Patch series "mm/damon: auto-tune DAMOS for NUMA setups including tiered
memory".
Utilizing DAMON for memory tiering usually requires manual tuning and/or
tedious controls. Let it self-tune hotness and coldness thresholds for
promotion and demotion aiming high utilization of high memory tiers, by
introducing new DAMOS quota goal metrics representing the used and the
free memory ratios of specific NUMA nodes. And introduce a sample DAMON
module that demonstrates how the new feature can be used for memory
tiering use cases.
Backgrounds
===========
A type of tiered memory system exposes the memory tiers as NUMA nodes. A
straightforward pages placement strategy for such systems is placing
access-hot and cold pages on upper and lower tiers, reespectively,
pursuing higher utilization of upper tiers. Since access temperature can
be dynamic, periodically finding and migrating hot pages and cold pages to
proper tiers (promoting and demoting) is also required. Linux kernel
provides several features for such dynamic and transparent pages
placement.
Page Faults and LRU
-------------------
One widely known way is using NUMA balancing in tiering mode (a.k.a
NUMAB-2) and reclaim-based demotion features. In the setup, NUMAB-2 finds
hot pages using access check-purpose page faults (a.k.a prot_none) and
promote those inside each process' context, until there is no more pages
to promote, or the upper tier is filled up and memory pressure happens.
In the latter case, LRU-based reclaim logic wakes up as a response to the
memory pressure and demotes cold pages to lower tiers in asynchronous
(kswapd) and/or synchronous ways (direct reclaim).
DAMON
-----
Yet another available solution is using DAMOS with migrate_hot and
migrate_cold DAMOS actions for promotions and demotions, respectively. To
make it optimum, users need to specify aggressiveness and access
temperature thresholds for promotions and demotions in a good balance that
results in high utilization of upper tiers. The number of parameters is
not small, and optimum parameter values depend on characteristics of the
underlying hardware and the workload. As a result, it often requires
manual, time consuming and repetitive tuning of the DAMOS schemes for
given workloads and systems combinations.
Self-tuned DAMON-based Memory Tiering
=====================================
To solve such manual tuning problems, DAMOS provides aim-oriented
feedback-driven quotas self-tuning. Using the feature, we design a
self-tuned DAMON-based memory tiering for general multi-tier memory
systems.
For each memory tier node, if it has a lower tier, run a DAMOS scheme that
demotes cold pages of the node, auto-tuning the aggressiveness aiming an
amount of free space of the node. The free space is for keeping the
headroom that avoids significant memory pressure during upper tier memory
usage spike, and promoting hot pages from the lower tier.
For each memory tier node, if it has an upper tier, run a DAMOS scheme
that promotes hot pages of the current node to the upper tier, auto-tuning
the aggressiveness aiming a high utilization ratio of the upper tier. The
target ratio is to ensure higher tiers are utilized as much as possible.
It should match with the headroom for demotion scheme, but have slight
overlap, to ensure promotion and demotion are not entirely stopped.
The aim-oriented aggressiveness auto-tuning of DAMOS is already available.
Hence, to make such tiering solution implementation, only new quota goal
metrics for utilization and free space ratio of specific NUMA node need to
be developed.
Discussions
===========
The design imposes below discussion points.
Expected Behaviors
------------------
The system will let upper tier memory node accommodates as many hot data
as possible. If total amount of the data is less than the top tier
memory's promotion/demotion target utilization, entire data will be just
placed on the top tier. Promotion scheme will do nothing since there is
no data to promote. Demotion scheme will also do nothing since the free
space ratio of the top tier is higher than the goal.
Only if the amount of data is larger than the top tier's utilization
ratio, demotion scheme will demote cold pages and ensure the headroom free
space. Since the promotion and demotion schemes for a single node has
small overlap at their target utilization and free space goals, promotions
and demotions will continue working with a moderate aggressiveness level.
It will keep all data is placed on access hotness under dynamic access
pattern, while minimizing the migration overhead.
In any case, each node will keep headroom free space and as many upper
tiers are utilized as possible.
Ease of Use
-----------
Users still need to set the target utilization and free space ratio, but
it will be easier to set. We argue 99.7 % utilization and 0.5 % free
space ratios can be good default values. It can be easily adjusted based
on desired headroom size of given use case. Users are also still required
to answer the minimum coldness and hotness thresholds. Together with
monitoring intervals auto-tuning[2], DAMON will always show meaningful
amount of hot and cold memory. And DAMOS quota's prioritization mechanism
will make good decision as long as the source information is that
colorful. Hence, users can very naively set the minimum criterias. We
believe any access observation and no access observation within last one
aggregation interval is enough for minimum hot and cold regions criterias.
General Tiered Memory Setup Applicability
-----------------------------------------
The design can be applied to any number of tiers having any performance
characteristics, as long as they can be hierarchical. Hence, applying the
system to different tiered memory system will be straightforward. Note
that this assumes only single CPU NUMA node case. Because today's DAMON
is not aware of which CPU made each access, applying this on systems
having multiple CPU NUMA nodes can be complicated. We are planning to
extend DAMON for the use case, but that's out of the scope of this patch
series.
How To Use
----------
Users can implement the auto-tuned DAMON-based memory tiering using DAMON
sysfs interface. It can be easily done using DAMON user-space tool like
user-space tool. Below evaluation results section shows an example DAMON
user-space tool command for that.
For wider and simpler deployment, having a kernel module that sets up and
run the DAMOS schemes via DAMON kernel API can be useful. The module can
enable the memory tiering at boot time via kernel command line parameter
or at run time with single command. This patch series implements a sample
DAMON kernel module that shows how such module can be implemented.
Comparison To Page Faults and LRU-based Approaches
--------------------------------------------------
The existing page faults based promotion (NUMAB-2) does hot pages
detection and migration in the process context. When there are many pages
to promote, it can block the progress of the application's real works.
DAMOS works in asynchronous worker thread, so it doesn't block the real
works.
NUMAB-2 doesn't provide a way to control aggressiveness of promotion other
than the maximum amount of pages to promote per given time widnow. If hot
pages are found, promotions can happen in the upper-bound speed,
regardless of upper tier's memory pressure. If the maximum speed is not
well set for the given workload, it can result in slow promotion or
unnecessary memory pressure. Self-tuned DAMON-based memory tiering
alleviates the problem by adjusting the speed based on current utilization
of the upper tier.
LRU-based demotion can be triggered in both asynchronous (kswapd) and
synchronous (direct reclaim) ways. Other than the way of finding cold
pages, asynchronous LRU-based demotion and DAMON-based demotion has no big
difference. DAMON-based demotion can make a better balancing with
DAMON-based promotion, though. The LRU-based demotion can do better than
DAMON-based demotion when the tier is having significant memory pressure.
It would be wise to use DAMON-based demotion as a proactive and primary
one, but utilizing LRU-based demotions together as a fast backup solution.
Evaluation
==========
In short, under a setup that requires fast and frequent promotions,
self-tuned DAMON-based memory tiering's hot pages promotion improves
performance about 4.42 %. We believe this shows self-tuned DAMON-based
promotion's effectiveness. Meanwhile, NUMAB-2's hot pages promotion
degrades the performance about 7.34 %. We suspect the degradation is
mostly due to NUMAB-2's synchronous nature that can block the
application's progress, which highlights the advantage of DAMON-based
solution's asynchronous nature.
Note that the test was done with the RFC version of this patch series. We
don't run it again since this patch series got no meaningful change after
the RFC, while the test takes pretty long time.
Setup
-----
Hardware. Use a machine that equips 250 GiB DRAM memory tier and 50 GiB
CXL memory tier. The tiers are exposed as NUMA nodes 0 and 1,
respectively.
Kernel. Use Linux kernel v6.13 that modified as following. Add all DAMON
patches that available on mm tree of 2025-03-15, and this patch series.
Also modify it to ignore mempolicy() system calls, to avoid bad effects
from application's traditional NUMA systems assumed optimizations.
Workload. Use a modified version of Taobench benchmark[3] that available
on DCPerf benchmark suite. It represents an in-memory caching workload.
We set its 'memsize', 'warmup_time', and 'test_time' parameter as 340 GiB,
2,500 seconds and 1,440 seconds. The parameters are chosen to ensure the
workload uses more than DRAM memory tier. Its RSS under the parameter
grows to 270 GiB within the warmup time.
It turned out the workload has a very static access pattrn. Only about 13
% of the RSS is frequently accessed from the beginning to end. Hence
promotion shows no meaningful performance difference regardless of
different design and implementations. We therefore modify the kernel to
periodically demote up to 10 GiB hot pages and promote up to 10 GiB cold
pages once per minute. The intention is to simulate periodic access
pattern changes. The hotness and coldness threshold is very naively set
so that it is more like random access pattern change rather than strict
hot/cold pages exchange. This is why we call the workload as "modified".
It is implemented as two DAMOS schemes each running on an asynchronous
thread. It can be reproduced with DAMON user-space tool like below.
# ./damo start \
--ops paddr --numa_node 0 --monitoring_intervals 10s 200s 200s \
--damos_action migrate_hot 1 \
--damos_quota_interval 60s --damos_quota_space 10G \
--ops paddr --numa_node 1 --monitoring_intervals 10s 200s 200s \
--damos_action migrate_cold 0 \
--damos_quota_interval 60s --damos_quota_space 10G \
--nr_schemes 1 1 --nr_targets 1 1 --nr_ctxs 1 1
System configurations. Use below variant system configurations.
- Baseline. No memory tiering features are turned on.
- Numab_tiering. On the baseline, enable NUMAB-2 and relcaim-based
demotion. In detail, following command is executed:
echo 2 > /proc/sys/kernel/numa_balancing;
echo 1 > /sys/kernel/mm/numa/demotion_enabled;
echo 7 > /proc/sys/vm/zone_reclaim_mode
- DAMON_tiering. On the baseline, utilize self-tuned DAMON-based memory
tiering implementation via DAMON user-space tool. It utilizes two
kernel threads, namely promotion thread and demotion thread. Demotion
thread monitors access pattern of DRAM node using DAMON with
auto-tuned monitoring intervals aiming 4% DAMON-observed access ratio,
and demote coldest pages up to 200 MiB per second aiming 0.5% free
space of DRAM node. Promotion thread monitors CXL node using same
intervals auto-tuning, and promote hot pages in same way but aiming
for 99.7% utilization of DRAM node. Because DAMON provides only
best-effort accuracy, add young page DAMOS filters to allow only and
reject all young pages at promoting and demoting, respectively. It
can be reproduced with DAMON user-space tool like below.
# ./damo start \
--numa_node 0 --monitoring_intervals_goal 4% 3 5ms 10s \
--damos_action migrate_cold 1 --damos_access_rate 0% 0% \
--damos_apply_interval 1s \
--damos_quota_interval 1s --damos_quota_space 200MB \
--damos_quota_goal node_mem_free_bp 0.5% 0 \
--damos_filter reject young \
--numa_node 1 --monitoring_intervals_goal 4% 3 5ms 10s \
--damos_action migrate_hot 0 --damos_access_rate 5% max \
--damos_apply_interval 1s \
--damos_quota_interval 1s --damos_quota_space 200MB \
--damos_quota_goal node_mem_used_bp 99.7% 0 \
--damos_filter allow young \
--damos_nr_quota_goals 1 1 --damos_nr_filters 1 1 \
--nr_targets 1 1 --nr_schemes 1 1 --nr_ctxs 1 1
Measurment Results
------------------
On each system configuration, run the modified version of Taobench and
collect 'score'. 'score' is a metric that calculated and provided by
Taobench to represents the performance of the run on the system. To
handle the measurement errors, repeat the measurement five times. The
results are as below.
Config Score Stdev (%) Normalized
Baseline 1.6165 0.0319 1.9764 1.0000
Numab_tiering 1.4976 0.0452 3.0209 0.9264
DAMON_tiering 1.6881 0.0249 1.4767 1.0443
'Config' column shows the system config of the measurement. 'Score'
column shows the 'score' measurement in average of the five runs on the
system config. 'Stdev' column shows the standsard deviation of the five
measurements of the scores. '(%)' column shows the 'Stdev' to 'Score'
ratio in percentage. Finally, 'Normalized' column shows the averaged
score values of the configs that normalized to that of 'Baseline'.
The periodic hot pages demotion and cold pages promotion that was
conducted to simulate dynamic access pattern was started from the
beginning of the workload. It resulted in the DRAM tier utilization
always under the watermark, and hence no real demotion was happened for
all test runs. This means the above results show no difference between
LRU-based and DAMON-based demotions. Only difference between NUMAB-2 and
DAMON-based promotions are represented on the results.
Numab_tiering config degraded the performance about 7.36 %. We suspect
this happened because NUMAB-2's synchronous promotion was blocking the
Taobench's real work progress.
DAMON_tiering config improved the performance about 4.43 %. We believe
this shows effectiveness of DAMON-based promotion that didn't block
Taobench's real work progress due to its asynchronous nature. Also this
means DAMON's monitoring results are accurate enough to provide visible
amount of improvement.
Evaluation Limitations
----------------------
As mentioned above, this evaluation shows only comparison of promotion
mechanisms. DAMON-based tiering is recommended to be used together with
reclaim-based demotion as a faster backup under significant memory
pressure, though.
From some perspective, the modified version of Taobench may seems making
the picture distorted too much. It would be better to evaluate with more
realistic workload, or more finely tuned micro benchmarks.
Patch Sequence
==============
The first patch (patch 1) implements two new quota goal metrics on core
layer and expose it to DAMON core kernel API. The second and third ones
(patches 2 and 3) further link it to DAMON sysfs interface. Three
following patches (patches 4-6) document the new feature and sysfs file on
design, usage, and ABI documents. The final one (patch 7) implements a
working version of a self-tuned DAMON-based memory tiering solution in an
incomplete but easy to understand form as a kernel module under
samples/damon/ directory.
References
==========
[1] https://lore.kernel.org/
20231112195602.61525-1-sj@kernel.org/
[2] https://lore.kernel.org/
20250303221726.484227-1-sj@kernel.org
[3] https://github.com/facebookresearch/DCPerf/blob/main/packages/tao_bench/README.md
This patch (of 7):
Used and free space ratios for specific NUMA nodes can be useful inputs
for NUMA-specific DAMOS schemes' aggressiveness self-tuning feedback loop.
Implement DAMOS quota goal metrics for such self-tuned schemes.
Link: https://lkml.kernel.org/r/20250420194030.75838-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20250420194030.75838-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Yunjeong Mun <yunjeong.mun@sk.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
projects / thirdparty / linux.git / commitdiff
