> >  54.75   38.840351         110    352717           mmap
> >  44.55   31.607620          90    352716         1 munmap
> >   0.70    0.497668           0  25497667           brk
> >   0.00    0.000356           0    111920           write
> >   0.00    0.000280           0    111904           lseek
> >   0.00    0.000025           4         7           open
> >   0.00    0.000000           0       473           read
> >   0.00    0.000000           0         7           close
> >   0.00    0.000000           0         3           fstat
> >   0.00    0.000000           0         1           getpid
> >   0.00    0.000000           0         1           execve
> >   0.00    0.000000           0         1           uname
> >   0.00    0.000000           0         2           unlink
> >   0.00    0.000000           0         1           arch_prctl
> > ------ ----------- ----------- --------- --------- ----------------
> > 100.00   70.946300              26427420         1 total
> >
>
> I have no point of comparison here.  Is this performance good, or is
> the mmap-based approach still a lot more expensive?
>
>
> Compressing to non-mmap version improvement is 30 minutes against 130
minutes for total dump process. And kernel load is very minimal. So
definitely we need these patches.

-- 
Best regards,
Maxim Uvarov

[-- Attachment #2: Type: text/html, Size: 3646 bytes --]