Message-ID: <56CF1322.2040609@hpe.com>
Date:	Thu, 25 Feb 2016 09:43:46 -0500
From:	Waiman Long <waiman.long@....com>
To:	Ingo Molnar <mingo@...nel.org>
CC:	Jan Kara <jack@...e.cz>, Alexander Viro <viro@...iv.linux.org.uk>,
	Jan Kara <jack@...e.com>,
	Jeff Layton <jlayton@...chiereds.net>,
	"J. Bruce Fields" <bfields@...ldses.org>,
	Tejun Heo <tj@...nel.org>,
	Christoph Lameter <cl@...ux-foundation.org>,
	<linux-fsdevel@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
	Ingo Molnar <mingo@...hat.com>,
	Peter Zijlstra <peterz@...radead.org>,
	Andi Kleen <andi@...stfloor.org>,
	Dave Chinner <dchinner@...hat.com>,
	Scott J Norton <scott.norton@...com>,
	Douglas Hatch <doug.hatch@...com>
Subject: Re: [PATCH v3 3/3] vfs: Use per-cpu list for superblock's inode list

On 02/25/2016 03:06 AM, Ingo Molnar wrote:
> * Jan Kara<jack@...e.cz>  wrote:
>
>>>>> With an exit microbenchmark that creates a large number of threads,
>>>>> attachs many inodes to them and then exits. The runtimes of that
>>>>> microbenchmark with 1000 threads before and after the patch on a 4-socket
>>>>> Intel E7-4820 v3 system (40 cores, 80 threads) were as follows:
>>>>>
>>>>>    Kernel            Elapsed Time    System Time
>>>>>    ------            ------------    -----------
>>>>>    Vanilla 4.5-rc4      65.29s         82m14s
>>>>>    Patched 4.5-rc4      22.81s         23m03s
>>>>>
>>>>> Before the patch, spinlock contention at the inode_sb_list_add() function
>>>>> at the startup phase and the inode_sb_list_del() function at the exit
>>>>> phase were about 79% and 93% of total CPU time respectively (as measured
>>>>> by perf). After the patch, the percpu_list_add() function consumed only
>>>>> about 0.04% of CPU time at startup phase. The percpu_list_del() function
>>>>> consumed about 0.4% of CPU time at exit phase. There were still some
>>>>> spinlock contention, but they happened elsewhere.
>>>> While looking through this patch, I have noticed that the
>>>> list_for_each_entry_safe() iterations in evict_inodes() and
>>>> invalidate_inodes() are actually unnecessary. So if you first apply the
>>>> attached patch, you don't have to implement safe iteration variants at all.
>>>>
>>>> As a second comment, I'd note that this patch grows struct inode by 1
>>>> pointer. It is probably acceptable for large machines given the speedup but
>>>> it should be noted in the changelog. Furthermore for UP or even small SMP
>>>> systems this is IMHO undesired bloat since the speedup won't be noticeable.
>>>>
>>>> So for these small systems it would be good if per-cpu list magic would just
>>>> fall back to single linked list with a spinlock. Do you think that is
>>>> reasonably doable?
>>> Even many 'small' systems tend to be SMP these days.
>> Yes, I know. But my tablet with 4 ARM cores is unlikely to benefit from this
>> change either. [...]
> I'm not sure about that at all, the above numbers are showing a 3x-4x speedup in
> system time, which ought to be noticeable on smaller SMP systems as well.
>
> Waiman, could you please post the microbenchmark?
>
> Thanks,
>
> 	Ingo

The microbenchmark that I used is attached.

I do agree that performance benefit will decrease as the number of CPUs 
get smaller. The system that I used for testing have 4 sockets with 40 
cores (80 threads). Dave Chinner had run his fstests on a 16-core system 
(probably 2-socket) which showed modest improvement in performance 
(~4m40s vs 4m30s in runtime).

This patch enables parallel insertion and deletion to/from the inode 
list which used to be a serialized operation. So if that list operation 
is a bottleneck, you will see significant improvement. If it is not, we 
may not notice that much of a difference. For a single-socket 4-core 
system, I agree that the performance benefit, if any, will be limited.

Cheers,
Longman


View attachment "exit_test.c" of type "text/plain" (2666 bytes)

Powered by blists - more mailing lists