[Linux Memory Hotness and Promotion] Notes from February 26, 2026

[Linux Memory Hotness and Promotion] Notes from February 26, 2026

[David Rientjes]

Hi everybody,

Here are the notes from the last Linux Memory Hotness and Promotion call
that happened on Thursday, February 26.  Thanks to everybody who was 
involved!

These notes are intended to bring people up to speed who could not attend 
the call as well as keep the conversation going in between meetings.

----->o-----
Bharata provided an update on the status of his work.  He posted 
redis-memtier, graph500, and NPB-BT numbers for the RFC v5 patch series.  
He's planning posting v6 soon with a new set of numbers and in that series 
base pghot and hintfault source will be non-RFC.  It may even be possible 
to include other sources of hotness information at the same time.

Joshua Hahn noted that he was looking to test Bharata's series with his 
reclaim fairness patch series, now rebranded as tier-aware memcg limits.  
Gregory noted that this series would be focused only on top tier memory 
but extensible for future use cases.

----->o-----
We touched on the CXL Hotness Monitoring Unit (CHMU) and whether any work 
was on-going to abstract this in upstream Linux.  Both Google and Meta 
were not actively looking at this.  Yiannis suggested Jonathan Cameron may 
be looking at this for qemu and testing.  We concluded that not having 
CHMU support upstream is not currently holding anything back and it might 
be addressed in a year or so; it might also be solving a problem that 
nobody has yet.

----->o-----
We talked about LSF/MM/BPF topics that relate to this series of meetings.  
Joshua had proposed a topic for tier aware memcg limits looking for 
feedback on the overall approach and interface.  Joshua wanted to raise 
discussion points around actual multi-tiered system usage in production 
scenarios including three or four tiers of memory.  I chatted about 
potential future use cases but also focusing on the practicality of what 
is possible to deploy today.  Gregory suggested we need a better 
definition for what a coherent tier is given that NUMA nodes themselves 
are likely insufficient.  Managing nodes in userspace is likely not the 
end user support that will be required long term.

Gregory suggested that until we have three tier systems to use and deploy 
that we won't likely be able to foresee the problems with that until we 
get there.

----->o-----
Yiannis brought up non-temporal stores for tiering and the possibility of 
finding time to work on it in the next few weeks.  He saw great value in 
this from the compression side and was trying to determine if this brings 
value to CXL or tiered systems in general.  The idea is to extend 
migrate_pages() for the demotion path so it uses non-temporal stores -- we 
don't want to warm up our cache for cold memory.  Gregory noted that we 
want to ensure that when allocating the folio as the migration target that 
we would also need to make sure that's not in the cache for this cold 
memory.  I noted that Shivank from AMD had previously presented to this 
group about enlightening migrate_pages() for hardware assists and using 
the "reason" field of migrate_pages() to differentiate different use 
cases.

----->o-----
NOTE!!  Daylight Savings Time starts on Sunday, March 8 so the time of the 
meeting may have changed for some participants.

Next meeting will be on Thursday, March 12 at 8:30am PDT (UTC-7),
everybody is welcome: https://meet.google.com/jak-ytdx-hnm

Topics for the next meeting:

 - v6 of Bharata's patch series including pghot and hintfault
 - LSF/MM/BPF 2026 topics to propose for discussion on hotness, promotion,
   and memory tiering overall
 - tier-aware memcg limits (reclaim fairness)
 - Gregory's testing of tier aware memcg limits with Bharata's changes
 - discuss generalized subsystem for providing bandwidth information
   independent of the underlying platform, ideally through resctrl,
   otherwise utilizing bandwidth information will be challenging
   + preferably this bandwidth monitoring is not per NUMA node but rather
     slow and fast
 - update on non-temporal stores enlightenment for memory tiering
 - enlightening migrate_pages() for hardware assists and how this work
   will be charged to userspace, including for memory compaction

Please let me know if you'd like to propose additional topics for
discussion, thank you!