xref: /linux/Documentation/mm/damon/design.rst (revision 71dfa617ea9f18e4585fe78364217cd32b1fc382)
1.. SPDX-License-Identifier: GPL-2.0
2
3======
4Design
5======
6
7
8.. _damon_design_execution_model_and_data_structures:
9
10Execution Model and Data Structures
11===================================
12
13The monitoring-related information including the monitoring request
14specification and DAMON-based operation schemes are stored in a data structure
15called DAMON ``context``.  DAMON executes each context with a kernel thread
16called ``kdamond``.  Multiple kdamonds could run in parallel, for different
17types of monitoring.
18
19
20Overall Architecture
21====================
22
23DAMON subsystem is configured with three layers including
24
25- Operations Set: Implements fundamental operations for DAMON that depends on
26  the given monitoring target address-space and available set of
27  software/hardware primitives,
28- Core: Implements core logics including monitoring overhead/accurach control
29  and access-aware system operations on top of the operations set layer, and
30- Modules: Implements kernel modules for various purposes that provides
31  interfaces for the user space, on top of the core layer.
32
33
34.. _damon_design_configurable_operations_set:
35
36Configurable Operations Set
37---------------------------
38
39For data access monitoring and additional low level work, DAMON needs a set of
40implementations for specific operations that are dependent on and optimized for
41the given target address space.  On the other hand, the accuracy and overhead
42tradeoff mechanism, which is the core logic of DAMON, is in the pure logic
43space.  DAMON separates the two parts in different layers, namely DAMON
44Operations Set and DAMON Core Logics Layers, respectively.  It further defines
45the interface between the layers to allow various operations sets to be
46configured with the core logic.
47
48Due to this design, users can extend DAMON for any address space by configuring
49the core logic to use the appropriate operations set.  If any appropriate set
50is unavailable, users can implement one on their own.
51
52For example, physical memory, virtual memory, swap space, those for specific
53processes, NUMA nodes, files, and backing memory devices would be supportable.
54Also, if some architectures or devices supporting special optimized access
55check primitives, those will be easily configurable.
56
57
58Programmable Modules
59--------------------
60
61Core layer of DAMON is implemented as a framework, and exposes its application
62programming interface to all kernel space components such as subsystems and
63modules.  For common use cases of DAMON, DAMON subsystem provides kernel
64modules that built on top of the core layer using the API, which can be easily
65used by the user space end users.
66
67
68.. _damon_operations_set:
69
70Operations Set Layer
71====================
72
73The monitoring operations are defined in two parts:
74
751. Identification of the monitoring target address range for the address space.
762. Access check of specific address range in the target space.
77
78DAMON currently provides below three operation sets.  Below two subsections
79describe how those work.
80
81 - vaddr: Monitor virtual address spaces of specific processes
82 - fvaddr: Monitor fixed virtual address ranges
83 - paddr: Monitor the physical address space of the system
84
85
86 .. _damon_design_vaddr_target_regions_construction:
87
88VMA-based Target Address Range Construction
89-------------------------------------------
90
91A mechanism of ``vaddr`` DAMON operations set that automatically initializes
92and updates the monitoring target address regions so that entire memory
93mappings of the target processes can be covered.
94
95This mechanism is only for the ``vaddr`` operations set.  In cases of
96``fvaddr`` and ``paddr`` operation sets, users are asked to manually set the
97monitoring target address ranges.
98
99Only small parts in the super-huge virtual address space of the processes are
100mapped to the physical memory and accessed.  Thus, tracking the unmapped
101address regions is just wasteful.  However, because DAMON can deal with some
102level of noise using the adaptive regions adjustment mechanism, tracking every
103mapping is not strictly required but could even incur a high overhead in some
104cases.  That said, too huge unmapped areas inside the monitoring target should
105be removed to not take the time for the adaptive mechanism.
106
107For the reason, this implementation converts the complex mappings to three
108distinct regions that cover every mapped area of the address space.  The two
109gaps between the three regions are the two biggest unmapped areas in the given
110address space.  The two biggest unmapped areas would be the gap between the
111heap and the uppermost mmap()-ed region, and the gap between the lowermost
112mmap()-ed region and the stack in most of the cases.  Because these gaps are
113exceptionally huge in usual address spaces, excluding these will be sufficient
114to make a reasonable trade-off.  Below shows this in detail::
115
116    <heap>
117    <BIG UNMAPPED REGION 1>
118    <uppermost mmap()-ed region>
119    (small mmap()-ed regions and munmap()-ed regions)
120    <lowermost mmap()-ed region>
121    <BIG UNMAPPED REGION 2>
122    <stack>
123
124
125PTE Accessed-bit Based Access Check
126-----------------------------------
127
128Both of the implementations for physical and virtual address spaces use PTE
129Accessed-bit for basic access checks.  Only one difference is the way of
130finding the relevant PTE Accessed bit(s) from the address.  While the
131implementation for the virtual address walks the page table for the target task
132of the address, the implementation for the physical address walks every page
133table having a mapping to the address.  In this way, the implementations find
134and clear the bit(s) for next sampling target address and checks whether the
135bit(s) set again after one sampling period.  This could disturb other kernel
136subsystems using the Accessed bits, namely Idle page tracking and the reclaim
137logic.  DAMON does nothing to avoid disturbing Idle page tracking, so handling
138the interference is the responsibility of sysadmins.  However, it solves the
139conflict with the reclaim logic using ``PG_idle`` and ``PG_young`` page flags,
140as Idle page tracking does.
141
142
143Core Logics
144===========
145
146
147Monitoring
148----------
149
150Below four sections describe each of the DAMON core mechanisms and the five
151monitoring attributes, ``sampling interval``, ``aggregation interval``,
152``update interval``, ``minimum number of regions``, and ``maximum number of
153regions``.
154
155
156Access Frequency Monitoring
157~~~~~~~~~~~~~~~~~~~~~~~~~~~
158
159The output of DAMON says what pages are how frequently accessed for a given
160duration.  The resolution of the access frequency is controlled by setting
161``sampling interval`` and ``aggregation interval``.  In detail, DAMON checks
162access to each page per ``sampling interval`` and aggregates the results.  In
163other words, counts the number of the accesses to each page.  After each
164``aggregation interval`` passes, DAMON calls callback functions that previously
165registered by users so that users can read the aggregated results and then
166clears the results.  This can be described in below simple pseudo-code::
167
168    while monitoring_on:
169        for page in monitoring_target:
170            if accessed(page):
171                nr_accesses[page] += 1
172        if time() % aggregation_interval == 0:
173            for callback in user_registered_callbacks:
174                callback(monitoring_target, nr_accesses)
175            for page in monitoring_target:
176                nr_accesses[page] = 0
177        sleep(sampling interval)
178
179The monitoring overhead of this mechanism will arbitrarily increase as the
180size of the target workload grows.
181
182
183.. _damon_design_region_based_sampling:
184
185Region Based Sampling
186~~~~~~~~~~~~~~~~~~~~~
187
188To avoid the unbounded increase of the overhead, DAMON groups adjacent pages
189that assumed to have the same access frequencies into a region.  As long as the
190assumption (pages in a region have the same access frequencies) is kept, only
191one page in the region is required to be checked.  Thus, for each ``sampling
192interval``, DAMON randomly picks one page in each region, waits for one
193``sampling interval``, checks whether the page is accessed meanwhile, and
194increases the access frequency counter of the region if so.  The counter is
195called ``nr_regions`` of the region.  Therefore, the monitoring overhead is
196controllable by setting the number of regions.  DAMON allows users to set the
197minimum and the maximum number of regions for the trade-off.
198
199This scheme, however, cannot preserve the quality of the output if the
200assumption is not guaranteed.
201
202
203Adaptive Regions Adjustment
204~~~~~~~~~~~~~~~~~~~~~~~~~~~
205
206Even somehow the initial monitoring target regions are well constructed to
207fulfill the assumption (pages in same region have similar access frequencies),
208the data access pattern can be dynamically changed.  This will result in low
209monitoring quality.  To keep the assumption as much as possible, DAMON
210adaptively merges and splits each region based on their access frequency.
211
212For each ``aggregation interval``, it compares the access frequencies of
213adjacent regions and merges those if the frequency difference is small.  Then,
214after it reports and clears the aggregated access frequency of each region, it
215splits each region into two or three regions if the total number of regions
216will not exceed the user-specified maximum number of regions after the split.
217
218In this way, DAMON provides its best-effort quality and minimal overhead while
219keeping the bounds users set for their trade-off.
220
221
222.. _damon_design_age_tracking:
223
224Age Tracking
225~~~~~~~~~~~~
226
227By analyzing the monitoring results, users can also find how long the current
228access pattern of a region has maintained.  That could be used for good
229understanding of the access pattern.  For example, page placement algorithm
230utilizing both the frequency and the recency could be implemented using that.
231To make such access pattern maintained period analysis easier, DAMON maintains
232yet another counter called ``age`` in each region.  For each ``aggregation
233interval``, DAMON checks if the region's size and access frequency
234(``nr_accesses``) has significantly changed.  If so, the counter is reset to
235zero.  Otherwise, the counter is increased.
236
237
238Dynamic Target Space Updates Handling
239~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
240
241The monitoring target address range could dynamically changed.  For example,
242virtual memory could be dynamically mapped and unmapped.  Physical memory could
243be hot-plugged.
244
245As the changes could be quite frequent in some cases, DAMON allows the
246monitoring operations to check dynamic changes including memory mapping changes
247and applies it to monitoring operations-related data structures such as the
248abstracted monitoring target memory area only for each of a user-specified time
249interval (``update interval``).
250
251
252.. _damon_design_damos:
253
254Operation Schemes
255-----------------
256
257One common purpose of data access monitoring is access-aware system efficiency
258optimizations.  For example,
259
260    paging out memory regions that are not accessed for more than two minutes
261
262or
263
264    using THP for memory regions that are larger than 2 MiB and showing a high
265    access frequency for more than one minute.
266
267One straightforward approach for such schemes would be profile-guided
268optimizations.  That is, getting data access monitoring results of the
269workloads or the system using DAMON, finding memory regions of special
270characteristics by profiling the monitoring results, and making system
271operation changes for the regions.  The changes could be made by modifying or
272providing advice to the software (the application and/or the kernel), or
273reconfiguring the hardware.  Both offline and online approaches could be
274available.
275
276Among those, providing advice to the kernel at runtime would be flexible and
277effective, and therefore widely be used.   However, implementing such schemes
278could impose unnecessary redundancy and inefficiency.  The profiling could be
279redundant if the type of interest is common.  Exchanging the information
280including monitoring results and operation advice between kernel and user
281spaces could be inefficient.
282
283To allow users to reduce such redundancy and inefficiencies by offloading the
284works, DAMON provides a feature called Data Access Monitoring-based Operation
285Schemes (DAMOS).  It lets users specify their desired schemes at a high
286level.  For such specifications, DAMON starts monitoring, finds regions having
287the access pattern of interest, and applies the user-desired operation actions
288to the regions, for every user-specified time interval called
289``apply_interval``.
290
291
292.. _damon_design_damos_action:
293
294Operation Action
295~~~~~~~~~~~~~~~~
296
297The management action that the users desire to apply to the regions of their
298interest.  For example, paging out, prioritizing for next reclamation victim
299selection, advising ``khugepaged`` to collapse or split, or doing nothing but
300collecting statistics of the regions.
301
302The list of supported actions is defined in DAMOS, but the implementation of
303each action is in the DAMON operations set layer because the implementation
304normally depends on the monitoring target address space.  For example, the code
305for paging specific virtual address ranges out would be different from that for
306physical address ranges.  And the monitoring operations implementation sets are
307not mandated to support all actions of the list.  Hence, the availability of
308specific DAMOS action depends on what operations set is selected to be used
309together.
310
311The list of the supported actions, their meaning, and DAMON operations sets
312that supports each action are as below.
313
314 - ``willneed``: Call ``madvise()`` for the region with ``MADV_WILLNEED``.
315   Supported by ``vaddr`` and ``fvaddr`` operations set.
316 - ``cold``: Call ``madvise()`` for the region with ``MADV_COLD``.
317   Supported by ``vaddr`` and ``fvaddr`` operations set.
318 - ``pageout``: Reclaim the region.
319   Supported by ``vaddr``, ``fvaddr`` and ``paddr`` operations set.
320 - ``hugepage``: Call ``madvise()`` for the region with ``MADV_HUGEPAGE``.
321   Supported by ``vaddr`` and ``fvaddr`` operations set.
322 - ``nohugepage``: Call ``madvise()`` for the region with ``MADV_NOHUGEPAGE``.
323   Supported by ``vaddr`` and ``fvaddr`` operations set.
324 - ``lru_prio``: Prioritize the region on its LRU lists.
325   Supported by ``paddr`` operations set.
326 - ``lru_deprio``: Deprioritize the region on its LRU lists.
327   Supported by ``paddr`` operations set.
328 - ``stat``: Do nothing but count the statistics.
329   Supported by all operations sets.
330
331Applying the actions except ``stat`` to a region is considered as changing the
332region's characteristics.  Hence, DAMOS resets the age of regions when any such
333actions are applied to those.
334
335
336.. _damon_design_damos_access_pattern:
337
338Target Access Pattern
339~~~~~~~~~~~~~~~~~~~~~
340
341The access pattern of the schemes' interest.  The patterns are constructed with
342the properties that DAMON's monitoring results provide, specifically the size,
343the access frequency, and the age.  Users can describe their access pattern of
344interest by setting minimum and maximum values of the three properties.  If a
345region's three properties are in the ranges, DAMOS classifies it as one of the
346regions that the scheme is having an interest in.
347
348
349.. _damon_design_damos_quotas:
350
351Quotas
352~~~~~~
353
354DAMOS upper-bound overhead control feature.  DAMOS could incur high overhead if
355the target access pattern is not properly tuned.  For example, if a huge memory
356region having the access pattern of interest is found, applying the scheme's
357action to all pages of the huge region could consume unacceptably large system
358resources.  Preventing such issues by tuning the access pattern could be
359challenging, especially if the access patterns of the workloads are highly
360dynamic.
361
362To mitigate that situation, DAMOS provides an upper-bound overhead control
363feature called quotas.  It lets users specify an upper limit of time that DAMOS
364can use for applying the action, and/or a maximum bytes of memory regions that
365the action can be applied within a user-specified time duration.
366
367
368.. _damon_design_damos_quotas_prioritization:
369
370Prioritization
371^^^^^^^^^^^^^^
372
373A mechanism for making a good decision under the quotas.  When the action
374cannot be applied to all regions of interest due to the quotas, DAMOS
375prioritizes regions and applies the action to only regions having high enough
376priorities so that it will not exceed the quotas.
377
378The prioritization mechanism should be different for each action.  For example,
379rarely accessed (colder) memory regions would be prioritized for page-out
380scheme action.  In contrast, the colder regions would be deprioritized for huge
381page collapse scheme action.  Hence, the prioritization mechanisms for each
382action are implemented in each DAMON operations set, together with the actions.
383
384Though the implementation is up to the DAMON operations set, it would be common
385to calculate the priority using the access pattern properties of the regions.
386Some users would want the mechanisms to be personalized for their specific
387case.  For example, some users would want the mechanism to weigh the recency
388(``age``) more than the access frequency (``nr_accesses``).  DAMOS allows users
389to specify the weight of each access pattern property and passes the
390information to the underlying mechanism.  Nevertheless, how and even whether
391the weight will be respected are up to the underlying prioritization mechanism
392implementation.
393
394
395.. _damon_design_damos_quotas_auto_tuning:
396
397Aim-oriented Feedback-driven Auto-tuning
398^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
399
400Automatic feedback-driven quota tuning.  Instead of setting the absolute quota
401value, users can specify the metric of their interest, and what target value
402they want the metric value to be.  DAMOS then automatically tunes the
403aggressiveness (the quota) of the corresponding scheme.  For example, if DAMOS
404is under achieving the goal, DAMOS automatically increases the quota.  If DAMOS
405is over achieving the goal, it decreases the quota.
406
407The goal can be specified with three parameters, namely ``target_metric``,
408``target_value``, and ``current_value``.  The auto-tuning mechanism tries to
409make ``current_value`` of ``target_metric`` be same to ``target_value``.
410Currently, two ``target_metric`` are provided.
411
412- ``user_input``: User-provided value.  Users could use any metric that they
413  has interest in for the value.  Use space main workload's latency or
414  throughput, system metrics like free memory ratio or memory pressure stall
415  time (PSI) could be examples.  Note that users should explicitly set
416  ``current_value`` on their own in this case.  In other words, users should
417  repeatedly provide the feedback.
418- ``some_mem_psi_us``: System-wide ``some`` memory pressure stall information
419  in microseconds that measured from last quota reset to next quota reset.
420  DAMOS does the measurement on its own, so only ``target_value`` need to be
421  set by users at the initial time.  In other words, DAMOS does self-feedback.
422
423
424.. _damon_design_damos_watermarks:
425
426Watermarks
427~~~~~~~~~~
428
429Conditional DAMOS (de)activation automation.  Users might want DAMOS to run
430only under certain situations.  For example, when a sufficient amount of free
431memory is guaranteed, running a scheme for proactive reclamation would only
432consume unnecessary system resources.  To avoid such consumption, the user would
433need to manually monitor some metrics such as free memory ratio, and turn
434DAMON/DAMOS on or off.
435
436DAMOS allows users to offload such works using three watermarks.  It allows the
437users to configure the metric of their interest, and three watermark values,
438namely high, middle, and low.  If the value of the metric becomes above the
439high watermark or below the low watermark, the scheme is deactivated.  If the
440metric becomes below the mid watermark but above the low watermark, the scheme
441is activated.  If all schemes are deactivated by the watermarks, the monitoring
442is also deactivated.  In this case, the DAMON worker thread only periodically
443checks the watermarks and therefore incurs nearly zero overhead.
444
445
446.. _damon_design_damos_filters:
447
448Filters
449~~~~~~~
450
451Non-access pattern-based target memory regions filtering.  If users run
452self-written programs or have good profiling tools, they could know something
453more than the kernel, such as future access patterns or some special
454requirements for specific types of memory. For example, some users may know
455only anonymous pages can impact their program's performance.  They can also
456have a list of latency-critical processes.
457
458To let users optimize DAMOS schemes with such special knowledge, DAMOS provides
459a feature called DAMOS filters.  The feature allows users to set an arbitrary
460number of filters for each scheme.  Each filter specifies the type of target
461memory, and whether it should exclude the memory of the type (filter-out), or
462all except the memory of the type (filter-in).
463
464Currently, anonymous page, memory cgroup, address range, and DAMON monitoring
465target type filters are supported by the feature.  Some filter target types
466require additional arguments.  The memory cgroup filter type asks users to
467specify the file path of the memory cgroup for the filter.  The address range
468type asks the start and end addresses of the range.  The DAMON monitoring
469target type asks the index of the target from the context's monitoring targets
470list.  Hence, users can apply specific schemes to only anonymous pages,
471non-anonymous pages, pages of specific cgroups, all pages excluding those of
472specific cgroups, pages in specific address range, pages in specific DAMON
473monitoring targets, and any combination of those.
474
475To handle filters efficiently, the address range and DAMON monitoring target
476type filters are handled by the core layer, while others are handled by
477operations set.  If a memory region is filtered by a core layer-handled filter,
478it is not counted as the scheme has tried to the region.  In contrast, if a
479memory regions is filtered by an operations set layer-handled filter, it is
480counted as the scheme has tried.  The difference in accounting leads to changes
481in the statistics.
482
483
484Application Programming Interface
485---------------------------------
486
487The programming interface for kernel space data access-aware applications.
488DAMON is a framework, so it does nothing by itself.  Instead, it only helps
489other kernel components such as subsystems and modules building their data
490access-aware applications using DAMON's core features.  For this, DAMON exposes
491its all features to other kernel components via its application programming
492interface, namely ``include/linux/damon.h``.  Please refer to the API
493:doc:`document </mm/damon/api>` for details of the interface.
494
495
496Modules
497=======
498
499Because the core of DAMON is a framework for kernel components, it doesn't
500provide any direct interface for the user space.  Such interfaces should be
501implemented by each DAMON API user kernel components, instead.  DAMON subsystem
502itself implements such DAMON API user modules, which are supposed to be used
503for general purpose DAMON control and special purpose data access-aware system
504operations, and provides stable application binary interfaces (ABI) for the
505user space.  The user space can build their efficient data access-aware
506applications using the interfaces.
507
508
509General Purpose User Interface Modules
510--------------------------------------
511
512DAMON modules that provide user space ABIs for general purpose DAMON usage in
513runtime.
514
515DAMON user interface modules, namely 'DAMON sysfs interface' and 'DAMON debugfs
516interface' are DAMON API user kernel modules that provide ABIs to the
517user-space.  Please note that DAMON debugfs interface is currently deprecated.
518
519Like many other ABIs, the modules create files on sysfs and debugfs, allow
520users to specify their requests to and get the answers from DAMON by writing to
521and reading from the files.  As a response to such I/O, DAMON user interface
522modules control DAMON and retrieve the results as user requested via the DAMON
523API, and return the results to the user-space.
524
525The ABIs are designed to be used for user space applications development,
526rather than human beings' fingers.  Human users are recommended to use such
527user space tools.  One such Python-written user space tool is available at
528Github (https://github.com/awslabs/damo), Pypi
529(https://pypistats.org/packages/damo), and Fedora
530(https://packages.fedoraproject.org/pkgs/python-damo/damo/).
531
532Please refer to the ABI :doc:`document </admin-guide/mm/damon/usage>` for
533details of the interfaces.
534
535
536Special-Purpose Access-aware Kernel Modules
537-------------------------------------------
538
539DAMON modules that provide user space ABI for specific purpose DAMON usage.
540
541DAMON sysfs/debugfs user interfaces are for full control of all DAMON features
542in runtime.  For each special-purpose system-wide data access-aware system
543operations such as proactive reclamation or LRU lists balancing, the interfaces
544could be simplified by removing unnecessary knobs for the specific purpose, and
545extended for boot-time and even compile time control.  Default values of DAMON
546control parameters for the usage would also need to be optimized for the
547purpose.
548
549To support such cases, yet more DAMON API user kernel modules that provide more
550simple and optimized user space interfaces are available.  Currently, two
551modules for proactive reclamation and LRU lists manipulation are provided.  For
552more detail, please read the usage documents for those
553(:doc:`/admin-guide/mm/damon/reclaim` and
554:doc:`/admin-guide/mm/damon/lru_sort`).
555