Subject: Re: [PATCH v2 00/12] mm: page migration enhancement for thp
To: Zi Yan <zi.yan@cs.rutgers.edu>,
        Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
References: <1478561517-4317-1-git-send-email-n-horiguchi@ah.jp.nec.com>
 <5822FB60.5040905@linux.vnet.ibm.com>
 <20161109235223.GA31285@hori1.linux.bs1.fc.nec.co.jp>
 <D34FA575-7C5D-4E9D-B337-A925F1A89C66@cs.rutgers.edu>
Cc: "linux-mm@kvack.org" <linux-mm@kvack.org>,
        "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>,
        Hugh Dickins <hughd@google.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Dave Hansen <dave.hansen@intel.com>,
        Andrea Arcangeli <aarcange@redhat.com>,
        Mel Gorman <mgorman@techsingularity.net>,
        Michal Hocko <mhocko@kernel.org>, Vlastimil Babka <vbabka@suse.cz>,
        Pavel Emelyanov <xemul@parallels.com>,
        Balbir Singh <bsingharora@gmail.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Naoya Horiguchi <nao.horiguchi@gmail.com>
From: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Date: Fri, 11 Nov 2016 09:18:44 +0530
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101
 Thunderbird/38.1.0
MIME-Version: 1.0
In-Reply-To: <D34FA575-7C5D-4E9D-B337-A925F1A89C66@cs.rutgers.edu>
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 8bit
Message-Id: <58253F9C.6040307@linux.vnet.ibm.com>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2732
Lines: 66

On 11/10/2016 07:31 PM, Zi Yan wrote:
> On 9 Nov 2016, at 18:52, Naoya Horiguchi wrote:
> 
>> Hi Anshuman,
>>
>> On Wed, Nov 09, 2016 at 04:03:04PM +0530, Anshuman Khandual wrote:
>>> On 11/08/2016 05:01 AM, Naoya Horiguchi wrote:
>>>> Hi everyone,
>>>>
>>>> I've updated thp migration patches for v4.9-rc2-mmotm-2016-10-27-18-27
>>>> with feedbacks for ver.1.
>>>>
>>>> General description (no change since ver.1)
>>>> ===========================================
>>>>
>>>> This patchset enhances page migration functionality to handle thp migration
>>>> for various page migration's callers:
>>>>  - mbind(2)
>>>>  - move_pages(2)
>>>>  - migrate_pages(2)
>>>>  - cgroup/cpuset migration
>>>>  - memory hotremove
>>>>  - soft offline
>>>>
>>>> The main benefit is that we can avoid unnecessary thp splits, which helps us
>>>> avoid performance decrease when your applications handles NUMA optimization on
>>>> their own.
>>>>
>>>> The implementation is similar to that of normal page migration, the key point
>>>> is that we modify a pmd to a pmd migration entry in swap-entry like format.
>>>
>>> Will it be better to have new THP_MIGRATE_SUCCESS and THP_MIGRATE_FAIL
>>> VM events to capture how many times the migration worked without first
>>> splitting the huge page and how many time it did not work ?
>>
>> Thank you for the suggestion.
>> I think that's helpful, so will try it in next version.
>>
>>> Also do you
>>> have a test case which demonstrates this THP migration and kind of shows
>>> its better than the present split and move method ?
>>
>> I don't have test cases which compare thp migration and split-then-migration
>> with some numbers. Maybe measuring/comparing the overhead of migration is
>> a good start point, although I think the real benefit of thp migration comes
>> from workload "after migration" by avoiding thp split.
> 
> Migrating 4KB pages has much lower (~1/3) throughput than 2MB pages.

I assume the 2MB throughput you mentioned is with this THP migration
feature enabled.

> 
> What I get is that on average it takes 1987.38 us to migrate 512 4KB pages and
>                                        658.54  us to migrate 1   2MB page.
> 
> I did the test in a two-socket Intel Xeon E5-2640v4 box. I used migrate_pages()
> system call to migrate pages. MADV_NOHUGEPAGE and MADV_HUGEPAGE are used to
> make 4KB and 2MB pages and each page’s flags are checked to make sure the page
> size is 4KB or 2MB THP.
> 
> There is no split page. But the page migration time already tells the story.

Right. Just wondering if we can add a test case which measures just
this migration time improvement by avoiding the split not the TLB
based improvement which the workload will receive as an addition.