Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <39f59994-cf2e-f8c4-fd57-988e8c7e05e5@molgen.mpg.de>
Date:   Wed, 31 Aug 2022 13:59:38 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.2.0
Subject: Re: [PATCH -next 0/3] md/raid10: reduce lock contention for io
Content-Language: en-US
To:     Yu Kuai <yukuai1@huaweicloud.com>
Cc:     song@kernel.org, linux-raid@vger.kernel.org,
        linux-kernel@vger.kernel.org, yi.zhang@huawei.com,
        yukuai3@huawei.com
References: <20220829131502.165356-1-yukuai1@huaweicloud.com>
 <70172016-8ae5-92b8-2a43-8f0d4a2f6668@molgen.mpg.de>
 <0e22c51b-d36a-2216-4bfd-522049d3ed44@huaweicloud.com>
From:   Paul Menzel <pmenzel@molgen.mpg.de>
In-Reply-To: <0e22c51b-d36a-2216-4bfd-522049d3ed44@huaweicloud.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk

Dear Yu,


Am 30.08.22 um 03:09 schrieb Yu Kuai:

> 在 2022/08/29 21:58, Paul Menzel 写道:

>> Am 29.08.22 um 15:14 schrieb Yu Kuai:
>>> From: Yu Kuai <yukuai3@huawei.com>
>>>
>>> patch 1 is a small problem found by code review.
>>> patch 2 avoid holding resync_lock in fast path.
>>> patch 3 avoid holding lock in wake_up() in fast path.
>>>
>>> Test environment:
>>>
>>> Architecture: aarch64
>>> Cpu: Huawei KUNPENG 920, there are four numa nodes
>>>
>>> Raid10 initialize:
>>> mdadm --create /dev/md0 --level 10 --bitmap none --raid-devices 4 
>>> /dev/nvme0n1 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme3n1
>>>
>>> Test cmd:
>>> fio -name=0 -ioengine=libaio -direct=1 -group_reporting=1 
>>> -randseed=2022 -rwmixread=70 -refill_buffers -filename=/dev/md0 
>>> -numjobs=16 -runtime=60s -bs=4k -iodepth=256 -rw=randread
>>>
>>> Test result:
>>> before this patchset:    2.9 GiB/s
>>> after this patchset:    6.6 Gib/s
>>
>> Could you please give more details about the test setup, like the 
>> drives used?
> 
> test setup is described above, four nvme disks is used.

I was wondering about the model to be able to reproduce it.

>> Did you use some tools like ftrace to figure out the bottleneck?
> 
> Yes, I'm sure the bottleneck is spin_lock(), specifically threads from
> multiple nodes try to grab the same lock. By the way, if I bind the
> threads to the same node, performance can also improve to 6.6 Gib/s
> without this patchset.

Interesting. Maybe you could add all that to the commit message of the 
second patch.


Kind regards,

Paul


>>> Please noted that in kunpeng-920, memory access latency is very bad
>>> accross nodes compare to local node, and in other architecture
>>> performance improvement might not be significant.
>>>
>>> Yu Kuai (3):
>>>    md/raid10: fix improper BUG_ON() in raise_barrier()
>>>    md/raid10: convert resync_lock to use seqlock
>>>    md/raid10: prevent unnecessary calls to wake_up() in fast path
>>>
>>>   drivers/md/raid10.c | 88 +++++++++++++++++++++++++++++----------------
>>>   drivers/md/raid10.h |  2 +-
>>>   2 files changed, 59 insertions(+), 31 deletions(-)