Re: [PATCH -next] mm/hugetlb_cgroup: introduce peak and rsvd.peak to v2

[xiujianfeng <xiujianfeng@huawei.com>]

On 2024/7/9 0:04, Michal Hocko wrote:
> On Mon 08-07-24 21:40:39, xiujianfeng wrote:
>>
>>
>> On 2024/7/8 20:48, Michal Hocko wrote:
>>> On Wed 03-07-24 13:38:04, Andrew Morton wrote:
>>>> On Wed, 3 Jul 2024 10:45:56 +0800 xiujianfeng <xiujianfeng@huawei.com> wrote:
>>>>
>>>>>
>>>>>
>>>>> On 2024/7/3 9:58, Andrew Morton wrote:
>>>>>> On Tue, 2 Jul 2024 12:57:28 +0000 Xiu Jianfeng <xiujianfeng@huawei.com> wrote:
>>>>>>
>>>>>>> Introduce peak and rsvd.peak to v2 to show the historical maximum
>>>>>>> usage of resources, as in some scenarios it is necessary to configure
>>>>>>> the value of max/rsvd.max based on the peak usage of resources.
>>>>>>
>>>>>> "in some scenarios it is necessary" is not a strong statement.  It
>>>>>> would be helpful to fully describe these scenarios so that others can
>>>>>> better understand the value of this change.
>>>>>>
>>>>>
>>>>> Hi Andrew,
>>>>>
>>>>> Is the following description acceptable for you?
>>>>>
>>>>>
>>>>> Since HugeTLB doesn't support page reclaim, enforcing the limit at
>>>>> page fault time implies that, the application will get SIGBUS signal
>>>>> if it tries to fault in HugeTLB pages beyond its limit. Therefore the
>>>>> application needs to know exactly how many HugeTLB pages it uses before
>>>>> hand, and the sysadmin needs to make sure that there are enough
>>>>> available on the machine for all the users to avoid processes getting
>>>>> SIGBUS.
>>>
>>> yes, this is pretty much a definition of hugetlb.
>>>
>>>>> When running some open-source software, it may not be possible to know
>>>>> the exact amount of hugetlb it consumes, so cannot correctly configure
>>>>> the max value. If there is a peak metric, we can run the open-source
>>>>> software first and then configure the max based on the peak value.
>>>
>>> I would push back on this. Hugetlb workloads pretty much require to know
>>> the number of hugetlb pages ahead of time. Because you need to
>>> preallocate them for the global hugetlb pool. What I am really missing
>>> in the above justification is an explanation of how come you know how to
>>> configure the global pool but you do not know that for a particular
>>> cgroup. How exactly do you configure the global pool then?
>>
>> Yes, in this scenario, it's indeed challenging to determine the
>> appropriate size for the global pool. Therefore, a feasible approach is
>> to initially configure a larger value. Once the software is running
>> within the container successfully, the maximum value for the container
>> and the size of the system's global pool can be determined based on the
>> peak value, otherwise, increase the size of the global pool and try
>> again. so I believe the peak metric is useful for this scenario.
> 
> This sounds really backwards to me. Not that I care much about peak
> value itself. It is not really anything disruptive to add nor maintain
> but this approach to configuring the system just feels completely wrong.
> You shouldn't be really using hugetlb cgroup controller if you do not
> have a very specific idea about expected and therefore allowed hugetlb
> pool consumption.
> 

Thanks for sharing your thoughts.

Since the peak metric exists in the legacy hugetlb controller, do you
have any idea what scenario it's used for? I found it was introduced by
commit abb8206cb077 ("hugetlb/cgroup: add hugetlb cgroup control
files"), however there is no any description about the scenario.