mirror of
https://github.com/torvalds/linux.git
synced 2024-11-25 05:32:00 +00:00
32d32ef140
This patch improves the design doc. Specifically, 1. add a section for the per-memcg mm_struct list, and 2. add a section for the PID controller. Link: https://lkml.kernel.org/r/20230214035445.1250139-2-talumbau@google.com Signed-off-by: T.J. Alumbaugh <talumbau@google.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
270 lines
12 KiB
ReStructuredText
270 lines
12 KiB
ReStructuredText
.. SPDX-License-Identifier: GPL-2.0
|
|
|
|
=============
|
|
Multi-Gen LRU
|
|
=============
|
|
The multi-gen LRU is an alternative LRU implementation that optimizes
|
|
page reclaim and improves performance under memory pressure. Page
|
|
reclaim decides the kernel's caching policy and ability to overcommit
|
|
memory. It directly impacts the kswapd CPU usage and RAM efficiency.
|
|
|
|
Design overview
|
|
===============
|
|
Objectives
|
|
----------
|
|
The design objectives are:
|
|
|
|
* Good representation of access recency
|
|
* Try to profit from spatial locality
|
|
* Fast paths to make obvious choices
|
|
* Simple self-correcting heuristics
|
|
|
|
The representation of access recency is at the core of all LRU
|
|
implementations. In the multi-gen LRU, each generation represents a
|
|
group of pages with similar access recency. Generations establish a
|
|
(time-based) common frame of reference and therefore help make better
|
|
choices, e.g., between different memcgs on a computer or different
|
|
computers in a data center (for job scheduling).
|
|
|
|
Exploiting spatial locality improves efficiency when gathering the
|
|
accessed bit. A rmap walk targets a single page and does not try to
|
|
profit from discovering a young PTE. A page table walk can sweep all
|
|
the young PTEs in an address space, but the address space can be too
|
|
sparse to make a profit. The key is to optimize both methods and use
|
|
them in combination.
|
|
|
|
Fast paths reduce code complexity and runtime overhead. Unmapped pages
|
|
do not require TLB flushes; clean pages do not require writeback.
|
|
These facts are only helpful when other conditions, e.g., access
|
|
recency, are similar. With generations as a common frame of reference,
|
|
additional factors stand out. But obvious choices might not be good
|
|
choices; thus self-correction is necessary.
|
|
|
|
The benefits of simple self-correcting heuristics are self-evident.
|
|
Again, with generations as a common frame of reference, this becomes
|
|
attainable. Specifically, pages in the same generation can be
|
|
categorized based on additional factors, and a feedback loop can
|
|
statistically compare the refault percentages across those categories
|
|
and infer which of them are better choices.
|
|
|
|
Assumptions
|
|
-----------
|
|
The protection of hot pages and the selection of cold pages are based
|
|
on page access channels and patterns. There are two access channels:
|
|
|
|
* Accesses through page tables
|
|
* Accesses through file descriptors
|
|
|
|
The protection of the former channel is by design stronger because:
|
|
|
|
1. The uncertainty in determining the access patterns of the former
|
|
channel is higher due to the approximation of the accessed bit.
|
|
2. The cost of evicting the former channel is higher due to the TLB
|
|
flushes required and the likelihood of encountering the dirty bit.
|
|
3. The penalty of underprotecting the former channel is higher because
|
|
applications usually do not prepare themselves for major page
|
|
faults like they do for blocked I/O. E.g., GUI applications
|
|
commonly use dedicated I/O threads to avoid blocking rendering
|
|
threads.
|
|
|
|
There are also two access patterns:
|
|
|
|
* Accesses exhibiting temporal locality
|
|
* Accesses not exhibiting temporal locality
|
|
|
|
For the reasons listed above, the former channel is assumed to follow
|
|
the former pattern unless ``VM_SEQ_READ`` or ``VM_RAND_READ`` is
|
|
present, and the latter channel is assumed to follow the latter
|
|
pattern unless outlying refaults have been observed.
|
|
|
|
Workflow overview
|
|
=================
|
|
Evictable pages are divided into multiple generations for each
|
|
``lruvec``. The youngest generation number is stored in
|
|
``lrugen->max_seq`` for both anon and file types as they are aged on
|
|
an equal footing. The oldest generation numbers are stored in
|
|
``lrugen->min_seq[]`` separately for anon and file types as clean file
|
|
pages can be evicted regardless of swap constraints. These three
|
|
variables are monotonically increasing.
|
|
|
|
Generation numbers are truncated into ``order_base_2(MAX_NR_GENS+1)``
|
|
bits in order to fit into the gen counter in ``folio->flags``. Each
|
|
truncated generation number is an index to ``lrugen->folios[]``. The
|
|
sliding window technique is used to track at least ``MIN_NR_GENS`` and
|
|
at most ``MAX_NR_GENS`` generations. The gen counter stores a value
|
|
within ``[1, MAX_NR_GENS]`` while a page is on one of
|
|
``lrugen->folios[]``; otherwise it stores zero.
|
|
|
|
Each generation is divided into multiple tiers. A page accessed ``N``
|
|
times through file descriptors is in tier ``order_base_2(N)``. Unlike
|
|
generations, tiers do not have dedicated ``lrugen->folios[]``. In
|
|
contrast to moving across generations, which requires the LRU lock,
|
|
moving across tiers only involves atomic operations on
|
|
``folio->flags`` and therefore has a negligible cost. A feedback loop
|
|
modeled after the PID controller monitors refaults over all the tiers
|
|
from anon and file types and decides which tiers from which types to
|
|
evict or protect. The desired effect is to balance refault percentages
|
|
between anon and file types proportional to the swappiness level.
|
|
|
|
There are two conceptually independent procedures: the aging and the
|
|
eviction. They form a closed-loop system, i.e., the page reclaim.
|
|
|
|
Aging
|
|
-----
|
|
The aging produces young generations. Given an ``lruvec``, it
|
|
increments ``max_seq`` when ``max_seq-min_seq+1`` approaches
|
|
``MIN_NR_GENS``. The aging promotes hot pages to the youngest
|
|
generation when it finds them accessed through page tables; the
|
|
demotion of cold pages happens consequently when it increments
|
|
``max_seq``. The aging uses page table walks and rmap walks to find
|
|
young PTEs. For the former, it iterates ``lruvec_memcg()->mm_list``
|
|
and calls ``walk_page_range()`` with each ``mm_struct`` on this list
|
|
to scan PTEs, and after each iteration, it increments ``max_seq``. For
|
|
the latter, when the eviction walks the rmap and finds a young PTE,
|
|
the aging scans the adjacent PTEs. For both, on finding a young PTE,
|
|
the aging clears the accessed bit and updates the gen counter of the
|
|
page mapped by this PTE to ``(max_seq%MAX_NR_GENS)+1``.
|
|
|
|
Eviction
|
|
--------
|
|
The eviction consumes old generations. Given an ``lruvec``, it
|
|
increments ``min_seq`` when ``lrugen->folios[]`` indexed by
|
|
``min_seq%MAX_NR_GENS`` becomes empty. To select a type and a tier to
|
|
evict from, it first compares ``min_seq[]`` to select the older type.
|
|
If both types are equally old, it selects the one whose first tier has
|
|
a lower refault percentage. The first tier contains single-use
|
|
unmapped clean pages, which are the best bet. The eviction sorts a
|
|
page according to its gen counter if the aging has found this page
|
|
accessed through page tables and updated its gen counter. It also
|
|
moves a page to the next generation, i.e., ``min_seq+1``, if this page
|
|
was accessed multiple times through file descriptors and the feedback
|
|
loop has detected outlying refaults from the tier this page is in. To
|
|
this end, the feedback loop uses the first tier as the baseline, for
|
|
the reason stated earlier.
|
|
|
|
Working set protection
|
|
----------------------
|
|
Each generation is timestamped at birth. If ``lru_gen_min_ttl`` is
|
|
set, an ``lruvec`` is protected from the eviction when its oldest
|
|
generation was born within ``lru_gen_min_ttl`` milliseconds. In other
|
|
words, it prevents the working set of ``lru_gen_min_ttl`` milliseconds
|
|
from getting evicted. The OOM killer is triggered if this working set
|
|
cannot be kept in memory.
|
|
|
|
This time-based approach has the following advantages:
|
|
|
|
1. It is easier to configure because it is agnostic to applications
|
|
and memory sizes.
|
|
2. It is more reliable because it is directly wired to the OOM killer.
|
|
|
|
``mm_struct`` list
|
|
------------------
|
|
An ``mm_struct`` list is maintained for each memcg, and an
|
|
``mm_struct`` follows its owner task to the new memcg when this task
|
|
is migrated.
|
|
|
|
A page table walker iterates ``lruvec_memcg()->mm_list`` and calls
|
|
``walk_page_range()`` with each ``mm_struct`` on this list to scan
|
|
PTEs. When multiple page table walkers iterate the same list, each of
|
|
them gets a unique ``mm_struct``, and therefore they can run in
|
|
parallel.
|
|
|
|
Page table walkers ignore any misplaced pages, e.g., if an
|
|
``mm_struct`` was migrated, pages left in the previous memcg will be
|
|
ignored when the current memcg is under reclaim. Similarly, page table
|
|
walkers will ignore pages from nodes other than the one under reclaim.
|
|
|
|
This infrastructure also tracks the usage of ``mm_struct`` between
|
|
context switches so that page table walkers can skip processes that
|
|
have been sleeping since the last iteration.
|
|
|
|
Rmap/PT walk feedback
|
|
---------------------
|
|
Searching the rmap for PTEs mapping each page on an LRU list (to test
|
|
and clear the accessed bit) can be expensive because pages from
|
|
different VMAs (PA space) are not cache friendly to the rmap (VA
|
|
space). For workloads mostly using mapped pages, searching the rmap
|
|
can incur the highest CPU cost in the reclaim path.
|
|
|
|
``lru_gen_look_around()`` exploits spatial locality to reduce the
|
|
trips into the rmap. It scans the adjacent PTEs of a young PTE and
|
|
promotes hot pages. If the scan was done cacheline efficiently, it
|
|
adds the PMD entry pointing to the PTE table to the Bloom filter. This
|
|
forms a feedback loop between the eviction and the aging.
|
|
|
|
Bloom filters
|
|
-------------
|
|
Bloom filters are a space and memory efficient data structure for set
|
|
membership test, i.e., test if an element is not in the set or may be
|
|
in the set.
|
|
|
|
In the eviction path, specifically, in ``lru_gen_look_around()``, if a
|
|
PMD has a sufficient number of hot pages, its address is placed in the
|
|
filter. In the aging path, set membership means that the PTE range
|
|
will be scanned for young pages.
|
|
|
|
Note that Bloom filters are probabilistic on set membership. If a test
|
|
is false positive, the cost is an additional scan of a range of PTEs,
|
|
which may yield hot pages anyway. Parameters of the filter itself can
|
|
control the false positive rate in the limit.
|
|
|
|
PID controller
|
|
--------------
|
|
A feedback loop modeled after the Proportional-Integral-Derivative
|
|
(PID) controller monitors refaults over anon and file types and
|
|
decides which type to evict when both types are available from the
|
|
same generation.
|
|
|
|
The PID controller uses generations rather than the wall clock as the
|
|
time domain because a CPU can scan pages at different rates under
|
|
varying memory pressure. It calculates a moving average for each new
|
|
generation to avoid being permanently locked in a suboptimal state.
|
|
|
|
Memcg LRU
|
|
---------
|
|
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
|
|
since each node and memcg combination has an LRU of folios (see
|
|
``mem_cgroup_lruvec()``). Its goal is to improve the scalability of
|
|
global reclaim, which is critical to system-wide memory overcommit in
|
|
data centers. Note that memcg LRU only applies to global reclaim.
|
|
|
|
The basic structure of an memcg LRU can be understood by an analogy to
|
|
the active/inactive LRU (of folios):
|
|
|
|
1. It has the young and the old (generations), i.e., the counterparts
|
|
to the active and the inactive;
|
|
2. The increment of ``max_seq`` triggers promotion, i.e., the
|
|
counterpart to activation;
|
|
3. Other events trigger similar operations, e.g., offlining an memcg
|
|
triggers demotion, i.e., the counterpart to deactivation.
|
|
|
|
In terms of global reclaim, it has two distinct features:
|
|
|
|
1. Sharding, which allows each thread to start at a random memcg (in
|
|
the old generation) and improves parallelism;
|
|
2. Eventual fairness, which allows direct reclaim to bail out at will
|
|
and reduces latency without affecting fairness over some time.
|
|
|
|
In terms of traversing memcgs during global reclaim, it improves the
|
|
best-case complexity from O(n) to O(1) and does not affect the
|
|
worst-case complexity O(n). Therefore, on average, it has a sublinear
|
|
complexity.
|
|
|
|
Summary
|
|
-------
|
|
The multi-gen LRU (of folios) can be disassembled into the following
|
|
parts:
|
|
|
|
* Generations
|
|
* Rmap walks
|
|
* Page table walks via ``mm_struct`` list
|
|
* Bloom filters for rmap/PT walk feedback
|
|
* PID controller for refault feedback
|
|
|
|
The aging and the eviction form a producer-consumer model;
|
|
specifically, the latter drives the former by the sliding window over
|
|
generations. Within the aging, rmap walks drive page table walks by
|
|
inserting hot densely populated page tables to the Bloom filters.
|
|
Within the eviction, the PID controller uses refaults as the feedback
|
|
to select types to evict and tiers to protect.
|