Page replacement is handled in the Linux Kernel in one of two ways:
1) Asynchronously via kswapd
2) Synchronously, via direct reclaim
At page allocation time the allocating task is immediately given a page
from the zone free list allowing it to go right back to work doing
whatever it was doing; Probably directly or indirectly executing business
logic.
Just prior to satisfying the allocation, free pages is checked to see if
it has reached the zone low watermark and if so, kswapd is awakened.
Kswapd will start scanning pages looking for inactive pages to evict to
make room for new page allocations. The work of kswapd allows tasks to
continue allocating memory from their respective zone free list without
incurring any delay.
When the demand for free pages exceeds the rate that kswapd tasks can
supply them, page allocation works differently. Once the allocating task
finds that the number of free pages is at or below the zone min watermark,
the task will no longer pull pages from the free list. Instead, the task
will run the same CPU-bound routines as kswapd to satisfy its own
allocation by scanning and evicting pages. This is called a direct reclaim.
The time spent performing a direct reclaim can be substantial, often
taking tens to hundreds of milliseconds for small order0 allocations to
half a second or more for order9 huge-page allocations. In fact, kswapd is
not actually required on a linux system. It exists for the sole purpose of
optimizing performance by preventing direct reclaims.
When memory shortfall is sufficient to trigger direct reclaims, they can
occur in any task that is running on the system. A single aggressive
memory allocating task can set the stage for collateral damage to occur in
small tasks that rarely allocate additional memory. Consider the impact of
injecting an additional 100ms of latency when nscd allocates memory to
facilitate caching of a DNS query.
The presence of direct reclaims 10 years ago was a fairly reliable
indicator that too much was being asked of a Linux system. Kswapd was
likely wasting time scanning pages that were ineligible for eviction.
Adding RAM or reducing the working set size would usually make the problem
go away. Since then hardware has evolved to bring a new struggle for
kswapd. Storage speeds have increased by orders of magnitude while CPU
clock speeds stayed the same or even slowed down in exchange for more
cores per package. This presents a throughput problem for a single
threaded kswapd that will get worse with each generation of new hardware.
Test Details
NOTE: The tests below were run with shadow entries disabled. See the
associated patch and cover letter for details
The tests below were designed with the assumption that a kswapd bottleneck
is best demonstrated using filesystem reads. This way, the inactive list
will be full of clean pages, simplifying the analysis and allowing kswapd
to achieve the highest possible steal rate. Maximum steal rates for kswapd
are likely to be the same or lower for any other mix of page types on the
system.
Tests were run on a 2U Oracle X7-2L with 52 Intel Xeon Skylake 2GHz cores,
756GB of RAM and 8 x 3.6 TB NVMe Solid State Disk drives. Each drive has
an XFS file system mounted separately as /d0 through /d7. SSD drives
require multiple concurrent streams to show their potential, so I created
eleven 250GB zero-filled files on each drive so that I could test with
parallel reads.
The test script runs in multiple stages. At each stage, the number of dd
tasks run concurrently is increased by 2. I did not include all of the
test output for brevity.
During each stage dd tasks are launched to read from each drive in a round
robin fashion until the specified number of tasks for the stage has been
reached. Then iostat, vmstat and top are started in the background with 10
second intervals. After five minutes, all of the dd tasks are killed and
the iostat, vmstat and top output is parsed in order to report the
following:
CPU consumption
- sy - aggregate kernel mode CPU consumption from vmstat output. The value
doesn't tend to fluctuate much so I just grab the highest value.
Each sample is averaged over 10 seconds
- dd_cpu - for all of the dd tasks averaged across the top samples since
there is a lot of variation.
Throughput
- in Kbytes
- Command is iostat -x -d 10 -g total
This first test performs reads using O_DIRECT in order to show the maximum
throughput that can be obtained using these drives. It also demonstrates
how rapidly throughput scales as the number of dd tasks are increased.
The dd command for this test looks like this:
Command Used: dd iflag=direct if=/d${i}/$n of=/dev/null bs=4M
Test #1: Direct IO
dd sy dd_cpu throughput
6 0 2.33 14726026.40
10 1 2.95 19954974.80
16 1 2.63 24419689.30
22 1 2.63 25430303.20
28 1 2.91 26026513.20
34 1 2.53 26178618.00
40 1 2.18 26239229.20
46 1 1.91 26250550.40
52 1 1.69 26251845.60
58 1 1.54 26253205.60
64 1 1.43 26253780.80
70 1 1.31 26254154.80
76 1 1.21 26253660.80
82 1 1.12 26254214.80
88 1 1.07 26253770.00
90 1 1.04 26252406.40
Throughput was close to peak with only 22 dd tasks. Very little system CPU
was consumed as expected as the drives DMA directly into the user address
space when using direct IO.
In this next test, the iflag=direct option is removed and we only run the
test until the pgscan_kswapd from /proc/vmstat starts to increment. At
that point metrics are parsed and reported and the pagecache contents are
dropped prior to the next test. Lather, rinse, repeat.
Test #2: standard file system IO, no page replacement
dd sy dd_cpu throughput
6 2 28.78 5134316.40
10 3 31.40 8051218.40
16 5 34.73 11438106.80
22 7 33.65 14140596.40
28 8 31.24 16393455.20
34 10 29.88 18219463.60
40 11 28.33 19644159.60
46 11 25.05 20802497.60
52 13 26.92 22092370.00
58 13 23.29 22884881.20
64 14 23.12 23452248.80
70 15 22.40 23916468.00
76 16 22.06 24328737.20
82 17 20.97 24718693.20
88 16 18.57 25149404.40
90 16 18.31 25245565.60
Each read has to pause after the buffer in kernel space is populated while
those pages are added to the pagecache and copied into the user address
space. For this reason, more parallel streams are required to achieve peak
throughput. The copy operation consumes substantially more CPU than direct
IO as expected.
The next test measures throughput after kswapd starts running. This is the
same test only we wait for kswapd to wake up before we start collecting
metrics. The script actually keeps track of a few things that were not
mentioned earlier. It tracks direct reclaims and page scans by watching
the metrics in /proc/vmstat. CPU consumption for kswapd is tracked the
same way it is tracked for dd.
Since the test is 100% reads, you can assume that the page steal rate for
kswapd and direct reclaims is almost identical to the scan rate.
Test #3: 1 kswapd thread per node
dd sy dd_cpu kswapd0 kswapd1 throughput dr pgscan_kswapd pgscan_direct
10 4 26.07 28.56 27.03 7355924.40 0 459316976 0
16 7 34.94 69.33 69.66 10867895.20 0 872661643 0
22 10 36.03 93.99 99.33 13130613.60 489 1037654473 11268334
28 10 30.34 95.90 98.60 14601509.60 671 1182591373 15429142
34 14 34.77 97.50 99.23 16468012.00 10850 1069005644 249839515
40 17 36.32 91.49 97.11 17335987.60 18903 975417728 434467710
46 19 38.40 90.54 91.61 17705394.40 25369 855737040 582427973
52 22 40.88 83.97 83.70 17607680.40 31250 709532935 724282458
58 25 40.89 82.19 80.14 17976905.60 35060 657796473 804117540
64 28 41.77 73.49 75.20 18001910.00 39073 561813658 895289337
70 33 45.51 63.78 64.39 17061897.20 44523 379465571 1020726436
76 36 46.95 57.96 60.32 16964459.60 47717 291299464 1093172384
82 39 47.16 55.43 56.16 16949956.00 49479 247071062 1134163008
88 42 47.41 53.75 47.62 16930911.20 51521 195449924 1180442208
90 43 47.18 51.40 50.59 16864428.00 51618 190758156 1183203901
In the previous test where kswapd was not involved, the system-wide kernel
mode CPU consumption with 90 dd tasks was 16%. In this test CPU consumption
with 90 tasks is at 43%. With 52 cores, and two kswapd tasks (one per NUMA
node), kswapd can only be responsible for a little over 4% of the increase.
The rest is likely caused by 51,618 direct reclaims that scanned 1.2
billion pages over the five minute time period of the test.
Same test, more kswapd tasks:
Test #4: 4 kswapd threads per node
dd sy dd_cpu kswapd0 kswapd1 throughput dr pgscan_kswapd pgscan_direct
10 5 27.09 16.65 14.17 7842605.60 0 459105291 0
16 10 37.12 26.02 24.85 11352920.40 15 920527796 358515
22 11 36.94 37.13 35.82 13771869.60 0 1132169011 0
28 13 35.23 48.43 46.86 16089746.00 0 1312902070 0
34 15 33.37 53.02 55.69 18314856.40 0 1476169080 0
40 19 35.90 69.60 64.41 19836126.80 0 1629999149 0
46 22 36.82 88.55 57.20 20740216.40 0 1708478106 0
52 24 34.38 93.76 68.34 21758352.00 0 1794055559 0
58 24 30.51 79.20 82.33 22735594.00 0 1872794397 0
64 26 30.21 97.12 76.73 23302203.60 176 1916593721 4206821
70 33 32.92 92.91 92.87 23776588.00 3575 1817685086 85574159
76 37 31.62 91.20 89.83 24308196.80 4752 1812262569 113981763
82 29 25.53 93.23 92.33 24802791.20 306 2032093122 7350704
88 43 37.12 76.18 77.01 25145694.40 20310 1253204719 487048202
90 42 38.56 73.90 74.57 22516787.60 22774 1193637495 545463615
By increasing the number of kswapd threads, throughput increased by ~50%
while kernel mode CPU utilization decreased or stayed the same, likely due
to a decrease in the number of parallel tasks at any given time doing page
replacement.
Signed-off-by: Buddy Lumpkin <buddy.lumpkin@oracle.com>
Bug: 201263306
Link: https://lore.kernel.org/lkml/1522661062-39745-1-git-send-email-buddy.lumpkin@oracle.com
[charante@codeaurora.org]: Changes made to select number of kswapds through uapi
Signed-off-by: Charan Teja Reddy <charante@codeaurora.org>
[quic_vjitta@quicinc.com]: Changes made to move multiple kswapd threads logic to vendor hooks
Signed-off-by: Vijayanand Jitta <quic_vjitta@quicinc.com>
(cherry picked from commit 0d61a651e4dd3c61d1658cc92e0b0450c8374738)
Change-Id: I8425cab7f40cbeaf65af0ea118c1a9ac7da0930e
[quic_vjitta@quicinc.com]: Resolved minor merge conflicts
Signed-off-by: Vijayanand Jitta <quic_vjitta@quicinc.com>