On Thu, Jan 25, 2024 at 3:33 PM Chris Li <chrisl@xxxxxxxxxx> wrote:
>
> On Tue, Jan 23, 2024 at 10:46 AM Kairui Song <ryncsn@xxxxxxxxx> wrote:
> >
> > From: Kairui Song <kasong@xxxxxxxxxxx>
> >
> > Prefetch for inactive/active LRU have been long exiting, apply the same
> > optimization for MGLRU.
> >
> > Test 1: Ramdisk fio ro test in a 4G memcg on a EPYC 7K62:
> > fio -name=mglru --numjobs=16 --directory=/mnt --size=960m \
> > --buffered=1 --ioengine=io_uring --iodepth=128 \
> > --iodepth_batch_submit=32 --iodepth_batch_complete=32 \
> > --rw=randread --random_distribution=zipf:0.5 --norandommap \
> > --time_based --ramp_time=1m --runtime=6m --group_reporting
> >
> > Before this patch:
> > bw ( MiB/s): min= 7758, max= 9239, per=100.00%, avg=8747.59, stdev=16.51, samples=11488
> > iops : min=1986251, max=2365323, avg=2239380.87, stdev=4225.93, samples=11488
> >
> > After this patch (+7.2%):
> > bw ( MiB/s): min= 8360, max= 9771, per=100.00%, avg=9381.31, stdev=15.67, samples=11488
> > iops : min=2140296, max=2501385, avg=2401613.91, stdev=4010.41, samples=11488
> >
> > Test 2: Ramdisk fio hybrid test for 30m in a 4G memcg on a EPYC 7K62 (3 times):
> > fio --buffered=1 --numjobs=8 --size=960m --directory=/mnt \
> > --time_based --ramp_time=1m --runtime=30m \
> > --ioengine=io_uring --iodepth=128 --iodepth_batch_submit=32 \
> > --iodepth_batch_complete=32 --norandommap \
> > --name=mglru-ro --rw=randread --random_distribution=zipf:0.7 \
> > --name=mglru-rw --rw=randrw --random_distribution=zipf:0.7
> >
> > Before this patch:
> > READ: 6622.0 MiB/s. Stdev: 22.090722
> > WRITE: 1256.3 MiB/s. Stdev: 5.249339
> >
> > After this patch (+4.6%, +3.3%):
> > READ: 6926.6 MiB/s, Stdev: 37.950260
> > WRITE: 1297.3 MiB/s, Stdev: 7.408704
> >
> > Test 3: 30m of MySQL test in 6G memcg (12 times):
> > echo 'set GLOBAL innodb_buffer_pool_size=16106127360;' | \
> > mysql -u USER -h localhost --password=PASS
> >
> > sysbench /usr/share/sysbench/oltp_read_only.lua \
> > --mysql-user=USER --mysql-password=PASS --mysql-db=DB \
> > --tables=48 --table-size=2000000 --threads=16 --time=1800 run
> >
> > Before this patch
> > Avg: 134743.714545 qps. Stdev: 582.242189
> >
> > After this patch (+0.2%):
> > Avg: 135005.779091 qps. Stdev: 295.299027
> >
> > Test 4: Build linux kernel in 2G memcg with make -j48 with SSD swap
> > (for memory stress, 18 times):
> >
> > Before this patch:
> > Avg: 1456.768899 s. Stdev: 20.106973
> >
> > After this patch (+0.0%):
> > Avg: 1455.659254 s. Stdev: 15.274481
> >
> > Test 5: Memtier test in a 4G cgroup using brd as swap (18 times):
> > memcached -u nobody -m 16384 -s /tmp/memcached.socket \
> > -a 0766 -t 16 -B binary &
> > memtier_benchmark -S /tmp/memcached.socket \
> > -P memcache_binary -n allkeys \
> > --key-minimum=1 --key-maximum=16000000 -d 1024 \
> > --ratio=1:0 --key-pattern=P:P -c 1 -t 16 --pipeline 8 -x 3
> >
> > Before this patch:
> > Avg: 50317.984000 Ops/sec. Stdev: 2568.965458
> >
> > After this patch (-5.7%):
> > Avg: 47691.343500 Ops/sec. Stdev: 3925.772473
> >
> > It seems prefetch is helpful in most cases, but the memtier test is
> > either hitting a case where prefetch causes higher cache miss or it's
> > just too noisy (high stdev).
> >
> > Signed-off-by: Kairui Song <kasong@xxxxxxxxxxx>
> > ---
> > mm/vmscan.c | 30 ++++++++++++++++++++++++++----
> > 1 file changed, 26 insertions(+), 4 deletions(-)
> >
> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > index 4f9c854ce6cc..03631cedb3ab 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -3681,15 +3681,26 @@ static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
> > /* prevent cold/hot inversion if force_scan is true */
> > for (zone = 0; zone < MAX_NR_ZONES; zone++) {
> > struct list_head *head = &lrugen->folios[old_gen][type][zone];
> > + struct folio *prev = NULL;
> >
> > - while (!list_empty(head)) {
> > - struct folio *folio = lru_to_folio(head);
> > + if (!list_empty(head))
> > + prev = lru_to_folio(head);
> > +
> > + while (prev) {
> > + struct folio *folio = prev;
> >
> > VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
> > VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
> > VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
> > VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
> >
> > + if (unlikely(list_is_first(&folio->lru, head))) {
> > + prev = NULL;
> > + } else {
> > + prev = lru_to_folio(&folio->lru);
> > + prefetchw(&prev->flags);
> > + }
>
> This makes the code flow much harder to follow. Also for architecture
> that does not support prefetch, this will be a net loss.
>
> Can you use refetchw_prev_lru_folio() instead? It will make the code
> much easier to follow. It also turns into no-op when prefetch is not
> supported.
>
> Chris
>
Hi Chris,
Thanks for the suggestion.
Yes, that's doable, I made it this way because in previous series (V1
& V2) I applied the bulk move patch first which needed and introduced
the `prev` variable here, so the prefetch logic just used it.
For V3 I did a rebase and moved the prefetch commit to be the first
one, since it seems to be the most effective one, and just kept the
code style to avoid redundant change between patches.
I can update in V4 to make this individual patch better with your suggestion.
