Re: [PATCH 1/2] mm/slub: Introduce two counters for the partial objects

From: xunlei <xlpang@linux.alibaba.com>
To: Pekka Enberg <penberg@gmail.com>
Cc: Christoph Lameter <cl@linux.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	Wen Yang <wenyang@linux.alibaba.com>,
	Yang Shi <yang.shi@linux.alibaba.com>,
	Roman Gushchin <guro@fb.com>,
	"linux-mm@kvack.org" <linux-mm@kvack.org>,
	LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 1/2] mm/slub: Introduce two counters for the partial objects
Date: Fri, 31 Jul 2020 10:57:38 +0800	[thread overview]
Message-ID: <5eeb5c3d-1a34-ad96-9010-4d8a5ac32241@linux.alibaba.com> (raw)
In-Reply-To: <CAOJsxLFnY=4v6UQigyiZKTMTQXKakVOKf6KA+bCkMe-XVY6sqA@mail.gmail.com>

On 2020/7/7 下午11:23, Pekka Enberg wrote:
> Hi!
> 
> (Sorry for the delay, I missed your response.)
> 
> On Fri, Jul 3, 2020 at 12:38 PM xunlei <xlpang@linux.alibaba.com> wrote:
>>
>> On 2020/7/2 PM 7:59, Pekka Enberg wrote:
>>> On Thu, Jul 2, 2020 at 11:32 AM Xunlei Pang <xlpang@linux.alibaba.com> wrote:
>>>> The node list_lock in count_partial() spend long time iterating
>>>> in case of large amount of partial page lists, which can cause
>>>> thunder herd effect to the list_lock contention, e.g. it cause
>>>> business response-time jitters when accessing "/proc/slabinfo"
>>>> in our production environments.
>>>
>>> Would you have any numbers to share to quantify this jitter? I have no
>>
>> We have HSF RT(High-speed Service Framework Response-Time) monitors, the
>> RT figures fluctuated randomly, then we deployed a tool detecting "irq
>> off" and "preempt off" to dump the culprit's calltrace, capturing the
>> list_lock cost up to 100ms with irq off issued by "ss", this also caused
>> network timeouts.
> 
> Thanks for the follow up. This sounds like a good enough motivation
> for this patch, but please include it in the changelog.
> 
>>> objections to this approach, but I think the original design
>>> deliberately made reading "/proc/slabinfo" more expensive to avoid
>>> atomic operations in the allocation/deallocation paths. It would be
>>> good to understand what is the gain of this approach before we switch
>>> to it. Maybe even run some slab-related benchmark (not sure if there's
>>> something better than hackbench these days) to see if the overhead of
>>> this approach shows up.
>>
>> I thought that before, but most atomic operations are serialized by the
>> list_lock. Another possible way is to hold list_lock in __slab_free(),
>> then these two counters can be changed from atomic to long.
>>
>> I also have no idea what's the standard SLUB benchmark for the
>> regression test, any specific suggestion?
> 
> I don't know what people use these days. When I did benchmarking in
> the past, hackbench and netperf were known to be slab-allocation
> intensive macro-benchmarks. Christoph also had some SLUB
> micro-benchmarks, but I don't think we ever merged them into the tree.

I tested hackbench on 24-CPU machine, here are the results:

"hackbench 20 thread 1000"

== orignal(without any patch)
Time: 53.793
Time: 54.305
Time: 54.073

== with my patch1~2
Time: 54.036
Time: 53.840
Time: 54.066
Time: 53.449

== with my patch1~2, plus using a percpu partial free objects counter
Time: 53.303
Time: 52.994
Time: 53.218
Time: 53.268
Time: 53.739
Time: 53.072

The results show no performance regression, it's strange that the
figures even get a little better when using percpu counter.

Thanks,
Xunlei