mm: page_counter: relayout structure to reduce false sharing
Subsystems (3)
PATCH v2 1/1 mm: page_counter: relayout structure to reduce false sharing
When checking a memory cgroup related performance regression [1], from the perf c2c profiling data, we found high false sharing for accessing 'usage' and 'parent'.
On 64 bit system, the 'usage' and 'parent' are close to each other, and easy to be in one cacheline (for cacheline size == 64+ B). 'usage' is usally written, while 'parent' is usually read as the cgroup's hierarchical counting nature.
So move the 'parent' to the end of the structure to make sure they are in different cache lines.
Following are some performance data with the patch, against v5.11-rc1. [ In the data, A means a platform with 2 sockets 48C/96T, B is a platform of 4 sockests 72C/144T, and if a %stddev will be shown bigger than 2%, P100/P50 means number of test tasks equals to 100%/50% of nr_cpu]
will-it-scale/malloc1
---------------------
v5.11-rc1 v5.11-rc1+patch
A-P100 15782 ± 2% -0.1% 15765 ± 3% will-it-scale.per_process_ops
A-P50 21511 +8.9% 23432 will-it-scale.per_process_ops
B-P100 9155 +2.2% 9357 will-it-scale.per_process_ops
B-P50 10967 +7.1% 11751 ± 2% will-it-scale.per_process_ops
will-it-scale/pagefault2
------------------------
v5.11-rc1 v5.11-rc1+patch
A-P100 79028 +3.0% 81411 will-it-scale.per_process_ops A-P50 183960 ± 2% +4.4% 192078 ± 2% will-it-scale.per_process_ops B-P100 85966 +9.9% 94467 ± 3% will-it-scale.per_process_ops
B-P50 198195 +9.8% 217526 will-it-scale.per_process_ops
fio (4k/1M is block size)
-------------------------
v5.11-rc1 v5.11-rc1+patch
A-P50-r-4k 16881 ± 2% +1.2% 17081 ± 2% fio.read_bw_MBps
A-P50-w-4k 3931 +4.5% 4111 ± 2% fio.write_bw_MBps A-P50-r-1M 15178 -0.2% 15154 fio.read_bw_MBps A-P50-w-1M 3924 +0.1% 3929 fio.write_bw_MBps
[1].https://lore.kernel.org/lkml/20201102091543.GM31092@shao2-debian/
---
Changelogs:
v2:
* Adjust the format of performance data to be more readable, as suggested by Michal Hocko
include/linux/page_counter.h | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
include/linux/page_counter.h +8 -1
@@ -12,7 +12,6 @@ struct page_counter {unsigned long low;unsigned long high;unsigned long max; -struct page_counter *parent;/* effective memory.min and memory.min usage tracking */ unsigned long emin;
@@ -27,6 +26,14 @@ struct page_counter {/* legacy */ unsigned long watermark;unsigned long failcnt; + +/* +* 'parent' is placed here to be far from 'usage' to reduce +* cache false sharing, as 'usage' is written mostly while +* parent is frequently read for cgroup's hierarchical +* counting nature. +*/ +struct page_counter *parent; };#if BITS_PER_LONG ==32
Signature · 1 line
2.7.4
Quoted 46 lines · Feng Tang
When checking a memory cgroup related performance regression [1], from the perf c2c profiling data, we found high false sharing for accessing 'usage' and 'parent'.
On 64 bit system, the 'usage' and 'parent' are close to each other, and easy to be in one cacheline (for cacheline size == 64+ B). 'usage' is usally written, while 'parent' is usually read as the cgroup's hierarchical counting nature.
So move the 'parent' to the end of the structure to make sure they are in different cache lines.
Following are some performance data with the patch, against v5.11-rc1. [ In the data, A means a platform with 2 sockets 48C/96T, B is a platform of 4 sockests 72C/144T, and if a %stddev will be shown bigger than 2%, P100/P50 means number of test tasks equals to 100%/50% of
nr_cpu]will-it-scale/malloc1
---------------------v5.11-rc1 v5.11-rc1+patchA-P100 15782 ± 2% -0.1% 15765 ± 3% will-it-scale.per_process_ops
A-P50 21511 +8.9% 23432 will-it-scale.per_process_ops B-P100 9155 +2.2% 9357 will-it-scale.per_process_ops B-P50 10967 +7.1% 11751 ± 2% will-it-scale.per_process_opswill-it-scale/pagefault2
------------------------v5.11-rc1 v5.11-rc1+patchA-P100 79028 +3.0% 81411 will-it-scale.per_process_ops A-P50 183960 ± 2% +4.4% 192078 ± 2% will-it-scale.per_process_ops B-P100 85966 +9.9% 94467 ± 3% will-it-scale.per_process_opsB-P50 198195 +9.8% 217526 will-it-scale.per_process_ops
fio (4k/1M is block size)
-------------------------v5.11-rc1 v5.11-rc1+patchA-P50-r-4k 16881 ± 2% +1.2% 17081 ± 2% fio.read_bw_MBps
A-P50-w-4k 3931 +4.5% 4111 ± 2% fio.write_bw_MBps A-P50-r-1M 15178 -0.2% 15154 fio.read_bw_MBps A-P50-w-1M 3924 +0.1% 3929 fio.write_bw_MBps
Quoted 5 lines
[1].https://lore.kernel.org/lkml/20201102091543.GM31092@shao2-debian/
Signed-off-by Feng Tang <[email protected]>Reviewed-by Roman Gushchin <[email protected]>Cc Johannes Weiner <[email protected]>Cc Michal Hocko <[email protected]>
Quoted 45 lines · Feng Tang
When checking a memory cgroup related performance regression [1], from the perf c2c profiling data, we found high false sharing for accessing 'usage' and 'parent'.
On 64 bit system, the 'usage' and 'parent' are close to each other, and easy to be in one cacheline (for cacheline size == 64+ B). 'usage' is usally written, while 'parent' is usually read as the cgroup's hierarchical counting nature.
So move the 'parent' to the end of the structure to make sure they are in different cache lines.
Following are some performance data with the patch, against v5.11-rc1. [ In the data, A means a platform with 2 sockets 48C/96T, B is a platform of 4 sockests 72C/144T, and if a %stddev will be shown bigger than 2%, P100/P50 means number of test tasks equals to 100%/50% of
nr_cpu]will-it-scale/malloc1
---------------------v5.11-rc1 v5.11-rc1+patchA-P100 15782 ± 2% -0.1% 15765 ± 3% will-it-scale.per_process_ops
A-P50 21511 +8.9% 23432 will-it-scale.per_process_ops
B-P100 9155 +2.2% 9357 will-it-scale.per_process_ops
B-P50 10967 +7.1% 11751 ± 2% will-it-scale.per_process_opswill-it-scale/pagefault2
------------------------v5.11-rc1 v5.11-rc1+patchA-P100 79028 +3.0% 81411 will-it-scale.per_process_ops A-P50 183960 ± 2% +4.4% 192078 ± 2% will-it-scale.per_process_ops B-P100 85966 +9.9% 94467 ± 3% will-it-scale.per_process_opsB-P50 198195 +9.8% 217526 will-it-scale.per_process_ops
fio (4k/1M is block size)
-------------------------v5.11-rc1 v5.11-rc1+patchA-P50-r-4k 16881 ± 2% +1.2% 17081 ± 2% fio.read_bw_MBps
A-P50-w-4k 3931 +4.5% 4111 ± 2% fio.write_bw_MBps A-P50-r-1M 15178 -0.2% 15154 fio.read_bw_MBps A-P50-w-1M 3924 +0.1% 3929 fio.write_bw_MBps
Quoted 5 lines
[1].https://lore.kernel.org/lkml/20201102091543.GM31092@shao2-debian/
Signed-off-by Feng Tang <[email protected]>Reviewed-by Roman Gushchin <[email protected]>Cc Johannes Weiner <[email protected]>Cc Michal Hocko <[email protected]>
Thanks!
Quoted 36 lines · Feng Tang
When checking a memory cgroup related performance regression [1], from the perf c2c profiling data, we found high false sharing for accessing 'usage' and 'parent'.
On 64 bit system, the 'usage' and 'parent' are close to each other, and easy to be in one cacheline (for cacheline size == 64+ B). 'usage' is usally written, while 'parent' is usually read as the cgroup's hierarchical counting nature.
So move the 'parent' to the end of the structure to make sure they are in different cache lines.
Following are some performance data with the patch, against v5.11-rc1. [ In the data, A means a platform with 2 sockets 48C/96T, B is a platform of 4 sockests 72C/144T, and if a %stddev will be shown bigger than 2%, P100/P50 means number of test tasks equals to 100%/50% of
nr_cpu]will-it-scale/malloc1
---------------------v5.11-rc1 v5.11-rc1+patchA-P100 15782 ± 2% -0.1% 15765 ± 3% will-it-scale.per_process_ops
A-P50 21511 +8.9% 23432 will-it-scale.per_process_ops
B-P100 9155 +2.2% 9357 will-it-scale.per_process_ops
B-P50 10967 +7.1% 11751 ± 2% will-it-scale.per_process_opswill-it-scale/pagefault2
------------------------v5.11-rc1 v5.11-rc1+patchA-P100 79028 +3.0% 81411 will-it-scale.per_process_ops A-P50 183960 ± 2% +4.4% 192078 ± 2% will-it-scale.per_process_ops B-P100 85966 +9.9% 94467 ± 3% will-it-scale.per_process_opsB-P50 198195 +9.8% 217526 will-it-scale.per_process_ops
Quoted 8 lines
fio (4k/1M is block size)
-------------------------v5.11-rc1 v5.11-rc1+patchA-P50-r-4k 16881 ± 2% +1.2% 17081 ± 2% fio.read_bw_MBps
A-P50-w-4k 3931 +4.5% 4111 ± 2% fio.write_bw_MBps A-P50-r-1M 15178 -0.2% 15154 fio.read_bw_MBps A-P50-w-1M 3924 +0.1% 3929 fio.write_bw_MBps
Thanks for making results easier to read and understand.
Quoted 5 lines
[1].https://lore.kernel.org/lkml/20201102091543.GM31092@shao2-debian/
Signed-off-by Feng Tang <[email protected]>Reviewed-by Roman Gushchin <[email protected]>Cc Johannes Weiner <[email protected]>Cc Michal Hocko <[email protected]>
Thanks!
Quoted 40 lines
---
Changelogs:v2:
* Adjust the format of performance data to be more readable, as suggested by Michal Hockoinclude/linux/page_counter.h | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)include/linux/page_counter.h +8
-1@@ -12,7 +12,6 @@ struct page_counter {unsigned long low;unsigned long high;unsigned long max; -struct page_counter *parent;/* effective memory.min and memory.min usage tracking */ unsigned long emin;@@ -27,6 +26,14 @@ struct page_counter {/* legacy */ unsigned long watermark;unsigned long failcnt; + +/* +* 'parent' is placed here to be far from 'usage' to reduce +* cache false sharing, as 'usage' is written mostly while +* parent is frequently read for cgroup's hierarchical +* counting nature. +*/ +struct page_counter *parent; };#if BITS_PER_LONG ==32 Signature · 1 line
2.7.4
Signature · 2 lines
Michal Hocko SUSE Labs