Re: [PATCH v9 06/14] mm: multi-gen LRU: minimal implementation

[Yu Zhao <yuzhao@google.com>]

On Mon, Mar 21, 2022 at 7:01 AM Aneesh Kumar K.V
<aneesh.kumar@linux.ibm.com> wrote:
>
> Yu Zhao <yuzhao@google.com> writes:
>
> > To avoid confusion, the terms "promotion" and "demotion" will be
> > applied to the multi-gen LRU, as a new convention; the terms
> > "activation" and "deactivation" will be applied to the active/inactive
> > LRU, as usual.
> >
> > 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 has the complexity
> > O(nr_hot_pages), since it is only interested in hot pages. Promotion
> > in the aging path does not require any LRU list operations, only the
> > updates of the gen counter and lrugen->nr_pages[]; demotion, unless as
> > the result of the increment of max_seq, requires LRU list operations,
> > e.g., lru_deactivate_fn().
> >
> > The eviction consumes old generations. Given an lruvec, it increments
> > min_seq when the lists indexed by min_seq%MAX_NR_GENS become empty. A
> > feedback loop modeled after the PID controller monitors refaults over
> > anon and file types and decides which type to evict when both types
> > are available from the same generation.
> >
> > Each generation is divided into multiple tiers. Tiers represent
> > different ranges of numbers of accesses through file descriptors. A
> > page accessed N times through file descriptors is in tier
> > order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
> > bits in folio->flags. In contrast to moving across generations, which
> > requires the LRU lock, moving across tiers only involves operations on
> > folio->flags. The feedback loop also monitors refaults over all tiers
> > and decides when to protect pages in which tiers (N>1), using the
> > first tier (N=0,1) as a baseline. The first tier contains single-use
> > unmapped clean pages, which are most likely the best choices. The
> > eviction moves a page to the next generation, i.e., min_seq+1, if the
> > feedback loop decides so. This approach has the following advantages:
> > 1. It removes the cost of activation in the buffered access path by
> >    inferring whether pages accessed multiple times through file
> >    descriptors are statistically hot and thus worth protecting in the
> >    eviction path.
> > 2. It takes pages accessed through page tables into account and avoids
> >    overprotecting pages accessed multiple times through file
> >    descriptors. (Pages accessed through page tables are in the first
> >    tier, since N=0.)
> > 3. More tiers provide better protection for pages accessed more than
> >    twice through file descriptors, when under heavy buffered I/O
> >    workloads.
> >
> > Server benchmark results:
> >   Single workload:
> >     fio (buffered I/O): +[47, 49]%
> >                 IOPS         BW
> >       5.17-rc2: 2242k        8759MiB/s
> >       patch1-5: 3321k        12.7GiB/s
> >
> >   Single workload:
> >     memcached (anon): +[101, 105]%
> >                 Ops/sec      KB/sec
> >       5.17-rc2: 476771.79    18544.31
> >       patch1-5: 972526.07    37826.95
> >
> >   Configurations:
> >     CPU: two Xeon 6154
> >     Mem: total 256G
> >
> >     Node 1 was only used as a ram disk to reduce the variance in the
> >     results.
> >
> >     patch drivers/block/brd.c <<EOF
> >     99,100c99,100
> >     <         gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
> >     <         page = alloc_page(gfp_flags);
> >     ---
> >     >         gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> >     >         page = alloc_pages_node(1, gfp_flags, 0);
> >     EOF
> >
> >     cat >>/etc/systemd/system.conf <<EOF
> >     CPUAffinity=numa
> >     NUMAPolicy=bind
> >     NUMAMask=0
> >     EOF
> >
> >     cat >>/etc/memcached.conf <<EOF
> >     -m 184320
> >     -s /var/run/memcached/memcached.sock
> >     -a 0766
> >     -t 36
> >     -B binary
> >     EOF
> >
> >     cat fio.sh
> >     modprobe brd rd_nr=1 rd_size=113246208
> >     mkfs.ext4 /dev/ram0
> >     mount -t ext4 /dev/ram0 /mnt
> >
> >     mkdir /sys/fs/cgroup/user.slice/test
> >     echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
> >     echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
> >     fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
> >       --buffered=1 --ioengine=io_uring --iodepth=128 \
> >       --iodepth_batch_submit=32 --iodepth_batch_complete=32 \
> >       --rw=randread --random_distribution=random --norandommap \
> >       --time_based --ramp_time=10m --runtime=5m --group_reporting
> >
> >     cat memcached.sh
> >     modprobe brd rd_nr=1 rd_size=113246208
> >     swapoff -a
> >     mkswap /dev/ram0
> >     swapon /dev/ram0
> >
> >     memtier_benchmark -S /var/run/memcached/memcached.sock \
> >       -P memcache_binary -n allkeys --key-minimum=1 \
> >       --key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
> >       --ratio 1:0 --pipeline 8 -d 2000
> >
> >     memtier_benchmark -S /var/run/memcached/memcached.sock \
> >       -P memcache_binary -n allkeys --key-minimum=1 \
> >       --key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
> >       --ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
> >
> > Client benchmark results:
> >   kswapd profiles:
> >     5.17-rc2
> >       38.05%  page_vma_mapped_walk
> >       20.86%  lzo1x_1_do_compress (real work)
> >        6.16%  do_raw_spin_lock
> >        4.61%  _raw_spin_unlock_irq
> >        2.20%  vma_interval_tree_iter_next
> >        2.19%  vma_interval_tree_subtree_search
> >        2.15%  page_referenced_one
> >        1.93%  anon_vma_interval_tree_iter_first
> >        1.65%  ptep_clear_flush
> >        1.00%  __zram_bvec_write
> >
> >     patch1-5
> >       39.73%  lzo1x_1_do_compress (real work)
> >       14.96%  page_vma_mapped_walk
> >        6.97%  _raw_spin_unlock_irq
> >        3.07%  do_raw_spin_lock
> >        2.53%  anon_vma_interval_tree_iter_first
> >        2.04%  ptep_clear_flush
> >        1.82%  __zram_bvec_write
> >        1.76%  __anon_vma_interval_tree_subtree_search
> >        1.57%  memmove
> >        1.45%  free_unref_page_list
> >
> >   Configurations:
> >     CPU: single Snapdragon 7c
> >     Mem: total 4G
> >
> >     Chrome OS MemoryPressure [1]
> >
> > [1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
> >
>
> In shrink_active_list we do preferential treatment of VM_EXEC pages.
> Do we do similar thing with MGLRU? if not why is that not needed?

No, because MGLRU has a different set of assumptions than the
active/inactive LRU does [1]. It provides mmapped pages with equal
opportunities, and the tradeoff was discussed here [2].

Note that even with this preferential treatment of executable pages,
plus other heuristics added since then, executable pages are still
underprotected for at least desktop workloads [3]. And I can confirm
the problem reported is genuine -- we recently accidentally removed
our private patch that works around the problem for the last 12 years,
and observed immediate consequences on a small portion of devices not
using MGLRU [4].

[1] https://lore.kernel.org/linux-mm/20220309021230.721028-15-yuzhao@google.com/
[2] https://lore.kernel.org/linux-mm/20220208081902.3550911-5-yuzhao@google.com/
[3] https://lore.kernel.org/linux-mm/2dc51fc8-f14e-17ed-a8c6-0ec70423bf54@valdikss.org.ru/
[4] https://chromium-review.googlesource.com/c/chromiumos/third_party/kernel/+/3429559