[RFC,0/4] DAMON based 2-tier memory management for CXL memory

Message ID 20240115045253.1775-1-honggyu.kim@sk.com
Headers
Series DAMON based 2-tier memory management for CXL memory |

Message

Honggyu Kim Jan. 15, 2024, 4:52 a.m. UTC
  There was an RFC IDEA "DAMOS-based Tiered-Memory Management" previously
posted at [1].

It says there is no implementation of the demote/promote DAMOS action
are made.  This RFC is about its implementation for physical address
space.


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

With the advent of CXL/PCIe attached DRAM, which will be called simply
as CXL memory in this cover letter, some systems are becoming more
heterogenous having memory systems with different latency and bandwidth
characteristics.  They are usually handled as different NUMA nodes in
separate memory tiers and CXL memory is used as slow tiers because of
its protocol overhead compared to local DRAM.

In this kind of systems, we need to be careful placing memory pages on
proper NUMA nodes based on the memory access frequency.  Otherwise, some
frequently accessed pages might reside on slow tiers and it makes
performance degradation unexpectedly.  Moreover, the memory access
patterns can be changed at runtime.

To handle this problem, we need a way to monitor the memory access
patterns and migrate pages based on their access temperature.  The
DAMON(Data Access MONitor) framework and its DAMOS(DAMON-based Operation
Schemes) can be useful features for monitoring and migrating pages.
DAMOS provides multiple actions based on DAMON monitoring results and it
can be used for proactive reclaim, which means swapping cold pages out
with DAMOS_PAGEOUT action, but it doesn't support migration actions such
as demotion and promotion between tiered memory nodes.

This series supports two new DAMOS actions; DAMOS_DEMOTE for demotion
from fast tiers and DAMOS_PROMOTE for promotion from slow tiers.  This
prevents hot pages from being stuck on slow tiers, which makes
performance degradation and cold pages can be proactively demoted to
slow tiers so that the system can increase the chance to allocate more
hot pages to fast tiers.

The DAMON provides various tuning knobs but we found that the proactive
demotion for cold pages is especially useful when the system is running
out of memory on its fast tier nodes.

Our evaluation result shows that it reduces the performance slowdown
compared to the default memory policy from 15~17% to 4~5% when the
system runs under high memory pressure on its fast tier DRAM nodes.


DAMON configuration
===================
The specific DAMON configuration doesn't have to be in the scope of this
patch series, but some rough idea is better to be shared to explain the
evaluation result.

The DAMON provides many knobs for fine tuning but its configuration file
is generated by HMSDK[2].  It includes gen_config.py script that
generates a json file with the full config of DAMON knobs and it creates
multiple kdamonds for each NUMA node when the DAMON is enabled so that
it can run hot/cold based migration for tiered memory.


Evaluation Workload
===================

The performance evaluation is done with redis[3], which is a widely used
in-memory database and the memory access patterns are generated via
YCSB[4].  We have measured two different workloads with zipfian and
latest distributions but their configs are slightly modified to make
memory usage higher and execution time longer for better evaluation.

The idea of evaluation using these demote and promote actions covers
system-wide memory management rather than partitioning hot/cold pages of
a single workload.  The default memory allocation policy creates pages
to the fast tier DRAM node first, then allocates newly created pages to
the slow tier CXL node when the DRAM node has insufficient free space.
Once the page allocation is done then those pages never move between
NUMA nodes.  It's not true when using numa balancing, but it is not the
scope of this DAMON based 2-tier memory management support.

If the working set of redis can be fit fully into the DRAM node, then
the redis will access the fast DRAM only.  Since the performance of DRAM
only is faster than partially accessing CXL memory in slow tiers, this
environment is not useful to evaluate this patch series.

To make pages of redis be distributed across fast DRAM node and slow
CXL node to evaluate our demote and promote actions, we pre-allocate
some cold memory externally using mmap and memset before launching
redis-server.  We assumed that there are enough amount of cold memory in
datacenters as TMO[5] and TPP[6] papers mentioned.

The evaluation sequence is as follows.

1. Turn on DAMON with DAMOS_DEMOTE action for DRAM node and
   DAMOS_PROMOTE action for CXL node.  It demotes cold pages on DRAM
   node and promotes hot pages on CXL node in a regular interval.
2. Allocate a huge block of cold memory by calling mmap and memset at
   the fast tier DRAM node, then make the process sleep to make the fast
   tier has insufficient memory for redis-server.
3. Launch redis-server and load prebaked snapshot image, dump.rdb.  The
   redis-server consumes 52GB of anon pages and 33GB of file pages, but
   due to the cold memory allocated at 2, it fails allocating the entire
   memory of redis-server on the fast tier DRAM node so it partially
   allocates the remaining on the slow tier CXL node.  The ratio of
   DRAM:CXL depends on the size of the pre-allocated cold memory.
4. Run YCSB to make zipfian or latest distribution of memory accesses to
   redis-server, then measure its execution time when it's completed.
5. Repeat 4 over 50 times to measure the average execution time for each
   run.
6. Increase the cold memory size then repeat goes to 2.

For each test at 4 took about a minute so repeating it 50 times almost
took about 1 hour for each test with a specific cold memory from 440GB
to 500GB in 10GB increments for each evaluation.  So it took about more
than 10 hours for both zipfian and latest workloads to get the entire
evaluation results.  Repeating the same test set multiple times doesn't
show much difference so I think it might be enough to make the result
reliable.


Evaluation Results
==================

All the result values are normalized to DRAM-only execution time because
the workload cannot be faster than DRAM-only unless the workload hits
the bandwidth peak but our redis test doesn't go beyond the bandwidth
limit.

So the DRAM-only execution time is the ideal result without affected by
the gap between DRAM and CXL performance difference.  The NUMA node
environment is as follows.

  node0 - local DRAM, 512GB with a CPU socket (fast tier)
  node1 - disabled
  node2 - CXL DRAM, 96GB, no CPU attached (slow tier)

The following is the result of generating zipfian distribution to
redis-server and the numbers are averaged by 50 times of execution.

  1. YCSB zipfian distribution read only workload
  memory pressure with cold memory on node0 with 512GB of local DRAM.
  =============+================================================+=========
               |       cold memory occupied by mmap and memset  |
               |   0G  440G  450G  460G  470G  480G  490G  500G |
  =============+================================================+=========
  Execution time normalized to DRAM-only values                 | GEOMEAN
  -------------+------------------------------------------------+---------
  DRAM-only    | 1.00     -     -     -     -     -     -     - | 1.00
  CXL-only     | 1.21     -     -     -     -     -     -     - | 1.21
  default      |    -  1.09  1.10  1.13  1.15  1.18  1.21  1.21 | 1.15
  DAMON 2-tier |    -  1.02  1.04  1.05  1.04  1.05  1.05  1.06 | 1.04
  =============+================================================+=========
  CXL usage of redis-server in GB                               | AVERAGE
  -------------+------------------------------------------------+---------
  DRAM-only    |  0.0     -     -     -     -     -     -     - |  0.0
  CXL-only     | 52.6     -     -     -     -     -     -     - | 52.6
  default      |    -  19.4  26.1  32.3  38.5  44.7  50.5  50.3 | 37.4
  DAMON 2-tier |    -   0.1   1.6   5.2   8.0   9.1  11.8  13.6 |  7.1
  =============+================================================+=========

Each test result is based on the exeuction environment as follows.

  DRAM-only   : redis-server uses only local DRAM memory.
  CXL-only    : redis-server uses only CXL memory.
  default     : default memory policy(MPOL_DEFAULT).
                numa balancing disabled.
  DAMON 2-tier: DAMON enabled with DAMOS_DEMOTE for DRAM nodes and
                DAMOS_PROMOTE for CXL nodes.

The above result shows the "default" execution time goes up as the size
of cold memory is increased from 440G to 500G because the more cold
memory used, the more CXL memory is used for the target redis workload
and this makes the execution time increase.

However, "DAMON 2-tier" result shows less slowdown because the
DAMOS_DEMOTE action at DRAM node proactively demotes pre-allocated cold
memory to CXL node and this free space at DRAM increases more chance to
allocate hot or warm pages of redis-server to fast DRAM node.  Moreover,
DEMOS_PROMOTE action at CXL node also promotes hot pages of redis-server
to DRAM node actively.

As a result, it makes more memory of redis-server stay in DRAM node
compared to "default" memory policy and this makes the performance
improvement.

The following result of latest distribution workload shows similar data.

  2. YCSB latest distribution read only workload
  memory pressure with cold memory on node0 with 512GB of local DRAM.
  =============+================================================+=========
               |       cold memory occupied by mmap and memset  |
               |   0G  440G  450G  460G  470G  480G  490G  500G |
  =============+================================================+=========
  Execution time normalized to DRAM-only values                 | GEOMEAN
  -------------+------------------------------------------------+---------
  DRAM-only    | 1.00  -     -     -     -     -     -     -    | 1.00
  CXL-only     | 1.18  -     -     -     -     -     -     -    | 1.18
  default      | -     1.16  1.15  1.17  1.18  1.16  1.18  1.15 | 1.17
  DAMON 2-tier | -     1.04  1.04  1.05  1.05  1.06  1.05  1.06 | 1.05
  =============+================================================+=========
  CXL usage of redis-server in GB                               | AVERAGE
  -------------+------------------------------------------------+---------
  DRAM-only    |  0.0     -     -     -     -     -     -     - |  0.0
  CXL-only     | 52.6     -     -     -     -     -     -     - | 52.6
  default      |    -  19.3  26.1  32.2  38.5  44.6  50.5  50.6 | 37.4
  DAMON 2-tier |    -   1.3   3.8   7.0   4.1   9.4  12.5  16.7 |  7.8
  =============+================================================+=========

In summary of both results, our evaluation shows that "DAMON 2-tier"
memory management reduces the performance slowdown compared to the
"default" memory policy from 15~17% to 4~5% when the system runs with
high memory pressure on its fast tier DRAM nodes.

The similar evaluation was done in another machine that has 256GB of
local DRAM and 96GB of CXL memory.  The performance slowdown is reduced
from 20~24% for "default" to 5~7% for "DAMON 2-tier".

Having these DAMOS_DEMOTE and DAMOS_PROMOTE actions can make 2-tier
memory systems run more efficiently under high memory pressures.

Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
Signed-off-by: Rakie Kim <rakie.kim@sk.com>

[1] https://lore.kernel.org/damon/20231112195602.61525-1-sj@kernel.org
[2] https://github.com/skhynix/hmsdk
[3] https://github.com/redis/redis/tree/7.0.0
[4] https://github.com/brianfrankcooper/YCSB/tree/0.17.0
[5] https://dl.acm.org/doi/10.1145/3503222.3507731
[6] https://dl.acm.org/doi/10.1145/3582016.3582063

Honggyu Kim (2):
  mm/vmscan: refactor reclaim_pages with reclaim_or_migrate_folios
  mm/damon: introduce DAMOS_DEMOTE action for demotion

Hyeongtak Ji (2):
  mm/memory-tiers: add next_promotion_node to find promotion target
  mm/damon: introduce DAMOS_PROMOTE action for promotion

 include/linux/damon.h          |   4 +
 include/linux/memory-tiers.h   |  11 ++
 include/linux/migrate_mode.h   |   1 +
 include/linux/vm_event_item.h  |   1 +
 include/trace/events/migrate.h |   3 +-
 mm/damon/paddr.c               |  46 ++++++-
 mm/damon/sysfs-schemes.c       |   2 +
 mm/internal.h                  |   2 +
 mm/memory-tiers.c              |  43 ++++++
 mm/vmscan.c                    | 231 +++++++++++++++++++++++++++++++--
 mm/vmstat.c                    |   1 +
 11 files changed, 330 insertions(+), 15 deletions(-)


base-commit: 0dd3ee31125508cd67f7e7172247f05b7fd1753a
  

Comments

SeongJae Park Jan. 16, 2024, 8:31 p.m. UTC | #1
Hello,

On Mon, 15 Jan 2024 13:52:48 +0900 Honggyu Kim <honggyu.kim@sk.com> wrote:

> There was an RFC IDEA "DAMOS-based Tiered-Memory Management" previously
> posted at [1].
> 
> It says there is no implementation of the demote/promote DAMOS action
> are made.  This RFC is about its implementation for physical address
> space.
> 
[...]
> Evaluation Results
> ==================
> 
[...]
> In summary of both results, our evaluation shows that "DAMON 2-tier"
> memory management reduces the performance slowdown compared to the
> "default" memory policy from 15~17% to 4~5% when the system runs with
> high memory pressure on its fast tier DRAM nodes.
> 
> The similar evaluation was done in another machine that has 256GB of
> local DRAM and 96GB of CXL memory.  The performance slowdown is reduced
> from 20~24% for "default" to 5~7% for "DAMON 2-tier".
> 
> Having these DAMOS_DEMOTE and DAMOS_PROMOTE actions can make 2-tier
> memory systems run more efficiently under high memory pressures.


Thank you so much for this great patches and the above nice test results.  I
believe the test setup and results make sense, and merging a revised version of
this patchset would provide real benefits to the users.

In a high level, I think it might better to separate DAMON internal changes
from DAMON external changes.

For DAMON part changes, I have no big concern other than trivial coding style
level comments.

For DAMON-external changes that implementing demote_pages() and
promote_pages(), I'm unsure if the implementation is reusing appropriate
functions, and if those are placee in right source file.  Especially, I'm
unsure if vmscan.c is the right place for promotion code.  Also I don't know if
there is a good agreement on the promotion/demotion target node decision.  That
should be because I'm not that familiar with the areas and the files, but I
feel this might because our discussions on the promotion and the demotion
operations are having rooms for being more matured.  Because I'm not very
faimiliar with the part, I'd like to hear others' comments, too.

To this end, I feel the problem might be able to be simpler, because this
patchset is trying to provide two sophisticated operations, while I think a
simpler approach might be possible.  My humble simpler idea is adding a DAMOS
operation for moving pages to a given node (like sys_move_phy_pages RFC[1]),
instead of the promote/demote.  Because the general pages migration can handle
multiple cases including the promote/demote in my humble assumption.  In more
detail, users could decide which is the appropriate node for promotion or
demotion and use the new DAMOS action to do promotion and demotion.  Users
would requested to decide which node is the proper promotion/demotion target
nodes, but that decision wouldn't be that hard in my opinion.

For this, 'struct damos' would need to be updated for such argument-dependent
actions, like 'struct damos_filter' is haing a union.

In future, we could extend the operation to the promotion and the demotion
after the dicussion around the promotion and demotion is matured, if required.
And assuming DAMON be extended for originating CPU-aware access monitoring, the
new DAMOS action would also cover more use cases such as general NUMA nodes
balancing (extending DAMON for CPU-aware monitoring would required), and some
complex configurations where having both CPU affinity and tiered memory.  I
also think that may well fit with my RFC idea[2] for tiered memory management.

Looking forward to opinions from you and others.  I admig I miss many things,
and more than happy to be enlightened.

[1] https://lwn.net/Articles/944007/
[2] https://lore.kernel.org/damon/20231112195602.61525-1-sj@kernel.org/


Thanks,
SJ

> 
> Signed-off-by: Honggyu Kim <honggyu.kim@sk.com>
> Signed-off-by: Hyeongtak Ji <hyeongtak.ji@sk.com>
> Signed-off-by: Rakie Kim <rakie.kim@sk.com>
> 
> [1] https://lore.kernel.org/damon/20231112195602.61525-1-sj@kernel.org
> [2] https://github.com/skhynix/hmsdk
> [3] https://github.com/redis/redis/tree/7.0.0
> [4] https://github.com/brianfrankcooper/YCSB/tree/0.17.0
> [5] https://dl.acm.org/doi/10.1145/3503222.3507731
> [6] https://dl.acm.org/doi/10.1145/3582016.3582063
> 
> Honggyu Kim (2):
>   mm/vmscan: refactor reclaim_pages with reclaim_or_migrate_folios
>   mm/damon: introduce DAMOS_DEMOTE action for demotion
> 
> Hyeongtak Ji (2):
>   mm/memory-tiers: add next_promotion_node to find promotion target
>   mm/damon: introduce DAMOS_PROMOTE action for promotion
> 
>  include/linux/damon.h          |   4 +
>  include/linux/memory-tiers.h   |  11 ++
>  include/linux/migrate_mode.h   |   1 +
>  include/linux/vm_event_item.h  |   1 +
>  include/trace/events/migrate.h |   3 +-
>  mm/damon/paddr.c               |  46 ++++++-
>  mm/damon/sysfs-schemes.c       |   2 +
>  mm/internal.h                  |   2 +
>  mm/memory-tiers.c              |  43 ++++++
>  mm/vmscan.c                    | 231 +++++++++++++++++++++++++++++++--
>  mm/vmstat.c                    |   1 +
>  11 files changed, 330 insertions(+), 15 deletions(-)
> 
> 
> base-commit: 0dd3ee31125508cd67f7e7172247f05b7fd1753a
> -- 
> 2.34.1
  
Honggyu Kim Jan. 17, 2024, 11:49 a.m. UTC | #2
Hi SeongJae,

Thanks very much for your comments in details.

On Tue, 16 Jan 2024 12:31:59 -0800 SeongJae Park <sj@kernel.org> wrote:

> Thank you so much for this great patches and the above nice test results.  I
> believe the test setup and results make sense, and merging a revised version of
> this patchset would provide real benefits to the users.

Glad to hear that!

> In a high level, I think it might better to separate DAMON internal changes
> from DAMON external changes.

I agree.  I can't guarantee but I can move all the external changes
inside mm/damon, but will try that as much as possible.

> For DAMON part changes, I have no big concern other than trivial coding style
> level comments.

Sure.  I will fix those.

> For DAMON-external changes that implementing demote_pages() and
> promote_pages(), I'm unsure if the implementation is reusing appropriate
> functions, and if those are placee in right source file.  Especially, I'm
> unsure if vmscan.c is the right place for promotion code.  Also I don't know if
> there is a good agreement on the promotion/demotion target node decision.  That
> should be because I'm not that familiar with the areas and the files, but I
> feel this might because our discussions on the promotion and the demotion
> operations are having rooms for being more matured.  Because I'm not very
> faimiliar with the part, I'd like to hear others' comments, too.

I would also like to hear others' comments, but this might not be needed
if most of external code can be moved to mm/damon.

> To this end, I feel the problem might be able te be simpler, because this
> patchset is trying to provide two sophisticated operations, while I think a
> simpler approach might be possible.  My humble simpler idea is adding a DAMOS
> operation for moving pages to a given node (like sys_move_phy_pages RFC[1]),
> instead of the promote/demote.  Because the general pages migration can handle
> multiple cases including the promote/demote in my humble assumption.

My initial implementation was similar but I found that it's not accurate
enough due to the nature of inaccuracy of DAMON regions.  I saw that
many pages were demoted and promoted back and forth because migration
target regions include both hot and cold pages together.

So I have implemented the demotion and promotion logics based on the
shrink_folio_list, which contains many corner case handling logics for
reclaim.

Having the current demotion and promotion logics makes the hot/cold
migration pretty accurate as expected.  We made a simple program called
"hot_cold" and it receives 2 arguments for hot size and cold size in MB.
For example, "hot_cold 200 500" allocates 200MB of hot memory and 500MB
of cold memory.  It basically allocates 2 large blocks of memory with
mmap, then repeat memset for the initial 200MB to make it accessed in an
infinite loop.

Let's say there are 3 nodes in the system and the first node0 and node1
are the first tier, and node2 is the second tier.

  $ cat /sys/devices/virtual/memory_tiering/memory_tier4/nodelist
  0-1

  $ cat /sys/devices/virtual/memory_tiering/memory_tier22/nodelist
  2

Here is the result of partitioning hot/cold memory and I put execution
command at the right side of numastat result.  I initially ran each
hot_cold program with preferred setting so that they initially allocate
memory on one of node0 or node2, but they gradually migrated based on
their access frequencies.

  $ numastat -c -p hot_cold
  Per-node process memory usage (in MBs) 
  PID              Node 0 Node 1 Node 2 Total 
  ---------------  ------ ------ ------ ----- 
  754 (hot_cold)     1800      0   2000  3800    <- hot_cold 1800 2000 
  1184 (hot_cold)     300      0    500   800    <- hot_cold 300 500 
  1818 (hot_cold)     801      0   3199  4000    <- hot_cold 800 3200 
  30289 (hot_cold)      4      0      5    10    <- hot_cold 3 5 
  30325 (hot_cold)     31      0     51    81    <- hot_cold 30 50 
  ---------------  ------ ------ ------ ----- 
  Total              2938      0   5756  8695

The final node placement result shows that DAMON accurately migrated
pages by their hotness for multiple processes.

> In more detail, users could decide which is the appropriate node for promotion
> or demotion and use the new DAMOS action to do promotion and demotion.  Users
> would requested to decide which node is the proper promotion/demotion target
> nodes, but that decision wouldn't be that hard in my opinion.
> 
> For this, 'struct damos' would need to be updated for such argument-dependent
> actions, like 'struct damos_filter' is haing a union.

That might be a better solution.  I will think about it.

> In future, we could extend the operation to the promotion and the demotion
> after the dicussion around the promotion and demotion is matured, if required.
> And assuming DAMON be extended for originating CPU-aware access monitoring, the
> new DAMOS action would also cover more use cases such as general NUMA nodes
> balancing (extending DAMON for CPU-aware monitoring would required), and some
> complex configurations where having both CPU affinity and tiered memory.  I
> also think that may well fit with my RFC idea[2] for tiered memory management.
> 
> Looking forward to opinions from you and others.  I admig I miss many things,
> and more than happy to be enlightened.
> 
> [1] https://lwn.net/Articles/944007/
> [2] https://lore.kernel.org/damon/20231112195602.61525-1-sj@kernel.org/

Thanks very much for your comments.  I will need a few more days for the
update but will try to address your concerns as much as possible.

Thanks,
Honggyu
  
SeongJae Park Jan. 17, 2024, 9:11 p.m. UTC | #3
Hi Honggyu,

On Wed, 17 Jan 2024 20:49:25 +0900 Honggyu Kim <honggyu.kim@sk.com> wrote:

> Hi SeongJae,
> 
> Thanks very much for your comments in details.
> 
> On Tue, 16 Jan 2024 12:31:59 -0800 SeongJae Park <sj@kernel.org> wrote:
> 
> > Thank you so much for this great patches and the above nice test results.  I
> > believe the test setup and results make sense, and merging a revised version of
> > this patchset would provide real benefits to the users.
> 
> Glad to hear that!
> 
> > In a high level, I think it might better to separate DAMON internal changes
> > from DAMON external changes.
> 
> I agree.  I can't guarantee but I can move all the external changes
> inside mm/damon, but will try that as much as possible.
> 
> > For DAMON part changes, I have no big concern other than trivial coding style
> > level comments.
> 
> Sure.  I will fix those.
> 
> > For DAMON-external changes that implementing demote_pages() and
> > promote_pages(), I'm unsure if the implementation is reusing appropriate
> > functions, and if those are placee in right source file.  Especially, I'm
> > unsure if vmscan.c is the right place for promotion code.  Also I don't know if
> > there is a good agreement on the promotion/demotion target node decision.  That
> > should be because I'm not that familiar with the areas and the files, but I
> > feel this might because our discussions on the promotion and the demotion
> > operations are having rooms for being more matured.  Because I'm not very
> > faimiliar with the part, I'd like to hear others' comments, too.
> 
> I would also like to hear others' comments, but this might not be needed
> if most of external code can be moved to mm/damon.
> 
> > To this end, I feel the problem might be able te be simpler, because this
> > patchset is trying to provide two sophisticated operations, while I think a
> > simpler approach might be possible.  My humble simpler idea is adding a DAMOS
> > operation for moving pages to a given node (like sys_move_phy_pages RFC[1]),
> > instead of the promote/demote.  Because the general pages migration can handle
> > multiple cases including the promote/demote in my humble assumption.
> 
> My initial implementation was similar but I found that it's not accurate
> enough due to the nature of inaccuracy of DAMON regions.  I saw that
> many pages were demoted and promoted back and forth because migration
> target regions include both hot and cold pages together.
> 
> So I have implemented the demotion and promotion logics based on the
> shrink_folio_list, which contains many corner case handling logics for
> reclaim.
> 
> Having the current demotion and promotion logics makes the hot/cold
> migration pretty accurate as expected.  We made a simple program called
> "hot_cold" and it receives 2 arguments for hot size and cold size in MB.
> For example, "hot_cold 200 500" allocates 200MB of hot memory and 500MB
> of cold memory.  It basically allocates 2 large blocks of memory with
> mmap, then repeat memset for the initial 200MB to make it accessed in an
> infinite loop.
> 
> Let's say there are 3 nodes in the system and the first node0 and node1
> are the first tier, and node2 is the second tier.
> 
>   $ cat /sys/devices/virtual/memory_tiering/memory_tier4/nodelist
>   0-1
> 
>   $ cat /sys/devices/virtual/memory_tiering/memory_tier22/nodelist
>   2
> 
> Here is the result of partitioning hot/cold memory and I put execution
> command at the right side of numastat result.  I initially ran each
> hot_cold program with preferred setting so that they initially allocate
> memory on one of node0 or node2, but they gradually migrated based on
> their access frequencies.
> 
>   $ numastat -c -p hot_cold
>   Per-node process memory usage (in MBs) 
>   PID              Node 0 Node 1 Node 2 Total 
>   ---------------  ------ ------ ------ ----- 
>   754 (hot_cold)     1800      0   2000  3800    <- hot_cold 1800 2000 
>   1184 (hot_cold)     300      0    500   800    <- hot_cold 300 500 
>   1818 (hot_cold)     801      0   3199  4000    <- hot_cold 800 3200 
>   30289 (hot_cold)      4      0      5    10    <- hot_cold 3 5 
>   30325 (hot_cold)     31      0     51    81    <- hot_cold 30 50 
>   ---------------  ------ ------ ------ ----- 
>   Total              2938      0   5756  8695
> 
> The final node placement result shows that DAMON accurately migrated
> pages by their hotness for multiple processes.

What was the result when the corner cases handling logics were not applied?

And, what are the corner cases handling logic that seemed essential?  I show
the page granularity active/reference check could indeed provide many
improvements, but that's only my humble assumption.

If the corner cases are indeed better to be applied in page granularity, I
agree we need some more efforts since DAMON monitoring results are not page
granularity aware by the design.  Users could increase min_nr_regions to make
it more accurate, and we have plan to support page granularity monitoring,
though.  But maybe the overhead could be unacceptable.

Ideal solution would be making DAMON more accurate while keeping current level
of overhead.  We indeed have TODO items for DAMON accuracy improvement, but
this may take some time that might unacceptable for your case.

If that's the case, I think the additional corner handling (or, page gran
additional access check) could be made as DAMOS filters[1], since DAMOS filters
can be applied in page granularity, and designed for this kind of handling of
information that DAMON monitoring results cannot provide.  More specifically,
we could have filters for promotion-qualifying pages and demotion-qualifying
pages.  In this way, I think we can keep the action more flexible while the
filters can be applied in creative ways.

[1] https://git.kernel.org/sj/c/98def236f63c66629fb6b2d4b69cecffc5b46539

> 
> > In more detail, users could decide which is the appropriate node for promotion
> > or demotion and use the new DAMOS action to do promotion and demotion.  Users
> > would requested to decide which node is the proper promotion/demotion target
> > nodes, but that decision wouldn't be that hard in my opinion.
> > 
> > For this, 'struct damos' would need to be updated for such argument-dependent
> > actions, like 'struct damos_filter' is haing a union.
> 
> That might be a better solution.  I will think about it.

More specifically, I think receiving an address range as the argument might
more flexible than just NUMA node.  Maybe we can imagine proactively migrating
cold movable pages from normal zones to movable zones, to avoid normal zone
memory pressure.

> 
> > In future, we could extend the operation to the promotion and the demotion
> > after the dicussion around the promotion and demotion is matured, if required.
> > And assuming DAMON be extended for originating CPU-aware access monitoring, the
> > new DAMOS action would also cover more use cases such as general NUMA nodes
> > balancing (extending DAMON for CPU-aware monitoring would required), and some
> > complex configurations where having both CPU affinity and tiered memory.  I
> > also think that may well fit with my RFC idea[2] for tiered memory management.
> > 
> > Looking forward to opinions from you and others.  I admig I miss many things,
> > and more than happy to be enlightened.
> > 
> > [1] https://lwn.net/Articles/944007/
> > [2] https://lore.kernel.org/damon/20231112195602.61525-1-sj@kernel.org/
> 
> Thanks very much for your comments.  I will need a few more days for the
> update but will try to address your concerns as much as possible.

No problem, please take your time.  I'm looking forward to the next version :)


Thanks,
SJ

> 
> Thanks,
> Honggyu
  
SeongJae Park Jan. 17, 2024, 9:24 p.m. UTC | #4
On Wed, 17 Jan 2024 13:11:03 -0800 SeongJae Park <sj@kernel.org> wrote:

[...]
> Hi Honggyu,
> 
> On Wed, 17 Jan 2024 20:49:25 +0900 Honggyu Kim <honggyu.kim@sk.com> wrote:
> 
> > Hi SeongJae,
> > 
> > Thanks very much for your comments in details.
> > 
> > On Tue, 16 Jan 2024 12:31:59 -0800 SeongJae Park <sj@kernel.org> wrote:
> > 
[...]
> > > To this end, I feel the problem might be able te be simpler, because this
> > > patchset is trying to provide two sophisticated operations, while I think a
> > > simpler approach might be possible.  My humble simpler idea is adding a DAMOS
> > > operation for moving pages to a given node (like sys_move_phy_pages RFC[1]),
> > > instead of the promote/demote.  Because the general pages migration can handle
> > > multiple cases including the promote/demote in my humble assumption.
[...]
> > > In more detail, users could decide which is the appropriate node for promotion
> > > or demotion and use the new DAMOS action to do promotion and demotion.  Users
> > > would requested to decide which node is the proper promotion/demotion target
> > > nodes, but that decision wouldn't be that hard in my opinion.
> > > 
> > > For this, 'struct damos' would need to be updated for such argument-dependent
> > > actions, like 'struct damos_filter' is haing a union.
> > 
> > That might be a better solution.  I will think about it.
> 
> More specifically, I think receiving an address range as the argument might
> more flexible than just NUMA node.  Maybe we can imagine proactively migrating
> cold movable pages from normal zones to movable zones, to avoid normal zone
> memory pressure.

Yet another crazy idea.  Finding hot regions in the middle of cold region and
move to besides of other hot pages.  As a result, memory is sorted by access
temperature even in same node, and the system gains more spatial locality,
which benefits general locality-based algorithms including DAMON's adaptive
regions adjustment.


Thanks,
SJ

[...]
  
Hyeongtak Ji Jan. 18, 2024, 10:40 a.m. UTC | #5
Hi SeongJae,

On Wed, 17 Jan 2024 SeongJae Park <sj@kernel.org> wrote:

[...]
>> Let's say there are 3 nodes in the system and the first node0 and node1
>> are the first tier, and node2 is the second tier.
>> 
>>   $ cat /sys/devices/virtual/memory_tiering/memory_tier4/nodelist
>>   0-1
>> 
>>   $ cat /sys/devices/virtual/memory_tiering/memory_tier22/nodelist
>>   2
>> 
>> Here is the result of partitioning hot/cold memory and I put execution
>> command at the right side of numastat result.  I initially ran each
>> hot_cold program with preferred setting so that they initially allocate
>> memory on one of node0 or node2, but they gradually migrated based on
>> their access frequencies.
>> 
>>   $ numastat -c -p hot_cold
>>   Per-node process memory usage (in MBs) 
>>   PID              Node 0 Node 1 Node 2 Total 
>>   ---------------  ------ ------ ------ ----- 
>>   754 (hot_cold)     1800      0   2000  3800    <- hot_cold 1800 2000 
>>   1184 (hot_cold)     300      0    500   800    <- hot_cold 300 500 
>>   1818 (hot_cold)     801      0   3199  4000    <- hot_cold 800 3200 
>>   30289 (hot_cold)      4      0      5    10    <- hot_cold 3 5 
>>   30325 (hot_cold)     31      0     51    81    <- hot_cold 30 50 
>>   ---------------  ------ ------ ------ ----- 
>>   Total              2938      0   5756  8695
>> 
>> The final node placement result shows that DAMON accurately migrated
>> pages by their hotness for multiple processes.
>
> What was the result when the corner cases handling logics were not applied?

This is the result of the same test that Honggyu did, but with an insufficient
corner cases handling logics.

  $ numastat -c -p hot_cold

  Per-node process memory usage (in MBs)
  PID             Node 0 Node 1 Node 2 Total
  --------------  ------ ------ ------ -----
  862 (hot_cold)    2256      0   1545  3801   <- hot_cold 1800 2000
  863 (hot_cold)     403      0    398   801   <- hot_cold 300 500
  864 (hot_cold)    1520      0   2482  4001   <- hot_cold 800 3200
  865 (hot_cold)       6      0      3     9   <- hot_cold 3 5
  866 (hot_cold)      29      0     52    81   <- hot_cold 30 50
  --------------  ------ ------ ------ -----
  Total             4215      0   4480  8695

As time goes by, DAMON keeps trying to split the hot/cold region, but it does
not seem to be enough.

  $ numastat -c -p hot_cold

  Per-node process memory usage (in MBs)
  PID             Node 0 Node 1 Node 2 Total
  --------------  ------ ------ ------ -----
  862 (hot_cold)    2022      0   1780  3801   <- hot_cold 1800 2000
  863 (hot_cold)     351      0    450   801   <- hot_cold 300 500
  864 (hot_cold)    1134      0   2868  4001   <- hot_cold 800 3200
  865 (hot_cold)       7      0      2     9   <- hot_cold 3 5
  866 (hot_cold)      43      0     39    81   <- hot_cold 30 50
  --------------  ------ ------ ------ -----
  Total             3557      0   5138  8695

>
> And, what are the corner cases handling logic that seemed essential?  I show
> the page granularity active/reference check could indeed provide many
> improvements, but that's only my humble assumption.

Yes, the page granularity active/reference check is essential.  To make the
above "insufficient" result, the only thing I did was to promote
inactive/not_referenced pages.

diff --git a/mm/vmscan.c b/mm/vmscan.c
index f03be320f9ad..c2aefb883c54 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1127,9 +1127,7 @@ static unsigned int __promote_folio_list(struct list_head *folio_list,
                VM_BUG_ON_FOLIO(folio_test_active(folio), folio);

                references = folio_check_references(folio, sc);
-               if (references == FOLIOREF_KEEP ||
-                   references == FOLIOREF_RECLAIM ||
-                   references == FOLIOREF_RECLAIM_CLEAN)
+               if (references == FOLIOREF_KEEP )
                        goto keep_locked;

                /* Relocate its contents to another node. */

>
> If the corner cases are indeed better to be applied in page granularity, I
> agree we need some more efforts since DAMON monitoring results are not page
> granularity aware by the design.  Users could increase min_nr_regions to make
> it more accurate, and we have plan to support page granularity monitoring,
> though.  But maybe the overhead could be unacceptable.
>
> Ideal solution would be making DAMON more accurate while keeping current level
> of overhead.  We indeed have TODO items for DAMON accuracy improvement, but
> this may take some time that might unacceptable for your case.
>
> If that's the case, I think the additional corner handling (or, page gran
> additional access check) could be made as DAMOS filters[1], since DAMOS filters
> can be applied in page granularity, and designed for this kind of handling of
> information that DAMON monitoring results cannot provide.  More specifically,
> we could have filters for promotion-qualifying pages and demotion-qualifying
> pages.  In this way, I think we can keep the action more flexible while the
> filters can be applied in creative ways.

Making corner handling as a new DAMOS filters is a good idea.  I'm just a bit
concerned if adding new filters might cause users to care more.

Kind regards,
Hyeongtak
  
SeongJae Park Jan. 18, 2024, 5:17 p.m. UTC | #6
On Thu, 18 Jan 2024 19:40:16 +0900 Hyeongtak Ji <hyeongtak.ji@sk.com> wrote:

> Hi SeongJae,
> 
> On Wed, 17 Jan 2024 SeongJae Park <sj@kernel.org> wrote:
> 
> [...]
> >> Let's say there are 3 nodes in the system and the first node0 and node1
> >> are the first tier, and node2 is the second tier.
> >> 
> >>   $ cat /sys/devices/virtual/memory_tiering/memory_tier4/nodelist
> >>   0-1
> >> 
> >>   $ cat /sys/devices/virtual/memory_tiering/memory_tier22/nodelist
> >>   2
> >> 
> >> Here is the result of partitioning hot/cold memory and I put execution
> >> command at the right side of numastat result.  I initially ran each
> >> hot_cold program with preferred setting so that they initially allocate
> >> memory on one of node0 or node2, but they gradually migrated based on
> >> their access frequencies.
> >> 
> >>   $ numastat -c -p hot_cold
> >>   Per-node process memory usage (in MBs) 
> >>   PID              Node 0 Node 1 Node 2 Total 
> >>   ---------------  ------ ------ ------ ----- 
> >>   754 (hot_cold)     1800      0   2000  3800    <- hot_cold 1800 2000 
> >>   1184 (hot_cold)     300      0    500   800    <- hot_cold 300 500 
> >>   1818 (hot_cold)     801      0   3199  4000    <- hot_cold 800 3200 
> >>   30289 (hot_cold)      4      0      5    10    <- hot_cold 3 5 
> >>   30325 (hot_cold)     31      0     51    81    <- hot_cold 30 50 
> >>   ---------------  ------ ------ ------ ----- 
> >>   Total              2938      0   5756  8695
> >> 
> >> The final node placement result shows that DAMON accurately migrated
> >> pages by their hotness for multiple processes.
> >
> > What was the result when the corner cases handling logics were not applied?
> 
> This is the result of the same test that Honggyu did, but with an insufficient
> corner cases handling logics.
> 
>   $ numastat -c -p hot_cold
> 
>   Per-node process memory usage (in MBs)
>   PID             Node 0 Node 1 Node 2 Total
>   --------------  ------ ------ ------ -----
>   862 (hot_cold)    2256      0   1545  3801   <- hot_cold 1800 2000
>   863 (hot_cold)     403      0    398   801   <- hot_cold 300 500
>   864 (hot_cold)    1520      0   2482  4001   <- hot_cold 800 3200
>   865 (hot_cold)       6      0      3     9   <- hot_cold 3 5
>   866 (hot_cold)      29      0     52    81   <- hot_cold 30 50
>   --------------  ------ ------ ------ -----
>   Total             4215      0   4480  8695
> 
> As time goes by, DAMON keeps trying to split the hot/cold region, but it does
> not seem to be enough.
> 
>   $ numastat -c -p hot_cold
> 
>   Per-node process memory usage (in MBs)
>   PID             Node 0 Node 1 Node 2 Total
>   --------------  ------ ------ ------ -----
>   862 (hot_cold)    2022      0   1780  3801   <- hot_cold 1800 2000
>   863 (hot_cold)     351      0    450   801   <- hot_cold 300 500
>   864 (hot_cold)    1134      0   2868  4001   <- hot_cold 800 3200
>   865 (hot_cold)       7      0      2     9   <- hot_cold 3 5
>   866 (hot_cold)      43      0     39    81   <- hot_cold 30 50
>   --------------  ------ ------ ------ -----
>   Total             3557      0   5138  8695
> 
> >
> > And, what are the corner cases handling logic that seemed essential?  I show
> > the page granularity active/reference check could indeed provide many
> > improvements, but that's only my humble assumption.
> 
> Yes, the page granularity active/reference check is essential.  To make the
> above "insufficient" result, the only thing I did was to promote
> inactive/not_referenced pages.
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index f03be320f9ad..c2aefb883c54 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -1127,9 +1127,7 @@ static unsigned int __promote_folio_list(struct list_head *folio_list,
>                 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
> 
>                 references = folio_check_references(folio, sc);
> -               if (references == FOLIOREF_KEEP ||
> -                   references == FOLIOREF_RECLAIM ||
> -                   references == FOLIOREF_RECLAIM_CLEAN)
> +               if (references == FOLIOREF_KEEP )
>                         goto keep_locked;
> 
>                 /* Relocate its contents to another node. */

Thank you for sharing the details :)  I think DAMOS filters based approach
could be worthy to try, then.

> 
> >
> > If the corner cases are indeed better to be applied in page granularity, I
> > agree we need some more efforts since DAMON monitoring results are not page
> > granularity aware by the design.  Users could increase min_nr_regions to make
> > it more accurate, and we have plan to support page granularity monitoring,
> > though.  But maybe the overhead could be unacceptable.
> >
> > Ideal solution would be making DAMON more accurate while keeping current level
> > of overhead.  We indeed have TODO items for DAMON accuracy improvement, but
> > this may take some time that might unacceptable for your case.
> >
> > If that's the case, I think the additional corner handling (or, page gran
> > additional access check) could be made as DAMOS filters[1], since DAMOS filters
> > can be applied in page granularity, and designed for this kind of handling of
> > information that DAMON monitoring results cannot provide.  More specifically,
> > we could have filters for promotion-qualifying pages and demotion-qualifying
> > pages.  In this way, I think we can keep the action more flexible while the
> > filters can be applied in creative ways.
> 
> Making corner handling as a new DAMOS filters is a good idea.  I'm just a bit
> concerned if adding new filters might cause users to care more.

I prefer keeping DAMON API and Sysfs interface flexible and easy to extended
even if it increases number of parameters, while providing simplified
high level interfaces for end users aiming to use DAMON for specific use cases,
like DAMON_RECLAIM, DAMON_LRU_SORT, and damo do.  Hence I'm not very concerned.


Thanks,
SJ

> 
> Kind regards,
> Hyeongtak