[LSF/MM/BPF TOPIC] Data Access-aware Transparent Hugepages

[LSF/MM/BPF TOPIC] Data Access-aware Transparent Hugepages

[SeongJae Park]

Hello,


Transparent Huge Pages (THP) could be more efficient if those are used only for
frequently accessed data.  DAMON is one of the ways for fidning such hot data,
and hence there were interests in utilizing DAMON for access-aware THP.  A
simple experiment leveraging MADV_[NO]HUGEPAGE based on DAMON-identifed hot and
cold memory regions was conducted as a proof of the concept.  The experiment
demonstrated [1] an 80% reduction in memory overhead while retaining 46% of THP
performance gains.

Multiple discussions for further developing production level access-aware THP
solutions have followed.  Most of the discussions were made in private, though
Asier recently proposed yet another approach [2] on the mailing list.  As a
result, we collected interesting challenges and ideas for the production level
data access-aware THP.  Some of the ideas are not yet public, could be
controversial, and missing important things that more eyeballs can find.  I
want to discuss those in LSF/MM/BPF to get more ideas and make high level
community alignments for next steps.

[1] https://assets.amazon.science/b7/2b/ce53222247739b174f2b54498d1a/daos-data-access-aware-operating-system.pdf
[2] https://lore.kernel.org/20260202145650.1795854-1-gutierrez.asier@huawei-partners.com


Thanks,
SJ

Re: [LSF/MM/BPF TOPIC] Data Access-aware Transparent Hugepages

[SeongJae Park]

Forgot Cc-ing damon@ and linux-kernel@.  Adding those.  Sorry for noise.

On Tue, 10 Feb 2026 18:28:55 SeongJae Park <sj@kernel.org> wrote:

> Hello,
> 
> 
> Transparent Huge Pages (THP) could be more efficient if those are used only for
> frequently accessed data.  DAMON is one of the ways for fidning such hot data,
> and hence there were interests in utilizing DAMON for access-aware THP.  A
> simple experiment leveraging MADV_[NO]HUGEPAGE based on DAMON-identifed hot and
> cold memory regions was conducted as a proof of the concept.  The experiment
> demonstrated [1] an 80% reduction in memory overhead while retaining 46% of THP
> performance gains.
> 
> Multiple discussions for further developing production level access-aware THP
> solutions have followed.  Most of the discussions were made in private, though
> Asier recently proposed yet another approach [2] on the mailing list.  As a
> result, we collected interesting challenges and ideas for the production level
> data access-aware THP.  Some of the ideas are not yet public, could be
> controversial, and missing important things that more eyeballs can find.  I
> want to discuss those in LSF/MM/BPF to get more ideas and make high level
> community alignments for next steps.
> 
> [1] https://assets.amazon.science/b7/2b/ce53222247739b174f2b54498d1a/daos-data-access-aware-operating-system.pdf
> [2] https://lore.kernel.org/20260202145650.1795854-1-gutierrez.asier@huawei-partners.com
> 
> 
> Thanks,
> SJ
> 

Sent using hkml (https://github.com/sjp38/hackermail)

Re: [LSF/MM/BPF TOPIC] Data Access-aware Transparent Hugepages

[SeongJae Park]

On Tue, 10 Feb 2026 20:45:33 -0800 SeongJae Park <sj@kernel.org> wrote:

> Forgot Cc-ing damon@ and linux-kernel@.  Adding those.  Sorry for noise.

And lsf-pc@ ...

> 
> On Tue, 10 Feb 2026 18:28:55 SeongJae Park <sj@kernel.org> wrote:
> 
> > Hello,
> > 
> > 
> > Transparent Huge Pages (THP) could be more efficient if those are used only for
> > frequently accessed data.  DAMON is one of the ways for fidning such hot data,
> > and hence there were interests in utilizing DAMON for access-aware THP.  A
> > simple experiment leveraging MADV_[NO]HUGEPAGE based on DAMON-identifed hot and
> > cold memory regions was conducted as a proof of the concept.  The experiment
> > demonstrated [1] an 80% reduction in memory overhead while retaining 46% of THP
> > performance gains.
> > 
> > Multiple discussions for further developing production level access-aware THP
> > solutions have followed.  Most of the discussions were made in private, though
> > Asier recently proposed yet another approach [2] on the mailing list.  As a
> > result, we collected interesting challenges and ideas for the production level
> > data access-aware THP.  Some of the ideas are not yet public, could be
> > controversial, and missing important things that more eyeballs can find.  I
> > want to discuss those in LSF/MM/BPF to get more ideas and make high level
> > community alignments for next steps.
> > 
> > [1] https://assets.amazon.science/b7/2b/ce53222247739b174f2b54498d1a/daos-data-access-aware-operating-system.pdf
> > [2] https://lore.kernel.org/20260202145650.1795854-1-gutierrez.asier@huawei-partners.com
> > 
> > 
> > Thanks,
> > SJ
> > 
> 
> Sent using hkml (https://github.com/sjp38/hackermail)

Sent using hkml (https://github.com/sjp38/hackermail)