A guide to DAMON tuning and results interpretation for hot pages
Table of Contents
The initial version of this post was initially posted to DAMON mailing list as https://lore.kernel.org/20241108232536.73843-1-sj@kernel.org
Posting it here too, for visibility and after-posting updates for any needs.
One of common issues that I received from DAMON users is that DAMON’s monitoring results show hot regions much less than expected. Specifically, the users find regions of only zero or low ’nr_accesses’ value from the DAMON-generated access pattern snapshots.
In some cases, it turned out the problem can be alleviated by tuning DAMON parameters or changing the way to interpret the results. I’d like to share the details and possible future improvements here.
Note that I’m not saying this is users’ tuning fault. I admit that the real root cause of the issue is the poor interface and lack of guides that makes correct tuning difficult, and the suboptimality of DAMON’s mechanisms. We will continue working on advancing it in long term. Sharing some of the plans and status at the end of this email.
TL; DR
Users show only low or zero nr_accesses regions mainly because they set ‘aggregation intrval’ too short compared to the workload’s memory access intensiveness. Please increase the aggregation interval, or treat ’nr_accesses’ zero regions of short ‘age’ as hot regions.
Now let’s walk through more details. The below sections assume you’re familiar with DAMON’s monitoring mechanisms including ‘Access Frequency Monitoring’, ‘Regions Based Sampling’, ‘Adaptive Regions Adjustment’, and ‘Age Tracking’. You should particularly be familiar with terms including ‘sampling interval’, ‘aggregation interval’, ’nr_accesses’, and ‘age’. If you’re not familiar with those, you can refer to the document (https://www.kernel.org/doc/html/next/mm/damon/design.html#monitoring).
Problem
Some users reported that they expected DAMON will be useful for finding both cold memory regions and hot memory regions. And they found it indeed works for finding cold memory regions. But, they found difficulties at finding hot memory regions. Some detailed reported cases are as below.
Case 1: Proactive Reclamation
A user tried to use DAMOS for proactively reclaiming cold memory regions. They hence specified maximum access rate (nr_accesses) and minimum age of target regions for the DAMOS scheme as zero and 2 minutes, respectively. It means asking DAMON to find regions that not accessed for two minutes or more and reclaim those as soon as found. If they use DAMON user-space tool, damo, they would used a command like below.
# damo start --damos_action pageout --damos_access rate 0% 0% --damos_age 2m max
The result was as the user expected. The user reported me that they were able to save memory without performance degradation.
My test setup also showed good results from similar DAMOS schemes, even though my setup was using even more aggressive approach (minimum age as 5 seconds).
Case 2: Heatmap Visualization
A user has recorded data access pattern of their workloads using DAMON user-space tool, damo, and draw heatmap using ‘damo report heatmap’ command. The workload was active and therefore the user expected the heatmap to show some heats. But the output was showing nearly zero heats. It was nearly always only dark.
On my test setup, some workloads indeed showed very dark heatmap. But on multiple workloads, meaningful access patterns were identifiable using the heatmaps (https://damonitor.github.io/test/result/visual/next/rec.heatmap.1.png.html).
){kind=link}
Case 3: Prioritizing Hot Pages
A user tried to use DAMOS for finding hot memory regions and make a prioritization action like backing the regions with THP, migrating the pages from CXL node to DRAM node, or moving the page from inactive LRU list to active LRU list. They hence specified minimum access rate (nr_accesses) for the DAMOS scheme with a high value. For example, a command like below was used:
# damo start --damos_action hugepage --damos_access_rate 50% max $(pidof workload)
However, they found DAMOS is doing nearly nothing. They reduced the minimum access_rate, and situation has been better, but still DAMOS was not finding expected amount of hot memory regions.
Case 2 and 3 mean that they expected to show regions of high ’nr_accesses’ value from DAMON-generated access pattern snapshots. But it showed only zero or low ’nr_accesses’ values.
On my test setup (https://damonitor.github.io/posts/damon_evaluation/), similar approach showed good results, though.
One common thing that I found from the reports was that they were using default values for ‘aggregation interval’, which is 100 milliseconds. My test setup is also using the default values.
Meaning of ‘aggregation interval’
For every ‘aggregation interval’, DAMON generates one complete access pattern snapshot. That means ‘aggregation interval’ can be though of as the time amount of information that a single snapshot can contain. The value should hence be decided depending on how much information the user wants each snapshot to have. If it is too short, each snapshot contains nearly no information, so not be useful. If it is too long, each snapshot contains too much information that coarsely cumulated, so again not useful.
The meaningful length of time depends on the data access intensiveness of the workload. The intensiveness should be measured with two factors: frequency and footprint.
The frequency is how frequently the workload is making accesses. If the workload makes zero or only a low number of accesses per given ‘aggregation interval’, the snapshot will of course show zero or low number of accesses, due to ‘Access Frequency Monitoring’ mechanism.
The footprint is the total amount of memory that the workload has accessed at least once. Due to ‘Region Based Sampling’ mechanism, it should be meaningfully large compared to the size of total monitoring target regions. For example, if DAMON is requested to monitor 1 TiB memory and the workload is accessing 1 MiB region of it, DAMON’s sampling based approach will have difficulty at finding the 1 MiB region. ‘Adaptive Regions Adjustment’ mechanism will make DAMON to find the 1 MiB region eventually. But it will take time. The workload could stop accessing the 1 MiB region and starts accessing another 1 MiB region before the ‘Adaptive Regions Adjustment’ mechanism finds it.
Note that the frequency and footprint of accesses from workloads would depend on not only the source code but also many factors. The system’s total memory bandwidth, the extent of load and noisy neighbor workloads could be a few examples.
So, if the desired maximum ’nr_accesses’ on each snapshot is fixed, ‘aggregation interval’ should be increased as the access aggressiveness is decreased, and vice versa.
Root Causes 1: Suboptimal Aggregation Interval Tuning
I suspect in many of the reported problematic cases, the ‘aggregation interval’ was too short. The default value is good enough for my test setup, but it would need tuning on different systems. Especially on large systems having limited bandwidth, 100 millisecond ‘aggregation interval’ could be not long enough to make meaningful amount of accesses in terms of both frequency and footprint within the interval.
I actually suggested some of the reporters to increase ‘aggregation interval’, and it was helpful at alleviating the issue in some degree. I cannot know exactly who are using DAMON with what parameters. But at least one open-sourced usage from HMSDK is setting ‘aggregation interval’ as two seconds (https://github.com/skhynix/hmsdk/blob/hmsdk-v3.0/tools/gen_migpol.py#L14).
Root Cause 2: Ignorance of Recency (‘age’)
The ’nr_accesses’ represents the access frequency, and hence it is natural to assume hot memory regions would have high ’nr_accesses’.
DAMON provides not only frequency. It also informs users how long the ’nr_accesses’ of the regions is maintained, namely ‘age’, using ‘Age Tracking’ mechanism. If ’nr_accesses’ is non-zero, ‘age’ can be used to calculate actual access hotness of the region (if we turn fire on for longer time, the temperature will be higher). If ’nr_accesses’ is zero, ‘age’ can represent a sort of the recency information (when the regions has last accessed).
The recency information can be useful for finding cold pages like case 1 (proactive reclaim). The opposite of finding cold pages is finding hot pages, so the recency information can also be used for finding relatively hot pages. In other words, if for any reason DAMON cannot generate a snapshot having enough non-zero ’nr_accesses’ divergence for given purpose, users could further differentiate hot regions among zero ’nr_accesses’ regions, using ‘age’ information. It would be not ideal, but still reasonable like a sort of LRU-based appraoches could be.
So I suspect some of the problems occur from not using ‘age’ for zero ’nr_accesses’ regions. Actually reports from case 3 (prioritizing hot pages), which were successful only on my test setup, were commonly using non-zero minimum ’nr_accesses’ for DAMOS schemes, so were ignoring ‘age’ of zero ’nr_accesses’ regions. Meanwhile, case 2 (proactive reclaim) was using ‘age’ information for zero ’nr_accesses’ regions, and no negative results have reported so far.
This might seem like not addressing case 2 (heatmap visualization), because heatmap shows ’nr_accesses’ change of regions over time. But if the record is collected for long time, regions that shown non-zero ’nr_accesses’ for a short period may look a very small dot on the low-resolution picture that not easy to shown with human eyes. The users might be able to get different results using ‘age’ information on the collected snapshot, like recency or temperation based histogram (https://github.com/damonitor/damo/blob/v2.5.4/USAGE.md#access-report-styles).
Tuning Guide
Based on above root cause theories, I suggest to try below tuning guides.
If you show DAMON is not working well at finding hot pages,
- Ensure your workload is making meaningfully intensive data accesses.
- Gradually increase aggregation interval and show if it makes change.
- Try using ‘age’ information even if ’nr_accesses’ is zero.
- If nothing works, report the problem to sj@kernel.org, damon@lists.linux.dev and/or linux-mm@kvack.org.
If increasing aggregation interval alleviates your problem, you can further consider increasing ‘sampling interval’. If it doesn’t harm the quality of the access pattern snapshots, having low ‘sampling interval’ will only increase DAMON’s CPU usage.
For using ‘age’ information of zero ’nr_accesses’ regions, different approaches could be used for profiling use case and DAMOS use case. For profiling use case, users can try reading recency or access temperature based histograms (https://github.com/damonitor/damo/blob/v2.5.4/USAGE.md#access-report-styles) of snapshots from record, or live-captured snapshots.
If the use case is for DAMOS, applying the ‘age’ information on DAMOS target access pattern would be straightforward. Using DAMOS Quotas together can be useful, since it provides its own under-quota-prioritization logic that utilizes ‘age’ information for zero ’nr_accesses’ regions, and further provides auto-tuning of the quota for given target metric/value.
Tuning Example
The initial version of this section is also posted to the mailing list: https://lore.kernel.org/20241202175459.2005526-1-sj@kernel.org
I tried monitoring the access patterns on the physical address space of a system running a real-world server workload. I was able to reproduce the reported poor quality of hot pages detection using default parameter. And I was also able to improve the quality following the above tuning guide. I’m sharing the details as an example.
5ms/100ms intervals: Reproduce Problem
Initially, I captured the access pattern snapshot on the physical address space using DAMON, with the default interval parameters (5 milliseconds and 100 milliseconds for the sampling and the aggregation intervals, respectively). I wait ten minutes after starting DAMON, to show a meaningful time-wise access patterns.
# damo start
# sleep 600
# damo record --snapshot 0 1
# damo stop
Then, I listed the DAMON-found regions of different access patterns, sorted by the access temperature. Access temperature is calculated as a weighted sum of the access frequency and the age of the region. If the access frequency is 0 %, the temperature is multipled by minus one. That is, if a region is not accessed, it gets minus temperature and it gets lower as not accessed for longer time. The sorting is in temperature-ascendint order, so the region at the top of the list is the coldest, and the one at the bottom is the hottest one.
# damo report access --sort_regions_by temperature
0 addr 16.052 GiB size 5.985 GiB access 0 % age 5.900 s # coldest
1 addr 22.037 GiB size 6.029 GiB access 0 % age 5.300 s
2 addr 28.065 GiB size 6.045 GiB access 0 % age 5.200 s
3 addr 10.069 GiB size 5.983 GiB access 0 % age 4.500 s
4 addr 4.000 GiB size 6.069 GiB access 0 % age 4.400 s
5 addr 62.008 GiB size 3.992 GiB access 0 % age 3.700 s
6 addr 56.795 GiB size 5.213 GiB access 0 % age 3.300 s
7 addr 39.393 GiB size 6.096 GiB access 0 % age 2.800 s
8 addr 50.782 GiB size 6.012 GiB access 0 % age 2.800 s
9 addr 34.111 GiB size 5.282 GiB access 0 % age 2.300 s
10 addr 45.489 GiB size 5.293 GiB access 0 % age 1.800 s # hottest
total size: 62.000 GiB
The list shows not seemingly hot regions, and only minimum access pattern
diversity. Every region has zero access frequency. The number of region is
10, which is the default min_nr_regions value. Size of each region is also
nearly idential. We can suspect this is because “adaptive regions adjustment”
mechanism was not well working. As the guide suggested, we can get relative
hotness of regions using ‘age’ as the recency information. That would be
better than nothing, but given the fact that the longest age is only about 6
seconds while we waited about ten minuts, it is unclear how useful this will
be.
The temperature ranges to total size of regions of each range histogram visualization of the results also shows no interesting distribution pattern.
# damo report access --style temperature-sz-hist
<temperature> <total size>
[-,590,000,000, -,549,000,000) 5.985 GiB |********** |
[-,549,000,000, -,508,000,000) 12.074 GiB |********************|
[-,508,000,000, -,467,000,000) 0 B | |
[-,467,000,000, -,426,000,000) 12.052 GiB |********************|
[-,426,000,000, -,385,000,000) 0 B | |
[-,385,000,000, -,344,000,000) 3.992 GiB |******* |
[-,344,000,000, -,303,000,000) 5.213 GiB |********* |
[-,303,000,000, -,262,000,000) 12.109 GiB |********************|
[-,262,000,000, -,221,000,000) 5.282 GiB |********* |
[-,221,000,000, -,180,000,000) 0 B | |
[-,180,000,000, -,139,000,000) 5.293 GiB |********* |
total size: 62.000 GiB
In short, the result is very similar to the reported problems: poor quality monitoring results for hot regions detection. According to the above guide, this is due to the too short aggregation interval.
100ms/2s intervals: Starts Showing Small Hot Regions
Following the guide, I increased the interval 20 times (100 milliseocnds and 2 seconds for sampling and aggregation intervals, respectively).
# damo start -s 100ms -a 2s
# sleep 600
# damo record --snapshot 0 1
# damo stop
# damo report access --sort_regions_by temperature
0 addr 10.180 GiB size 6.117 GiB access 0 % age 7 m 8 s # coldest
1 addr 49.275 GiB size 6.195 GiB access 0 % age 6 m 14 s
2 addr 62.421 GiB size 3.579 GiB access 0 % age 6 m 4 s
3 addr 40.154 GiB size 6.127 GiB access 0 % age 5 m 40 s
4 addr 16.296 GiB size 6.182 GiB access 0 % age 5 m 32 s
5 addr 34.254 GiB size 5.899 GiB access 0 % age 5 m 24 s
6 addr 46.281 GiB size 2.995 GiB access 0 % age 5 m 20 s
7 addr 28.420 GiB size 5.835 GiB access 0 % age 5 m 6 s
8 addr 4.000 GiB size 6.180 GiB access 0 % age 4 m 16 s
9 addr 22.478 GiB size 5.942 GiB access 0 % age 3 m 58 s
10 addr 55.470 GiB size 915.645 MiB access 0 % age 3 m 6 s
11 addr 56.364 GiB size 6.056 GiB access 0 % age 2 m 8 s
12 addr 56.364 GiB size 4.000 KiB access 95 % age 16 s
13 addr 49.275 GiB size 4.000 KiB access 100 % age 8 m 24 s # hottest
total size: 62.000 GiB
# damo report access --style temperature-sz-hist
<temperature> <total size>
[-42,800,000,000, -33,479,999,000) 22.018 GiB |***************** |
[-33,479,999,000, -24,159,998,000) 27.090 GiB |********************|
[-24,159,998,000, -14,839,997,000) 6.836 GiB |****** |
[-14,839,997,000, -5,519,996,000) 6.056 GiB |***** |
[-5,519,996,000, 3,800,005,000) 4.000 KiB |* |
[3,800,005,000, 13,120,006,000) 0 B | |
[13,120,006,000, 22,440,007,000) 0 B | |
[22,440,007,000, 31,760,008,000) 0 B | |
[31,760,008,000, 41,080,009,000) 0 B | |
[41,080,009,000, 50,400,010,000) 0 B | |
[50,400,010,000, 59,720,011,000) 4.000 KiB |* |
total size: 62.000 GiB
DAMON found two distinct 4 KiB regions that pretty hot. The regions are also well aged. The hottest 4 KiB region was keeping the access frequency for about 8 minutes, and the coldest region was keeping no access for about 7 minutes. The distribution on the histogram also looks like having a pattern.
Especially, the finding of the 4 KiB regions shows DAMON’s adaptive regions adjustment is working as designed.
Still the number of regions is close to the min_nr_regions, and sizes of cold
regions are similar, though. Apparently it is improved, but it still has rooms
to improve.
400ms/8s intervals: Pretty Improved Results
I further increased the intervals four times (400 milliseconds and 8 seconds for sampling and aggregation intervals, respectively).
# damo start -s 400ms -a 8s
# sleep 600
# damo record --snapshot 0 1
# damo stop
# damo report access --sort_regions_by temperature
0 addr 64.492 GiB size 1.508 GiB access 0 % age 6 m 48 s # coldest
1 addr 21.749 GiB size 5.674 GiB access 0 % age 6 m 8 s
2 addr 27.422 GiB size 5.801 GiB access 0 % age 6 m
3 addr 49.431 GiB size 8.675 GiB access 0 % age 5 m 28 s
4 addr 33.223 GiB size 5.645 GiB access 0 % age 5 m 12 s
5 addr 58.321 GiB size 6.170 GiB access 0 % age 5 m 4 s
[...]
25 addr 6.615 GiB size 297.531 MiB access 15 % age 0 ns
26 addr 9.513 GiB size 12.000 KiB access 20 % age 0 ns
27 addr 9.511 GiB size 108.000 KiB access 25 % age 0 ns
28 addr 9.513 GiB size 20.000 KiB access 25 % age 0 ns
29 addr 9.511 GiB size 12.000 KiB access 30 % age 0 ns
30 addr 9.520 GiB size 4.000 KiB access 40 % age 0 ns
[...]
41 addr 9.520 GiB size 4.000 KiB access 80 % age 56 s
42 addr 9.511 GiB size 12.000 KiB access 100 % age 6 m 16 s
43 addr 58.321 GiB size 4.000 KiB access 100 % age 6 m 24 s
44 addr 9.512 GiB size 4.000 KiB access 100 % age 6 m 48 s
45 addr 58.106 GiB size 4.000 KiB access 100 % age 6 m 48 s # hottest
total size: 62.000 GiB
# damo report access --style temperature-sz-hist
<temperature> <total size>
[-40,800,000,000, -32,639,999,000) 21.657 GiB |********************|
[-32,639,999,000, -24,479,998,000) 17.938 GiB |***************** |
[-24,479,998,000, -16,319,997,000) 16.885 GiB |**************** |
[-16,319,997,000, -8,159,996,000) 586.879 MiB |* |
[-8,159,996,000, 5,000) 4.946 GiB |***** |
[5,000, 8,160,006,000) 260.000 KiB |* |
[8,160,006,000, 16,320,007,000) 0 B | |
[16,320,007,000, 24,480,008,000) 0 B | |
[24,480,008,000, 32,640,009,000) 0 B | |
[32,640,009,000, 40,800,010,000) 16.000 KiB |* |
[40,800,010,000, 48,960,011,000) 8.000 KiB |* |
total size: 62.000 GiB
The number of regions having different access patterns has significantly increased. Size of each region is also more varied. Total size of non-zero access frequency regions is also significantly increased. Maybe this is already good enough to make some meaningful memory management efficieny changes.
800ms/16s intervals: Another bias
Further doubling the intervals (800 milliseconds and 16 seconds for sampling and aggregation intervals, respectively) mor improved the hot regions detection, but starts looking degrading cold regions detection.
# damo start -s 800ms -a 16s
# sleep 600
# damo record --snapshot 0 1
# damo stop
# damo report access --sort_regions_by temperature
0 addr 64.781 GiB size 1.219 GiB access 0 % age 4 m 48 s
1 addr 24.505 GiB size 2.475 GiB access 0 % age 4 m 16 s
2 addr 26.980 GiB size 504.273 MiB access 0 % age 4 m
3 addr 29.443 GiB size 2.462 GiB access 0 % age 4 m
4 addr 37.264 GiB size 5.645 GiB access 0 % age 4 m
5 addr 31.905 GiB size 5.359 GiB access 0 % age 3 m 44 s
[...]
20 addr 8.711 GiB size 40.000 KiB access 5 % age 2 m 40 s
21 addr 27.473 GiB size 1.970 GiB access 5 % age 4 m
22 addr 48.185 GiB size 4.625 GiB access 5 % age 4 m
23 addr 47.304 GiB size 902.117 MiB access 10 % age 4 m
24 addr 8.711 GiB size 4.000 KiB access 100 % age 4 m
25 addr 20.793 GiB size 3.713 GiB access 5 % age 4 m 16 s
26 addr 8.773 GiB size 4.000 KiB access 100 % age 4 m 16 s
total size: 62.000 GiB
# damo report access --style temperature-sz-hist
<temperature> <total size>
[-28,800,000,000, -23,359,999,000) 12.294 GiB |***************** |
[-23,359,999,000, -17,919,998,000) 9.753 GiB |************* |
[-17,919,998,000, -12,479,997,000) 15.131 GiB |********************|
[-12,479,997,000, -7,039,996,000) 0 B | |
[-7,039,996,000, -1,599,995,000) 7.506 GiB |********** |
[-1,599,995,000, 3,840,006,000) 6.127 GiB |********* |
[3,840,006,000, 9,280,007,000) 0 B | |
[9,280,007,000, 14,720,008,000) 136.000 KiB |* |
[14,720,008,000, 20,160,009,000) 40.000 KiB |* |
[20,160,009,000, 25,600,010,000) 11.188 GiB |*************** |
[25,600,010,000, 31,040,011,000) 4.000 KiB |* |
total size: 62.000 GiB
It found more non-zero access frequency regions. The number of regions is
still much higher than the min_nr_regions, but it is reduced from that of the
previous setup. And apparently the distribution seems bit biased to hot
regions.
Conclusion
Because the workload is live, the above results are not always consistent. But, the tendency of the quality for the interval changes was consistent.
With the above experimental tuning results, I conclude the theory and the guide makes sense to at least this workload, and could be applied to similar cases.
This also gives us an idea for automated tuning of the intervals. If the interval is too short, results are biased to cold regions. If the interval is too long, results are biased to hot regions. Maybe DAMON can moitor to which direction the current snapshot is biased, and adjust the intervals. I will develop the idea more.
Future Plans
Again, I admit that the real root cause of the issue is the poor interface and lack of guides that makes correct tuning difficult, and the suboptimality of DAMON’s mechanisms. We will continue working on advancing it in long term.
For easy use of ‘age’ information for zero ’nr_accesses’ regions, DAMON is already providing its Quotas feature with auto-tuning. We will continue making the auto-tuning more advanced, and adding new features for ease of uses.
One obvious hidden root cause is the absence of the guideline. I will collect more inputs and write an official document for this. I actually thought about writing the official document first, but this writing this mail took already over two weeks. So posting this mail first.
For interval setting, we are planning a sort of auto-tuning. Like DAMOS quota auto-tuning, users will be able to set more intuitive knobs (e.g., how much diversity of regions they want) instead of directly setting the intervals. Then, DAMON will automatically tune the intervals and other low level knobs. This is in very early stage, so no specific design is made so far, and will take long time. Don’t expect delivery of this in near future. Use the tuning guide for short term and/or ask prioritization of this project if needed.
In my humble opinion, ‘Adaptive Regions Adjustment’ mechanism is not a root cause of the issue. Nonetheless, it also has many rooms for improvement that can make DAMON more lightweight and accurate. And any DAMON accuracy improvements would alleviate the hot page detection issue. We plan to do this together, and this is again a long term project that has no specific design yet. Nonetheless, we recently shared two simple short term features for this (https://lore.kernel.org/20241027204910.155254-1-sj@kernel.org). If you are interested in implementing the short term features, please step up.