mm: introduce Data Access MONitor (DAMON)

mm: introduce Data Access MONitor (DAMON)

Patch series "Introduce Data Access MONitor (DAMON)", v34.

Introduction
============

DAMON is a data access monitoring framework for the Linux kernel.  The
core mechanisms of DAMON called 'region based sampling' and 'adaptive
regions adjustment' (refer to 'mechanisms.rst' in the 11th patch of this
patchset for the detail) make it

- accurate (The monitored information is useful for DRAM level memory
  management.  It might not appropriate for Cache-level accuracy,
  though.),

- light-weight (The monitoring overhead is low enough to be applied
  online while making no impact on the performance of the target
  workloads.), and

- scalable (the upper-bound of the instrumentation overhead is
  controllable regardless of the size of target workloads.).

Using this framework, therefore, several memory management mechanisms such
as reclamation and THP can be optimized to aware real data access
patterns.  Experimental access pattern aware memory management
optimization works that incurring high instrumentation overhead will be
able to have another try.

Though DAMON is for kernel subsystems, it can be easily exposed to the
user space by writing a DAMON-wrapper kernel subsystem.  Then, user space
users who have some special workloads will be able to write personalized
tools or applications for deeper understanding and specialized
optimizations of their systems.

DAMON is also merged in two public Amazon Linux kernel trees that based on
v5.4.y[1] and v5.10.y[2].

[1] https://github.com/amazonlinux/linux/tree/amazon-5.4.y/master/mm/damon
[2] https://github.com/amazonlinux/linux/tree/amazon-5.10.y/master/mm/damon

The userspace tool[1] is available, released under GPLv2, and actively
being maintained.  I am also planning to implement another basic user
interface in perf[2].  Also, the basic test suite for DAMON is available
under GPLv2[3].

[1] https://github.com/awslabs/damo
[2] https://lore.kernel.org/linux-mm/20210107120729.22328-1-sjpark@amazon.com/
[3] https://github.com/awslabs/damon-tests

Long-term Plan
--------------

DAMON is a part of a project called Data Access-aware Operating System
(DAOS).  As the name implies, I want to improve the performance and
efficiency of systems using fine-grained data access patterns.  The
optimizations are for both kernel and user spaces.  I will therefore
modify or create kernel subsystems, export some of those to user space and
implement user space library / tools.  Below shows the layers and
components for the project.

    ---------------------------------------------------------------------------
    Primitives:     PTE Accessed bit, PG_idle, rmap, (Intel CMT), ...
    Framework:      DAMON
    Features:       DAMOS, virtual addr, physical addr, ...
    Applications:   DAMON-debugfs, (DARC), ...
    ^^^^^^^^^^^^^^^^^^^^^^^    KERNEL SPACE    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

    Raw Interface:  debugfs, (sysfs), (damonfs), tracepoints, (sys_damon), ...

    vvvvvvvvvvvvvvvvvvvvvvv    USER SPACE      vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
    Library:        (libdamon), ...
    Tools:          DAMO, (perf), ...
    ---------------------------------------------------------------------------

The components in parentheses or marked as '...' are not implemented yet
but in the future plan.  IOW, those are the TODO tasks of DAOS project.
For more detail, please refer to the plans:
https://lore.kernel.org/linux-mm/20201202082731.24828-1-sjpark@amazon.com/

Evaluations
===========

We evaluated DAMON's overhead, monitoring quality and usefulness using 24
realistic workloads on my QEMU/KVM based virtual machine running a kernel
that v24 DAMON patchset is applied.

DAMON is lightweight.  It increases system memory usage by 0.39% and slows
target workloads down by 1.16%.

DAMON is accurate and useful for memory management optimizations.  An
experimental DAMON-based operation scheme for THP, namely 'ethp', removes
76.15% of THP memory overheads while preserving 51.25% of THP speedup.
Another experimental DAMON-based 'proactive reclamation' implementation,
'prcl', reduces 93.38% of residential sets and 23.63% of system memory
footprint while incurring only 1.22% runtime overhead in the best case
(parsec3/freqmine).

NOTE that the experimental THP optimization and proactive reclamation are
not for production but only for proof of concepts.

Please refer to the official document[1] or "Documentation/admin-guide/mm:
Add a document for DAMON" patch in this patchset for detailed evaluation
setup and results.

[1] https://damonitor.github.io/doc/html/latest-damon/admin-guide/mm/damon/eval.html

Real-world User Story
=====================

In summary, DAMON has used on production systems and proved its usefulness.

DAMON as a profiler
-------------------

We analyzed characteristics of a large scale production systems of our
customers using DAMON.  The systems utilize 70GB DRAM and 36 CPUs.  From
this, we were able to find interesting things below.

There were obviously different access pattern under idle workload and
active workload.  Under the idle workload, it accessed large memory
regions with low frequency, while the active workload accessed small
memory regions with high freuqnecy.

DAMON found a 7GB memory region that showing obviously high access
frequency under the active workload.  We believe this is the
performance-effective working set and need to be protected.

There was a 4KB memory region that showing highest access frequency under
not only active but also idle workloads.  We think this must be a hottest
code section like thing that should never be paged out.

For this analysis, DAMON used only 0.3-1% of single CPU time.  Because we
used recording-based analysis, it consumed about 3-12 MB of disk space per
20 minutes.  This is only small amount of disk space, but we can further
reduce the disk usage by using non-recording-based DAMON features.  I'd
like to argue that only DAMON can do such detailed analysis (finding 4KB
highest region in 70GB memory) with the light overhead.

DAMON as a system optimization tool
-----------------------------------

We also found below potential performance problems on the systems and made
DAMON-based solutions.

The system doesn't want to make the workload suffer from the page
reclamation and thus it utilizes enough DRAM but no swap device.  However,
we found the system is actively reclaiming file-backed pages, because the
system has intensive file IO.  The file IO turned out to be not
performance critical for the workload, but the customer wanted to ensure
performance critical file-backed pages like code section to not mistakenly
be evicted.

Using direct IO should or `mlock()` would be a straightforward solution,
but modifying the user space code is not easy for the customer.
Alternatively, we could use DAMON-based operation scheme[1].  By using it,
we can ask DAMON to track access frequency of each region and make
'process_madvise(MADV_WILLNEED)[2]' call for regions having specific size
and access frequency for a time interval.

We also found the system is having high number of TLB misses.  We tried
'always' THP enabled policy and it greatly reduced TLB misses, but the
page reclamation also been more frequent due to the THP internal
fragmentation caused memory bloat.  We could try another DAMON-based
operation scheme that applies 'MADV_HUGEPAGE' to memory regions having
>=2MB size and high access frequency, while applying 'MADV_NOHUGEPAGE' to
regions having <2MB size and low access frequency.

We do not own the systems so we only reported the analysis results and
possible optimization solutions to the customers.  The customers satisfied
about the analysis results and promised to try the optimization guides.

[1] https://lore.kernel.org/linux-mm/20201006123931.5847-1-sjpark@amazon.com/
[2] https://lore.kernel.org/linux-api/20200622192900.22757-4-minchan@kernel.org/

Comparison with Idle Page Tracking
==================================

Idle Page Tracking allows users to set and read idleness of pages using a
bitmap file which represents each page with each bit of the file.  One
recommended usage of it is working set size detection.  Users can do that
by

    1. find PFN of each page for workloads in interest,
    2. set all the pages as idle by doing writes to the bitmap file,
    3. wait until the workload accesses its working set, and
    4. read the idleness of the pages again and count pages became not idle.

NOTE: While Idle Page Tracking is for user space users, DAMON is primarily
designed for kernel subsystems though it can easily exposed to the user
space.  Hence, this section only assumes such user space use of DAMON.

For what use cases Idle Page Tracking would be better?
------------------------------------------------------

1. Flexible usecases other than hotness monitoring.

Because Idle Page Tracking allows users to control the primitive (Page
idleness) by themselves, Idle Page Tracking users can do anything they
want.  Meanwhile, DAMON is primarily designed to monitor the hotness of
each memory region.  For this, DAMON asks users to provide sampling
interval and aggregation interval.  For the reason, there could be some
use case that using Idle Page Tracking is simpler.

2. Physical memory monitoring.

Idle Page Tracking receives PFN range as input, so natively supports
physical memory monitoring.

DAMON is designed to be extensible for multiple address spaces and use
cases by implementing and using primitives for the given use case.
Therefore, by theory, DAMON has no limitation in the type of target
address space as long as primitives for the given address space exists.
However, the default primitives introduced by this patchset supports only
virtual address spaces.

Therefore, for physical memory monitoring, you should implement your own
primitives and use it, or simply use Idle Page Tracking.

Nonetheless, RFC patchsets[1] for the physical memory address space
primitives is already available.  It also supports user memory same to
Idle Page Tracking.

[1] https://lore.kernel.org/linux-mm/20200831104730.28970-1-sjpark@amazon.com/

For what use cases DAMON is better?
-----------------------------------

1. Hotness Monitoring.

Idle Page Tracking let users know only if a page frame is accessed or not.
For hotness check, the user should write more code and use more memory.
DAMON do that by itself.

2. Low Monitoring Overhead

DAMON receives user's monitoring request with one step and then provide
the results.  So, roughly speaking, DAMON require only O(1) user/kernel
context switches.

In case of Idle Page Tracking, however, because the interface receives
contiguous page frames, the number of user/kernel context switches
increases as the monitoring target becomes complex and huge.  As a result,
the context switch overhead could be not negligible.

Moreover, DAMON is born to handle with the monitoring overhead.  Because
the core mechanism is pure logical, Idle Page Tracking users might be able
to implement the mechanism on their own, but it would be time consuming
and the user/kernel context switching will still more frequent than that
of DAMON.  Also, the kernel subsystems cannot use the logic in this case.

3. Page granularity working set size detection.

Until v22 of this patchset, this was categorized as the thing Idle Page
Tracking could do better, because DAMON basically maintains additional
metadata for each of the monitoring target regions.  So, in the page
granularity working set size detection use case, DAMON would incur (number
of monitoring target pages * size of metadata) memory overhead.  Size of
the single metadata item is about 54 bytes, so assuming 4KB pages, about
1.3% of monitoring target pages will be additionally used.

All essential metadata for Idle Page Tracking are embedded in 'struct
page' and page table entries.  Therefore, in this use case, only one
counter variable for working set size accounting is required if Idle Page
Tracking is used.

There are more details to consider, but roughly speaking, this is true in
most cases.

However, the situation changed from v23.  Now DAMON supports arbitrary
types of monitoring targets, which don't use the metadata.  Using that,
DAMON can do the working set size detection with no additional space
overhead but less user-kernel context switch.  A first draft for the
implementation of monitoring primitives for this usage is available in a
DAMON development tree[1].  An RFC patchset for it based on this patchset
will also be available soon.

Since v24, the arbitrary type support is dropped from this patchset
because this patchset doesn't introduce real use of the type.  You can
still get it from the DAMON development tree[2], though.

[1] https://github.com/sjp38/linux/tree/damon/pgidle_hack
[2] https://github.com/sjp38/linux/tree/damon/master

4. More future usecases

While Idle Page Tracking has tight coupling with base primitives (PG_Idle
and page table Accessed bits), DAMON is designed to be extensible for many
use cases and address spaces.  If you need some special address type or
want to use special h/w access check primitives, you can write your own
primitives for that and configure DAMON to use those.  Therefore, if your
use case could be changed a lot in future, using DAMON could be better.

Can I use both Idle Page Tracking and DAMON?
--------------------------------------------

Yes, though using them concurrently for overlapping memory regions could
result in interference to each other.  Nevertheless, such use case would
be rare or makes no sense at all.  Even in the case, the noise would bot
be really significant.  So, you can choose whatever you want depending on
the characteristics of your use cases.

More Information
================

We prepared a showcase web site[1] that you can get more information.
There are

- the official documentations[2],
- the heatmap format dynamic access pattern of various realistic workloads for
  heap area[3], mmap()-ed area[4], and stack[5] area,
- the dynamic working set size distribution[6] and chronological working set
  size changes[7], and
- the latest performance test results[8].

[1] https://damonitor.github.io/_index
[2] https://damonitor.github.io/doc/html/latest-damon
[3] https://damonitor.github.io/test/result/visual/latest/rec.heatmap.0.png.html
[4] https://damonitor.github.io/test/result/visual/latest/rec.heatmap.1.png.html
[5] https://damonitor.github.io/test/result/visual/latest/rec.heatmap.2.png.html
[6] https://damonitor.github.io/test/result/visual/latest/rec.wss_sz.png.html
[7] https://damonitor.github.io/test/result/visual/latest/rec.wss_time.png.html
[8] https://damonitor.github.io/test/result/perf/latest/html/index.html

Baseline and Complete Git Trees
===============================

The patches are based on the latest -mm tree, specifically
v5.14-rc1-mmots-2021-07-15-18-47 of https://github.com/hnaz/linux-mm.  You can
also clone the complete git tree:

    $ git clone git://github.com/sjp38/linux -b damon/patches/v34

The web is also available:
https://github.com/sjp38/linux/releases/tag/damon/patches/v34

Development Trees
-----------------

There are a couple of trees for entire DAMON patchset series and features
for future release.

- For latest release: https://github.com/sjp38/linux/tree/damon/master
- For next release: https://github.com/sjp38/linux/tree/damon/next

Long-term Support Trees
-----------------------

For people who want to test DAMON but using LTS kernels, there are another
couple of trees based on two latest LTS kernels respectively and
containing the 'damon/master' backports.

- For v5.4.y: https://github.com/sjp38/linux/tree/damon/for-v5.4.y
- For v5.10.y: https://github.com/sjp38/linux/tree/damon/for-v5.10.y

Amazon Linux Kernel Trees
-------------------------

DAMON is also merged in two public Amazon Linux kernel trees that based on
v5.4.y[1] and v5.10.y[2].

[1] https://github.com/amazonlinux/linux/tree/amazon-5.4.y/master/mm/damon
[2] https://github.com/amazonlinux/linux/tree/amazon-5.10.y/master/mm/damon

Git Tree for Diff of Patches
============================

For easy review of diff between different versions of each patch, I
prepared a git tree containing all versions of the DAMON patchset series:
https://github.com/sjp38/damon-patches

You can clone it and use 'diff' for easy review of changes between
different versions of the patchset.  For example:

    $ git clone https://github.com/sjp38/damon-patches && cd damon-patches
    $ diff -u damon/v33 damon/v34

Sequence Of Patches
===================

First three patches implement the core logics of DAMON.  The 1st patch
introduces basic sampling based hotness monitoring for arbitrary types of
targets.  Following two patches implement the core mechanisms for control
of overhead and accuracy, namely regions based sampling (patch 2) and
adaptive regions adjustment (patch 3).

Now the essential parts of DAMON is complete, but it cannot work unless
someone provides monitoring primitives for a specific use case.  The
following two patches make it just work for virtual address spaces
monitoring.  The 4th patch makes 'PG_idle' can be used by DAMON and the
5th patch implements the virtual memory address space specific monitoring
primitives using page table Accessed bits and the 'PG_idle' page flag.

Now DAMON just works for virtual address space monitoring via the kernel
space api.  To let the user space users can use DAMON, following four
patches add interfaces for them.  The 6th patch adds a tracepoint for
monitoring results.  The 7th patch implements a DAMON application kernel
module, namely damon-dbgfs, that simply wraps DAMON and exposes DAMON
interface to the user space via the debugfs interface.  The 8th patch
further exports pid of monitoring thread (kdamond) to user space for
easier cpu usage accounting, and the 9th patch makes the debugfs interface
to support multiple contexts.

Three patches for maintainability follows.  The 10th patch adds
documentations for both the user space and the kernel space.  The 11th
patch provides unit tests (based on the kunit) while the 12th patch adds
user space tests (based on the kselftest).

Finally, the last patch (13th) updates the MAINTAINERS file.

This patch (of 13):

DAMON is a data access monitoring framework for the Linux kernel.  The
core mechanisms of DAMON make it

 - accurate (the monitoring output is useful enough for DRAM level
   performance-centric memory management; It might be inappropriate for
   CPU cache levels, though),
 - light-weight (the monitoring overhead is normally low enough to be
   applied online), and
 - scalable (the upper-bound of the overhead is in constant range
   regardless of the size of target workloads).

Using this framework, hence, we can easily write efficient kernel space
data access monitoring applications.  For example, the kernel's memory
management mechanisms can make advanced decisions using this.
Experimental data access aware optimization works that incurring high
access monitoring overhead could again be implemented on top of this.

Due to its simple and flexible interface, providing user space interface
would be also easy.  Then, user space users who have some special
workloads can write personalized applications for better understanding and
optimizations of their workloads and systems.

===

Nevertheless, this commit is defining and implementing only basic access
check part without the overhead-accuracy handling core logic.  The basic
access check is as below.

The output of DAMON says what memory regions are how frequently accessed
for a given duration.  The resolution of the access frequency is
controlled by setting ``sampling interval`` and ``aggregation interval``.
In detail, DAMON checks access to each page per ``sampling interval`` and
aggregates the results.  In other words, counts the number of the accesses
to each region.  After each ``aggregation interval`` passes, DAMON calls
callback functions that previously registered by users so that users can
read the aggregated results and then clears the results.  This can be
described in below simple pseudo-code::

    init()
    while monitoring_on:
        for page in monitoring_target:
            if accessed(page):
                nr_accesses[page] += 1
        if time() % aggregation_interval == 0:
            for callback in user_registered_callbacks:
                callback(monitoring_target, nr_accesses)
            for page in monitoring_target:
                nr_accesses[page] = 0
        if time() % update_interval == 0:
            update()
        sleep(sampling interval)

The target regions constructed at the beginning of the monitoring and
updated after each ``regions_update_interval``, because the target regions
could be dynamically changed (e.g., mmap() or memory hotplug).  The
monitoring overhead of this mechanism will arbitrarily increase as the
size of the target workload grows.

The basic monitoring primitives for actual access check and dynamic target
regions construction aren't in the core part of DAMON.  Instead, it allows
users to implement their own primitives that are optimized for their use
case and configure DAMON to use those.  In other words, users cannot use
current version of DAMON without some additional works.

Following commits will implement the core mechanisms for the
overhead-accuracy control and default primitives implementations.

Link: https://lkml.kernel.org/r/20210716081449.22187-1-sj38.park@gmail.com
Link: https://lkml.kernel.org/r/20210716081449.22187-2-sj38.park@gmail.com
Signed-off-by: SeongJae Park <sjpark@amazon.de>
Reviewed-by: Leonard Foerster <foersleo@amazon.de>
Reviewed-by: Fernand Sieber <sieberf@amazon.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Amit Shah <amit@kernel.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Woodhouse <dwmw@amazon.com>
Cc: Marco Elver <elver@google.com>
Cc: Fan Du <fan.du@intel.com>
Cc: Greg Kroah-Hartman <greg@kroah.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Maximilian Heyne <mheyne@amazon.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Steven Rostedt (VMware) <rostedt@goodmis.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Brendan Higgins <brendanhiggins@google.com>
Cc: Markus Boehme <markubo@amazon.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>