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: <eaaab6c4-1719-5e21-ebf7-a0c17487314c@redhat.com>
Date:   Thu, 4 Aug 2022 10:12:00 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.11.0
Subject: Re: [RFC PATCH V4 1/1] mm: add last level page table numa info to
 /proc/pid/numa_pgtable
Content-Language: en-US
To:     haoxin <xhao@linux.alibaba.com>, willy@infradead.org
Cc:     akpm@linux-foundation.org, adobriyan@gmail.com,
        keescook@chromium.org, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org
References: <20220801121727.76186-1-xhao@linux.alibaba.com>
 <20220801121727.76186-2-xhao@linux.alibaba.com>
 <0c1f9e76-9b1d-7069-bb09-c18e4f19f0c4@redhat.com>
 <bc0d16a6-e340-e261-82a0-e17bd236c2d9@linux.alibaba.com>
From:   David Hildenbrand <david@redhat.com>
Organization: Red Hat
In-Reply-To: <bc0d16a6-e340-e261-82a0-e17bd236c2d9@linux.alibaba.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 04.08.22 10:04, haoxin wrote:
> 
> 在 2022/8/1 下午9:28, David Hildenbrand 写道:
>> On 01.08.22 14:17, Xin Hao wrote:
>>> In many data center servers, the shared memory architectures is
>>> Non-Uniform Memory Access (NUMA), remote numa node data access
>>> often brings a high latency problem, but what we are easy to ignore
>>> is that the page table remote numa access, It can also leads to a
>>> performance degradation.
>> Let me try rewriting:
>>
>> "
>> Many data center servers employ Non-Uniform Memory Access (NUMA)
>> architectures. Remote numa memory access results in high latency. While
>> memory placement is one issue, sub-optimal page table placement can also
>> result in surprise performance degradation.
>> "
> Thanks,  it reads more clearly.
> 
>>> So there add a new interface in /proc, This will help developers to
>>> get more info about performance issues if they are caused by cross-NUMA.
>>
>> Why do we only care about "last level page table", why not about the others?
>>
>> IMHO, we could emit something like "0, 1, 3, 0" instead for a given user
>> space address, showing the NUMA node the page table belongs to from
>> highest to lowest page table level.
> 
> I have planned to implement the PTE page table in this version first,  
> and then support other page tables in the next patch later.

If there are plans, let's do it all at once, to get a good and single
interface to expose that information.

-- 
Thanks,

David / dhildenb