NUMA + THP test suites?

NUMA + THP test suites?

[John Hubbard]

Hi,

I have some tentative solutions to the "NUMA balancing prerequisite" for
Ryan's "[PATCH v6 0/9] variable-order, large folios for anonymous
memory" [1]. However, I need to actually exercise the patches, in order
to post anything worth reading, and as usual that is the hard part.

After some initial searching around, it's time to simply ask: are there
any test suites that will exercise NUMA page faults with THP? I am
seeing hints of coverage, but with NUMA and THP separately covered, in
the mm selftests. And also stress-ng has some numa coverage.

So I was thinking of augmenting the selftests to get coverage (and the
pte-mapped THP would be a happier story with this as well), but maybe
tests along these lines already exist somewhere?


[1] https://lore.kernel.org/all/197819d2-d7fa-47ae-b893-ef2fce25b2be@nvidia.com/

thanks,
-- 
John Hubbard
NVIDIA

Re: NUMA + THP test suites?

[David Hildenbrand]

On 31.10.23 00:21, John Hubbard wrote:
> Hi,
> 
> I have some tentative solutions to the "NUMA balancing prerequisite" for
> Ryan's "[PATCH v6 0/9] variable-order, large folios for anonymous
> memory" [1]. However, I need to actually exercise the patches, in order
> to post anything worth reading, and as usual that is the hard part.
> 
> After some initial searching around, it's time to simply ask: are there
> any test suites that will exercise NUMA page faults with THP? I am
> seeing hints of coverage, but with NUMA and THP separately covered, in
> the mm selftests. And also stress-ng has some numa coverage.
> 
> So I was thinking of augmenting the selftests to get coverage (and the
> pte-mapped THP would be a happier story with this as well), but maybe
> tests along these lines already exist somewhere?

I suspect that you can run the autonuma-benchmark [1] (not sure if the 
official repository) with THP enabled.

[1] https://github.com/pholasek/autonuma-benchmark

-- 
Cheers,

David / dhildenb

Re: NUMA + THP test suites?

[John Hubbard]

On 10/31/23 03:28, David Hildenbrand wrote:
> On 31.10.23 00:21, John Hubbard wrote:
>> Hi,
>>
>> I have some tentative solutions to the "NUMA balancing prerequisite" for
>> Ryan's "[PATCH v6 0/9] variable-order, large folios for anonymous
>> memory" [1]. However, I need to actually exercise the patches, in order
>> to post anything worth reading, and as usual that is the hard part.
>>
>> After some initial searching around, it's time to simply ask: are there
>> any test suites that will exercise NUMA page faults with THP? I am
>> seeing hints of coverage, but with NUMA and THP separately covered, in
>> the mm selftests. And also stress-ng has some numa coverage.
>>
>> So I was thinking of augmenting the selftests to get coverage (and the
>> pte-mapped THP would be a happier story with this as well), but maybe
>> tests along these lines already exist somewhere?
> 
> I suspect that you can run the autonuma-benchmark [1] (not sure if the 
> official repository) with THP enabled.
> 
> [1] https://github.com/pholasek/autonuma-benchmark
> 

Good idea, and thanks for the quick response too.


thanks,
-- 
John Hubbard
NVIDIA