[RFC PLAN] Some humble ideas for DAMON future works

[SeongJae Park]

Hello,


After the DAMON code is merged (kudos to the community for the great helps), a
few people asked me about my plan for DAMON future works, and if DAMON will be
somewhat usable for their use cases.  I indeed have some humble plans, though
those are only in rough brainsorming level at the moment. so I'd like to share
those here before going forward and start coding, so that I can get some
feedback to fail fast.

User-space Policy or In-kernel Policy?  Both.
=============================================

When discussing about a sort of kernel involved system efficiency
optimizations, I show two kinds of people who have slightly different opinions.
The first party prefer to implement only simple but efficient mechanisms in the
kernel and export it to user space, so that users can make smart user space
policy.  Meanwhile, the second party prefer the kernel just works.  I agree
with both parties.

I think the first opinion makes sense as there are some valuable information
that only user space can know.  I think only such approaches could achieve the
ultimate efficiency in such cases.
I also agree to the second party, though, because there could be some people
who don't have special information that only their applications know, or
resources to do the additional work.  In-kernel simple policies will be still
beneficial for some users even though those are sub-optimal compared to the
highly tuned user space policy, if it provides some extent of efficiency gain
and no regressions for most cases.

I'd like to help both.  For the reason, I made DAMON as an in-kernel mechanism
for both user and kernel-space policies.  It provides highly tunable general
user space interface to help the first party.  It also provides in-kernel
policies which built on top of DAMON using its kernel-space API for specific
common use cases with conservative default parameters that assumed to incur no
regression but some extent of benefits in most cases, namely DAMON-based
proactive reclamation.  I will continue pursuing the two ways.

Imaginable DAMON-based Policies
===============================

I'd like to start from listing some imaginable data access-aware operation
policies that I hope to eventually be made.  The list will hopefully shed light
on how DAMON should be evolved to efficiently support the policies.

DAMON-based Proactive LRU-pages (de)Activation
----------------------------------------------

The reclamation mechanism which selects reclaim target using the
active/inactive LRU lists sometimes doesn't work well.  According to my
previous work, providing access pattern-based hints can significantly improve
the performance under memory pressure[1,2].

Proactive reclamation is known to be useful for many memory intensive systems,
and now we have a DAMON-based implementation of it[3].  However, the proactive
reclamation wouldn't be so welcome to some systems having high cost of I/O.
Also, even though the system runs proactive reclamation, memory pressure can
still occasionally triggered.

My idea for helping this situation is manipulating the orders of pages in LRU
lists using DAMON-provided monitoring results.  That is, making DAMON
proactively finds hot/cold memory regions and moves pages of the hot regions to
the head of the active list, while moving pages of the cold regions to the tail
of the inactive list.  This will help eventual reclamation under memory
pressure to evict cold pages first, so incur less additional page faults.

[1] https://www.usenix.org/conference/hotstorage19/presentation/park
[2] https://linuxplumbersconf.org/event/4/contributions/548/
[3] https://docs.kernel.org/admin-guide/mm/damon/reclaim.html

DAMON-based THP Coalesce/Split
------------------------------

THP is know to significantly improve performance, but also increase memory
footprint[1].  We can minimize the memory overhead while preserving the
performance benefit by asking DAMON to provide MADV_HUGEPAGE-like hints for hot
memory regions of >= 2MiB size, and MADV_NOHUGEPAGE-like hints for cold memory
regions.  Our experimental user space policy implementation[2] of this idea
removes 76.15% of THP memory waste while preserving 51.25% of THP speedup in
total.

[1] https://www.usenix.org/conference/osdi16/technical-sessions/presentation/kwon
[2] https://damonitor.github.io/doc/html/v34/vm/damon/eval.html

DAMON-based Tiered Memory (Pro|De)motion
----------------------------------------

In tiered memory systems utilizing DRAM and PMEM[1], we can promote hot pages to
DRAM and demote cold pages to PMEM using DAMON.  A patch for allowing
access-aware demotion user space policy development is already submitted[2] by
Baolin.

[1] https://www.intel.com/content/www/us/en/products/details/memory-storage/optane-memory.html
[2] https://lore.kernel.org/linux-mm/cover.1640171137.git.baolin.wang@xxxxxxxxxxxxxxxxx/

DAMON-based Proactive Compaction
--------------------------------

Compaction uses migration scanner to find migration source pages.  Hot pages
would be more likely to be unmovable compared to cold pages, so it would be
better to try migration of cold pages first.  DAMON could be used here.  That
is, proactively monitoring accesses via DAMON and start compaction so that the
migration scanner scan cold memory ranges first.  I should admit I'm not
familiar with compaction code and I have no PoC data for this but just the
groundless idea, though.

How We Can Implement These
--------------------------

Implementing most of the above mentioned policies wouldn't be too difficult
because we have DAMON-based Operation Schemes (DAMOS).  That is, we will need
to implement some more DAMOS action for each policy.  Some existing kernel
functions can be reused.  Such actions would include LRU (de)activation, THP
coalesce/split hints, memory (pro|de)motion, and cold pages first scanning
compaction.  Then, supporting those actions with the user space interface will
allows implementing user space policies.  If we find reasonably good default
DAMOS parameters and some kernel side control mechanism, we can further make
those as kernel policies in form of, say, builtin modules.

How DAMON Should Be Evolved For Supporting Those
================================================

Let's discuss what kind of changes in DAMON will be needed to efficiently
support above mentioned policies.

Simultaneously Monitoring Different Types of Address Spaces
-----------------------------------------------------------

It would be better to run all the above mentioned policies simultaneously on
single system.  As some policies such as LRU-pages (de)activation would better
to run on physical address space while some policies such as THP coalesce/split
would need to run on virtual address spaces, DAMON should support concurrently
monitoring different address spaces.  We can always do this by creating one
DAMON context for each address space and running those.  However, as the
address spaces will conflict, each other will be interfered.  Current idea for
avoiding this is allowing multiple DAMON contexts to run on a single thread,
forcing them to have same monitoring contexts.

Online Parameters Updates
-------------------------

Someone would also want to dynamically turn on/off and/or tune each policy.
This is impossible with current DAMON, because it prohibits updating any
parameter while it is running.  We disallow the online parameters update
mainly because we want to avoid doing additional synchronization between the
running kdamond and the parameters updater.  The idea for supporting the use
case while avoiding the additional synchronization is, allowing users to pause
DAMON and update parameters while it is paused.

A Better DAMON interface
------------------------

DAMON is currently exposing its major functionality to the user space via the
debugfs.  After all, DAMON is not for only debugging.  Also, this makes the
interface depends on debugfs unnecessarily, and considered unreliable.  Also,
the interface is quite unflexible for future interface extension.  I admit it
was not a good choice.

It would be better to implement another reliable and easily extensible
interface, and deprecate the debugfs interface.  The idea is exposing the
interface via sysfs using hierarchical Kobjects under mm_kobject.  For example,
the usage would be something like below:

    # cd /sys/kernel/mm/damon
    # echo 1 > nr_kdamonds
    # echo 1 > kdamond_1/contexts/nr_contexts
    # echo va > kdamond_1/contexts/context_1/target_type
    # echo 1 > kdamond_1/contexts/context_1/targets/nr_targets
    # echo $(pidof <workload>) > \
                    kdamond_1/contexts/context_1/targets/target_1/pid
    # echo Y > monitor_on

The underlying files hierarchy could be something like below.

    /sys/kernel/mm/damon/
    │ monitor_on
    │ kdamonds
    │ │ nr_kdamonds
    │ │ kdamond_1/
    │ │ │ kdamond_pid
    │ │ │ contexts
    │ │ │ │ nr_contexts
    │ │ │ │ context_1/
    │ │ │ │ │ target_type (va | pa)
    │ │ │ │ │ attrs/
    │ │ │ │ │ │ intervals/sampling,aggr,update
    │ │ │ │ │ │ nr_regions/min,max
    │ │ │ │ │ targets/
    │ │ │ │ │ │ nr_targets
    │ │ │ │ │ │ target_1/
    │ │ │ │ │ │ │ pid
    │ │ │ │ │ │ │ init_regions/
    │ │ │ │ │ │ │ │ region1/
    │ │ │ │ │ │ │ │ │ start,end
    │ │ │ │ │ │ │ │ ...
    │ │ │ │ │ │ ...
    │ │ │ │ │ schemes/
    │ │ │ │ │ │ nr_schemes
    │ │ │ │ │ │ scheme_1/
    │ │ │ │ │ │ │ action
    │ │ │ │ │ │ │ target_access_pattern/
    │ │ │ │ │ │ │ │ sz/min,max
    │ │ │ │ │ │ │ │ nr_accesses/min,max
    │ │ │ │ │ │ │ │ age/min,max
    │ │ │ │ │ │ │ quotas/
    │ │ │ │ │ │ │ │ ms,bytes,reset_interval
    │ │ │ │ │ │ │ │ prioritization_weights/
    │ │ │ │ │ │ │ │   sz,nr_accesses,age
    │ │ │ │ │ │ │ watermarks/
    │ │ │ │ │ │ │   metric,check_interval,high,mid,low
    │ │ │ │ │ │ │ stats/
    │ │ │ │ │ │ │ │ quota_exceeds
    │ │ │ │ │ │ │ │ tried/nr,sz
    │ │ │ │ │ │ │ │ applied/nr,sz
    │ │ │ │ │ │ │ ...
    │ │ │ │ ...
    │ │ ...

More DAMON Future Works
=======================

In addition to above mentioned things, there are many works to do.  It would be
better to extend DAMON for more use cases and address spaces support, including
page granularity, idleness only, read/write only, page cache only, and cgroups
monitoring supports.

Also it would be valuable to improve the accuracy of monitoring, using some
adaptive monitoring attributes tuning or some new fancy idea[1].

DAMOS could also be improved by utilizing its own autotuning feature, for
example, by monitoring PSI and other metrics related to the given action.

[1] https://linuxplumbersconf.org/event/11/contributions/984/


Thank you For Reading This
==========================

So, I shared current rough and immature plans off the top of my head here.
Hope this helps you understanding what I'm thinking about for the future of
DAMON.  Please note again that those are only in brainstorming level and some
are only groundless idea.  Some might be just insane ideas.  Hence, everything
is open for change or failure.  If you have any comment, please feel free to
let me know.


Thanks,
SJ