From: Solofo.Ramangalahy@bull.net
Subject: Re: jbd/jbd2 performance improvements
Date: Thu, 23 Oct 2008 12:42:48 +0200
Message-ID: <18688.21800.641991.52484@frecb006361.adech.frec.bull.fr>
References: <E1KnCO6-0001qG-SV@closure.thunk.org>
	<48F62893.9060606@redhat.com>
	<18678.55651.556822.187508@frecb006361.adech.frec.bull.fr>
	<48F72E5F.2050409@redhat.com>
	<48F735E8.7060803@redhat.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Cc: Ric Wheeler <rwheeler@redhat.com>, "Theodore Ts'o" <tytso@mit.edu>,
	linux-ext4@vger.kernel.org
To: Eric Sandeen <sandeen@redhat.com>
Return-path: <linux-ext4-owner@vger.kernel.org>
Received: from ecfrec.frec.bull.fr ([129.183.4.8]:33009 "EHLO
	ecfrec.frec.bull.fr" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1752366AbYJWKu7 (ORCPT
	<rfc822;linux-ext4@vger.kernel.org>); Thu, 23 Oct 2008 06:50:59 -0400
In-Reply-To: <48F735E8.7060803@redhat.com>
Sender: linux-ext4-owner@vger.kernel.org
List-ID: <linux-ext4.vger.kernel.org>

>>>>> On Thu, 16 Oct 2008 07:39:04 -0500, Eric Sandeen <sandeen@redhat.com> said:
    >> A very thorough test, but the results don't seem to point to a
    >> consistent winner.
    >> 
    >> I agree that running without KVM in the picture might be very
    >> interesting. Eric has some similar tests underway, I think that
    >> his results were also inconclusive so far...

    Eric> Yep, I've yet to find an fs_mark invocation, at least, which
    Eric> shows a clear winner.  I also ran w/ akpm's suggested
    Eric> io_schedule watcher patch and never see us waiting on this
    Eric> lock (I did set it to 1s though, which is probably too long
    Eric> for my storage).

I've redone the tests without kvm. Still no clear winner

To sum up:
. kernel ext4-stable
. mkfs (1.41.3) default options
. mount options: default, akpm, akpm_lock_hack
. scheduler default (cfq)
. 8 cpus, single 15K rpm disk.
. without the high latency detection patch
. a broad range of fs_mark (all the sync strategies, from 1 to 32
  threads, up to 10000 files/thread, several directories).
. a "tangled synchrony" workload as mentionned in the "Analysis and
  evolution of journaling file systems" paper discussed monday.

First things first, maybe I should have spent more time
reproducing Arjan behavior before testing.

This was not a complete waste of time though, as the following errors
were spotted during the runs:
1. EXT4-fs error (device sdb): ext4_free_inode: bit already cleared for inode 32769
2. EXT4-fs error (device sdb): ext4_init_inode_bitmap: Checksum bad for group 8
3. BUG: spinlock wrong CPU on CPU#3, fs_mark/1975
 lock: ffff88015a44f480, .magic: dead4ead, .owner: fs_mark/1975, .owner_cpu: 1
Pid: 1975, comm: fs_mark Not tainted 2.6.27.1-ext4-stable-gcov #1

Call Trace:
 [<ffffffff811a47a2>] spin_bug+0xa2/0xaa
 [<ffffffff811a481f>] _raw_spin_unlock+0x75/0x8a
 [<ffffffff814552c1>] _spin_unlock+0x26/0x2a
 [<ffffffffa00d4fd3>] ext4_read_inode_bitmap+0xfa/0x14e [ext4]
 [<ffffffffa00d564b>] ext4_new_inode+0x5d4/0xec4 [ext4]
 [<ffffffff810562db>] ? __lock_acquire+0x481/0x7d8
 [<ffffffffa00c2430>] ? jbd2_journal_start+0xef/0x11a [jbd2]
 [<ffffffffa00c2430>] ? jbd2_journal_start+0xef/0x11a [jbd2]
 [<ffffffffa00deb99>] ext4_create+0xc7/0x144 [ext4]
 [<ffffffff810b6734>] vfs_create+0xdf/0x155
 [<ffffffff810b8905>] do_filp_open+0x220/0x7fc
 [<ffffffff814552c1>] ? _spin_unlock+0x26/0x2a
 [<ffffffff810abe5a>] do_sys_open+0x53/0xd3
 [<ffffffff810abf03>] sys_open+0x1b/0x1d
 [<ffffffff8100bf8b>] system_call_fastpath+0x16/0x1b
 
Anybody seen this in their logs?

The "bit already cleared for inode" is triggered by:
fs_mark -v -d /mnt/test-ext4 -n10000 -D10 -N1000 -t8 -s4096 -S0

-- 
solofo