Design

Execution Model and Data Structures

The monitoring-related information including the monitoring request specification and DAMON-based operation schemes are stored in a data structure called DAMON context. DAMON executes each context with a kernel thread called kdamond. Multiple kdamonds could run in parallel, for different types of monitoring.

Overall Architecture

DAMON subsystem is configured with three layers including

  • Operations Set: Implements fundamental operations for DAMON that depends on the given monitoring target address-space and available set of software/hardware primitives,

  • Core: Implements core logics including monitoring overhead/accurach control and access-aware system operations on top of the operations set layer, and

  • Modules: Implements kernel modules for various purposes that provides interfaces for the user space, on top of the core layer.

Configurable Operations Set

For data access monitoring and additional low level work, DAMON needs a set of implementations for specific operations that are dependent on and optimized for the given target address space. On the other hand, the accuracy and overhead tradeoff mechanism, which is the core logic of DAMON, is in the pure logic space. DAMON separates the two parts in different layers, namely DAMON Operations Set and DAMON Core Logics Layers, respectively. It further defines the interface between the layers to allow various operations sets to be configured with the core logic.

Due to this design, users can extend DAMON for any address space by configuring the core logic to use the appropriate operations set. If any appropriate set is unavailable, users can implement one on their own.

For example, physical memory, virtual memory, swap space, those for specific processes, NUMA nodes, files, and backing memory devices would be supportable. Also, if some architectures or devices supporting special optimized access check primitives, those will be easily configurable.

Programmable Modules

Core layer of DAMON is implemented as a framework, and exposes its application programming interface to all kernel space components such as subsystems and modules. For common use cases of DAMON, DAMON subsystem provides kernel modules that built on top of the core layer using the API, which can be easily used by the user space end users.

Operations Set Layer

The monitoring operations are defined in two parts:

  1. Identification of the monitoring target address range for the address space.

  2. Access check of specific address range in the target space.

DAMON currently provides below three operation sets. Below two subsections describe how those work.

  • vaddr: Monitor virtual address spaces of specific processes

  • fvaddr: Monitor fixed virtual address ranges

  • paddr: Monitor the physical address space of the system

VMA-based Target Address Range Construction

A mechanism of vaddr DAMON operations set that automatically initializes and updates the monitoring target address regions so that entire memory mappings of the target processes can be covered.

This mechanism is only for the vaddr operations set. In cases of fvaddr and paddr operation sets, users are asked to manually set the monitoring target address ranges.

Only small parts in the super-huge virtual address space of the processes are mapped to the physical memory and accessed. Thus, tracking the unmapped address regions is just wasteful. However, because DAMON can deal with some level of noise using the adaptive regions adjustment mechanism, tracking every mapping is not strictly required but could even incur a high overhead in some cases. That said, too huge unmapped areas inside the monitoring target should be removed to not take the time for the adaptive mechanism.

For the reason, this implementation converts the complex mappings to three distinct regions that cover every mapped area of the address space. The two gaps between the three regions are the two biggest unmapped areas in the given address space. The two biggest unmapped areas would be the gap between the heap and the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed region and the stack in most of the cases. Because these gaps are exceptionally huge in usual address spaces, excluding these will be sufficient to make a reasonable trade-off. Below shows this in detail:

<heap>
<BIG UNMAPPED REGION 1>
<uppermost mmap()-ed region>
(small mmap()-ed regions and munmap()-ed regions)
<lowermost mmap()-ed region>
<BIG UNMAPPED REGION 2>
<stack>

PTE Accessed-bit Based Access Check

Both of the implementations for physical and virtual address spaces use PTE Accessed-bit for basic access checks. Only one difference is the way of finding the relevant PTE Accessed bit(s) from the address. While the implementation for the virtual address walks the page table for the target task of the address, the implementation for the physical address walks every page table having a mapping to the address. In this way, the implementations find and clear the bit(s) for next sampling target address and checks whether the bit(s) set again after one sampling period. This could disturb other kernel subsystems using the Accessed bits, namely Idle page tracking and the reclaim logic. DAMON does nothing to avoid disturbing Idle page tracking, so handling the interference is the responsibility of sysadmins. However, it solves the conflict with the reclaim logic using PG_idle and PG_young page flags, as Idle page tracking does.

Core Logics

Monitoring

Below four sections describe each of the DAMON core mechanisms and the five monitoring attributes, sampling interval, aggregation interval, update interval, minimum number of regions, and maximum number of regions.

Access Frequency Monitoring

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

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

The monitoring overhead of this mechanism will arbitrarily increase as the size of the target workload grows.

Region Based Sampling

To avoid the unbounded increase of the overhead, DAMON groups adjacent pages that assumed to have the same access frequencies into a region. As long as the assumption (pages in a region have the same access frequencies) is kept, only one page in the region is required to be checked. Thus, for each sampling interval, DAMON randomly picks one page in each region, waits for one sampling interval, checks whether the page is accessed meanwhile, and increases the access frequency counter of the region if so. The counter is called nr_regions of the region. Therefore, the monitoring overhead is controllable by setting the number of regions. DAMON allows users to set the minimum and the maximum number of regions for the trade-off.

This scheme, however, cannot preserve the quality of the output if the assumption is not guaranteed.

Adaptive Regions Adjustment

Even somehow the initial monitoring target regions are well constructed to fulfill the assumption (pages in same region have similar access frequencies), the data access pattern can be dynamically changed. This will result in low monitoring quality. To keep the assumption as much as possible, DAMON adaptively merges and splits each region based on their access frequency.

For each aggregation interval, it compares the access frequencies of adjacent regions and merges those if the frequency difference is small. Then, after it reports and clears the aggregated access frequency of each region, it splits each region into two or three regions if the total number of regions will not exceed the user-specified maximum number of regions after the split.

In this way, DAMON provides its best-effort quality and minimal overhead while keeping the bounds users set for their trade-off.

Age Tracking

By analyzing the monitoring results, users can also find how long the current access pattern of a region has maintained. That could be used for good understanding of the access pattern. For example, page placement algorithm utilizing both the frequency and the recency could be implemented using that. To make such access pattern maintained period analysis easier, DAMON maintains yet another counter called age in each region. For each aggregation interval, DAMON checks if the region’s size and access frequency (nr_accesses) has significantly changed. If so, the counter is reset to zero. Otherwise, the counter is increased.

Dynamic Target Space Updates Handling

The monitoring target address range could dynamically changed. For example, virtual memory could be dynamically mapped and unmapped. Physical memory could be hot-plugged.

As the changes could be quite frequent in some cases, DAMON allows the monitoring operations to check dynamic changes including memory mapping changes and applies it to monitoring operations-related data structures such as the abstracted monitoring target memory area only for each of a user-specified time interval (update interval).

Operation Schemes

One common purpose of data access monitoring is access-aware system efficiency optimizations. For example,

paging out memory regions that are not accessed for more than two minutes

or

using THP for memory regions that are larger than 2 MiB and showing a high access frequency for more than one minute.

One straightforward approach for such schemes would be profile-guided optimizations. That is, getting data access monitoring results of the workloads or the system using DAMON, finding memory regions of special characteristics by profiling the monitoring results, and making system operation changes for the regions. The changes could be made by modifying or providing advice to the software (the application and/or the kernel), or reconfiguring the hardware. Both offline and online approaches could be available.

Among those, providing advice to the kernel at runtime would be flexible and effective, and therefore widely be used. However, implementing such schemes could impose unnecessary redundancy and inefficiency. The profiling could be redundant if the type of interest is common. Exchanging the information including monitoring results and operation advice between kernel and user spaces could be inefficient.

To allow users to reduce such redundancy and inefficiencies by offloading the works, DAMON provides a feature called Data Access Monitoring-based Operation Schemes (DAMOS). It lets users specify their desired schemes at a high level. For such specifications, DAMON starts monitoring, finds regions having the access pattern of interest, and applies the user-desired operation actions to the regions, for every user-specified time interval called apply_interval.

Operation Action

The management action that the users desire to apply to the regions of their interest. For example, paging out, prioritizing for next reclamation victim selection, advising khugepaged to collapse or split, or doing nothing but collecting statistics of the regions.

The list of supported actions is defined in DAMOS, but the implementation of each action is in the DAMON operations set layer because the implementation normally depends on the monitoring target address space. For example, the code for paging specific virtual address ranges out would be different from that for physical address ranges. And the monitoring operations implementation sets are not mandated to support all actions of the list. Hence, the availability of specific DAMOS action depends on what operations set is selected to be used together.

The list of the supported actions, their meaning, and DAMON operations sets that supports each action are as below.

  • willneed: Call madvise() for the region with MADV_WILLNEED. Supported by vaddr and fvaddr operations set.

  • cold: Call madvise() for the region with MADV_COLD. Supported by vaddr and fvaddr operations set.

  • pageout: Reclaim the region. Supported by vaddr, fvaddr and paddr operations set.

  • hugepage: Call madvise() for the region with MADV_HUGEPAGE. Supported by vaddr and fvaddr operations set.

  • nohugepage: Call madvise() for the region with MADV_NOHUGEPAGE. Supported by vaddr and fvaddr operations set.

  • lru_prio: Prioritize the region on its LRU lists. Supported by paddr operations set.

  • lru_deprio: Deprioritize the region on its LRU lists. Supported by paddr operations set.

  • stat: Do nothing but count the statistics. Supported by all operations sets.

Applying the actions except stat to a region is considered as changing the region’s characteristics. Hence, DAMOS resets the age of regions when any such actions are applied to those.

Target Access Pattern

The access pattern of the schemes’ interest. The patterns are constructed with the properties that DAMON’s monitoring results provide, specifically the size, the access frequency, and the age. Users can describe their access pattern of interest by setting minimum and maximum values of the three properties. If a region’s three properties are in the ranges, DAMOS classifies it as one of the regions that the scheme is having an interest in.

Quotas

DAMOS upper-bound overhead control feature. DAMOS could incur high overhead if the target access pattern is not properly tuned. For example, if a huge memory region having the access pattern of interest is found, applying the scheme’s action to all pages of the huge region could consume unacceptably large system resources. Preventing such issues by tuning the access pattern could be challenging, especially if the access patterns of the workloads are highly dynamic.

To mitigate that situation, DAMOS provides an upper-bound overhead control feature called quotas. It lets users specify an upper limit of time that DAMOS can use for applying the action, and/or a maximum bytes of memory regions that the action can be applied within a user-specified time duration.

Prioritization

A mechanism for making a good decision under the quotas. When the action cannot be applied to all regions of interest due to the quotas, DAMOS prioritizes regions and applies the action to only regions having high enough priorities so that it will not exceed the quotas.

The prioritization mechanism should be different for each action. For example, rarely accessed (colder) memory regions would be prioritized for page-out scheme action. In contrast, the colder regions would be deprioritized for huge page collapse scheme action. Hence, the prioritization mechanisms for each action are implemented in each DAMON operations set, together with the actions.

Though the implementation is up to the DAMON operations set, it would be common to calculate the priority using the access pattern properties of the regions. Some users would want the mechanisms to be personalized for their specific case. For example, some users would want the mechanism to weigh the recency (age) more than the access frequency (nr_accesses). DAMOS allows users to specify the weight of each access pattern property and passes the information to the underlying mechanism. Nevertheless, how and even whether the weight will be respected are up to the underlying prioritization mechanism implementation.

Aim-oriented Feedback-driven Auto-tuning

Automatic feedback-driven quota tuning. Instead of setting the absolute quota value, users can specify the metric of their interest, and what target value they want the metric value to be. DAMOS then automatically tunes the aggressiveness (the quota) of the corresponding scheme. For example, if DAMOS is under achieving the goal, DAMOS automatically increases the quota. If DAMOS is over achieving the goal, it decreases the quota.

The goal can be specified with three parameters, namely target_metric, target_value, and current_value. The auto-tuning mechanism tries to make current_value of target_metric be same to target_value. Currently, two target_metric are provided.

  • user_input: User-provided value. Users could use any metric that they has interest in for the value. Use space main workload’s latency or throughput, system metrics like free memory ratio or memory pressure stall time (PSI) could be examples. Note that users should explicitly set current_value on their own in this case. In other words, users should repeatedly provide the feedback.

  • some_mem_psi_us: System-wide some memory pressure stall information in microseconds that measured from last quota reset to next quota reset. DAMOS does the measurement on its own, so only target_value need to be set by users at the initial time. In other words, DAMOS does self-feedback.

Watermarks

Conditional DAMOS (de)activation automation. Users might want DAMOS to run only under certain situations. For example, when a sufficient amount of free memory is guaranteed, running a scheme for proactive reclamation would only consume unnecessary system resources. To avoid such consumption, the user would need to manually monitor some metrics such as free memory ratio, and turn DAMON/DAMOS on or off.

DAMOS allows users to offload such works using three watermarks. It allows the users to configure the metric of their interest, and three watermark values, namely high, middle, and low. If the value of the metric becomes above the high watermark or below the low watermark, the scheme is deactivated. If the metric becomes below the mid watermark but above the low watermark, the scheme is activated. If all schemes are deactivated by the watermarks, the monitoring is also deactivated. In this case, the DAMON worker thread only periodically checks the watermarks and therefore incurs nearly zero overhead.

Filters

Non-access pattern-based target memory regions filtering. If users run self-written programs or have good profiling tools, they could know something more than the kernel, such as future access patterns or some special requirements for specific types of memory. For example, some users may know only anonymous pages can impact their program’s performance. They can also have a list of latency-critical processes.

To let users optimize DAMOS schemes with such special knowledge, DAMOS provides a feature called DAMOS filters. The feature allows users to set an arbitrary number of filters for each scheme. Each filter specifies the type of target memory, and whether it should exclude the memory of the type (filter-out), or all except the memory of the type (filter-in).

Currently, anonymous page, memory cgroup, address range, and DAMON monitoring target type filters are supported by the feature. Some filter target types require additional arguments. The memory cgroup filter type asks users to specify the file path of the memory cgroup for the filter. The address range type asks the start and end addresses of the range. The DAMON monitoring target type asks the index of the target from the context’s monitoring targets list. Hence, users can apply specific schemes to only anonymous pages, non-anonymous pages, pages of specific cgroups, all pages excluding those of specific cgroups, pages in specific address range, pages in specific DAMON monitoring targets, and any combination of those.

To handle filters efficiently, the address range and DAMON monitoring target type filters are handled by the core layer, while others are handled by operations set. If a memory region is filtered by a core layer-handled filter, it is not counted as the scheme has tried to the region. In contrast, if a memory regions is filtered by an operations set layer-handled filter, it is counted as the scheme has tried. The difference in accounting leads to changes in the statistics.

Application Programming Interface

The programming interface for kernel space data access-aware applications. DAMON is a framework, so it does nothing by itself. Instead, it only helps other kernel components such as subsystems and modules building their data access-aware applications using DAMON’s core features. For this, DAMON exposes its all features to other kernel components via its application programming interface, namely include/linux/damon.h. Please refer to the API document for details of the interface.

Modules

Because the core of DAMON is a framework for kernel components, it doesn’t provide any direct interface for the user space. Such interfaces should be implemented by each DAMON API user kernel components, instead. DAMON subsystem itself implements such DAMON API user modules, which are supposed to be used for general purpose DAMON control and special purpose data access-aware system operations, and provides stable application binary interfaces (ABI) for the user space. The user space can build their efficient data access-aware applications using the interfaces.

General Purpose User Interface Modules

DAMON modules that provide user space ABIs for general purpose DAMON usage in runtime.

DAMON user interface modules, namely ‘DAMON sysfs interface’ and ‘DAMON debugfs interface’ are DAMON API user kernel modules that provide ABIs to the user-space. Please note that DAMON debugfs interface is currently deprecated.

Like many other ABIs, the modules create files on sysfs and debugfs, allow users to specify their requests to and get the answers from DAMON by writing to and reading from the files. As a response to such I/O, DAMON user interface modules control DAMON and retrieve the results as user requested via the DAMON API, and return the results to the user-space.

The ABIs are designed to be used for user space applications development, rather than human beings’ fingers. Human users are recommended to use such user space tools. One such Python-written user space tool is available at Github (https://github.com/awslabs/damo), Pypi (https://pypistats.org/packages/damo), and Fedora (https://packages.fedoraproject.org/pkgs/python-damo/damo/).

Please refer to the ABI document for details of the interfaces.

Special-Purpose Access-aware Kernel Modules

DAMON modules that provide user space ABI for specific purpose DAMON usage.

DAMON sysfs/debugfs user interfaces are for full control of all DAMON features in runtime. For each special-purpose system-wide data access-aware system operations such as proactive reclamation or LRU lists balancing, the interfaces could be simplified by removing unnecessary knobs for the specific purpose, and extended for boot-time and even compile time control. Default values of DAMON control parameters for the usage would also need to be optimized for the purpose.

To support such cases, yet more DAMON API user kernel modules that provide more simple and optimized user space interfaces are available. Currently, two modules for proactive reclamation and LRU lists manipulation are provided. For more detail, please read the usage documents for those (DAMON-based Reclamation and DAMON-based LRU-lists Sorting).