[PATCH v7 1/5] drivers/base/node: Optimize memory block registration to reduce boot time

[David Hildenbrand]

On 04.06.25 15:17, Donet Tom wrote:
> 
> On 6/4/25 3:15 PM, David Hildenbrand wrote:
>> On 04.06.25 05:07, Andrew Morton wrote:
>>> On Wed, 28 May 2025 12:18:00 -0500 Donet Tom <donettom at linux.ibm.com>
>>> wrote:
>>>
>>>> During node device initialization, `memory blocks` are registered under
>>>> each NUMA node. The `memory blocks` to be registered are identified
>>>> using
>>>> the node’s start and end PFNs, which are obtained from the node's
>>>> pg_data
>>>
>>> It's quite unconventional to omit the [0/N] changelog.  This omission
>>> somewhat messed up my processes so I added a one-liner to this.
>>>
>>
>> Yeah, I was assuming that I simply did not get cc'ed on the cover
>> letter, but there is actually none.
>>
>> Donet please add that in the future. git can do this using
>> --cover-letter.
> 
> Sure,
> 
> I will add cover letter in next revision.
> 
> 
>>
>>>>
>>>> ...
>>>>
>>>> Test Results on My system with 32TB RAM
>>>> =======================================
>>>> 1. Boot time with CONFIG_DEFERRED_STRUCT_PAGE_INIT enabled.
>>>>
>>>> Without this patch
>>>> ------------------
>>>> Startup finished in 1min 16.528s (kernel)
>>>>
>>>> With this patch
>>>> ---------------
>>>> Startup finished in 17.236s (kernel) - 78% Improvement
>>>
>>> Well someone is in for a nice surprise.
>>>
>>>> 2. Boot time with CONFIG_DEFERRED_STRUCT_PAGE_INIT disabled.
>>>>
>>>> Without this patch
>>>> ------------------
>>>> Startup finished in 28.320s (kernel)
>>>
>>> what.  CONFIG_DEFERRED_STRUCT_PAGE_INIT is supposed to make bootup
>>> faster.
>>
>> Right, that's weird. Especially that it is still slower after these
>> changes.
>>
>> CONFIG_DEFERRED_STRUCT_PAGE_INIT should be initializing in parallel
>> which ... should be faster.
>>
>> @Donet, how many CPUs and nodes does your system have? Can you
>> identify what is taking longer than without
>> CONFIG_DEFERRED_STRUCT_PAGE_INIT?
> 
> 
> 
> My system has,
> 
> CPU      - 1528

Holy cow.

Pure speculation: are we parallelizing *too much* ? :)

That's ~95 CPUs per node on average.

Staring at deferred_init_memmap(), we do have

	max_threads = deferred_page_init_max_threads(cpumask);

And that calls cpumask_weight(), essentially using all CPUs on the node.

... not sure what exactly happens if there are no CPUs for a node.

> Node     - 16

Are any of these memory-less?

> Memory - 31TB



-- 
Cheers,

David / dhildenb