The page allocation performance requirements of different workloads
are often different. So, we need to tune the PCP (Per-CPU Pageset)
high on each CPU automatically to optimize the page allocation
performance.
The list of patches in series is as follows,
[1/9] mm, pcp: avoid to drain PCP when process exit
[2/9] cacheinfo: calculate per-CPU data cache size
[3/9] mm, pcp: reduce lock contention for draining high-order pages
[4/9] mm: restrict the pcp batch scale factor to avoid too long latency
[5/9] mm, page_alloc: scale the number of pages that are batch allocated
[6/9] mm: add framework for PCP high auto-tuning
[7/9] mm: tune PCP high automatically
[8/9] mm, pcp: decrease PCP high if free pages < high watermark
[9/9] mm, pcp: reduce detecting time of consecutive high order page freeing
Patch [1/9], [2/9], [3/9] optimize the PCP draining for consecutive
high-order pages freeing.
Patch [4/9], [5/9] optimize batch freeing and allocating.
Patch [6/9], [7/9], [8/9] implement and optimize a PCP high
auto-tuning method.
Patch [9/9] optimize the PCP draining for consecutive high order page
freeing based on PCP high auto-tuning.
The test results for patches with performance impact are as follows,
kbuild
======
On a 2-socket Intel server with 224 logical CPU, we run 8 kbuild
instances in parallel (each with `make -j 28`) in 8 cgroup. This
simulates the kbuild server that is used by 0-Day kbuild service.
build time lock contend% free_high alloc_zone
---------- ---------- --------- ----------
base 100.0 14.0 100.0 100.0
patch1 99.5 12.8 19.5 95.6
patch3 99.4 12.6 7.1 95.6
patch5 98.6 11.0 8.1 97.1
patch7 95.1 0.5 2.8 15.6
patch9 95.0 1.0 8.8 20.0
The PCP draining optimization (patch [1/9], [3/9]) and PCP batch
allocation optimization (patch [5/9]) reduces zone lock contention a
little. The PCP high auto-tuning (patch [7/9], [9/9]) reduces build
time visibly. Where the tuning target: the number of pages allocated
from zone reduces greatly. So, the zone contention cycles% reduces
greatly.
With PCP tuning patches (patch [7/9], [9/9]), the average used memory
during test increases up to 18.4% because more pages are cached in
PCP. But at the end of the test, the number of the used memory
decreases to the same level as that of the base patch. That is, the
pages cached in PCP will be released to zone after not being used
actively.
netperf SCTP_STREAM_MANY
========================
On a 2-socket Intel server with 128 logical CPU, we tested
SCTP_STREAM_MANY test case of netperf test suite with 64-pair
processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 2.1 100.0 100.0 1.3
patch1 99.4 2.1 99.4 99.4 1.3
patch3 106.4 1.3 13.3 106.3 1.3
patch5 106.0 1.2 13.2 105.9 1.3
patch7 103.4 1.9 6.7 90.3 7.6
patch9 108.6 1.3 13.7 108.6 1.3
The PCP draining optimization (patch [1/9]+[3/9]) improves
performance. The PCP high auto-tuning (patch [7/9]) reduces
performance a little because PCP draining cannot be triggered in time
sometimes. So, the cache miss rate% increases. The further PCP
draining optimization (patch [9/9]) based on PCP tuning restore the
performance.
lmbench3 UNIX (AF_UNIX)
=======================
On a 2-socket Intel server with 128 logical CPU, we tested UNIX
(AF_UNIX socket) test case of lmbench3 test suite with 16-pair
processes.
score lock contend% free_high alloc_zone cache miss rate%
----- ---------- --------- ---------- ----------------
base 100.0 51.4 100.0 100.0 0.2
patch1 116.8 46.1 69.5 104.3 0.2
patch3 199.1 21.3 7.0 104.9 0.2
patch5 200.0 20.8 7.1 106.9 0.3
patch7 191.6 19.9 6.8 103.8 2.8
patch9 193.4 21.7 7.0 104.7 2.1
The PCP draining optimization (patch [1/9], [3/9]) improves
performance much. The PCP tuning (patch [7/9]) reduces performance a
little because PCP draining cannot be triggered in time sometimes.
The further PCP draining optimization (patch [9/9]) based on PCP
tuning restores the performance partly.
The patchset adds several fields in struct per_cpu_pages. The struct
layout before/after the patchset is as follows,
base
====
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int batch; /* 12 4 */
short int free_factor; /* 16 2 */
short int expire; /* 18 2 */
/* XXX 4 bytes hole, try to pack */
struct list_head lists[13]; /* 24 208 */
/* size: 256, cachelines: 4, members: 7 */
/* sum members: 228, holes: 1, sum holes: 4 */
/* padding: 24 */
} __attribute__((__aligned__(64)));
patched
=======
struct per_cpu_pages {
spinlock_t lock; /* 0 4 */
int count; /* 4 4 */
int high; /* 8 4 */
int high_min; /* 12 4 */
int high_max; /* 16 4 */
int batch; /* 20 4 */
u8 flags; /* 24 1 */
u8 alloc_factor; /* 25 1 */
u8 expire; /* 26 1 */
/* XXX 1 byte hole, try to pack */
short int free_count; /* 28 2 */
/* XXX 2 bytes hole, try to pack */
struct list_head lists[13]; /* 32 208 */
/* size: 256, cachelines: 4, members: 11 */
/* sum members: 237, holes: 2, sum holes: 3 */
/* padding: 16 */
} __attribute__((__aligned__(64)));
The size of the struct doesn't changed with the patchset.
Changelog:
v3:
- Fix the per-CPU data cache slice size calculation in [3/9]. Thanks
Mel for pointing this out!
- Fix a PCP->high decrement amount issue when free pages become low.
- Dropped original "[9/10] mm, pcp: avoid to reduce PCP high
unnecessarily", because of no enough performance data to support it.
- Add some comments per Mel's comments. Thanks!
- Add a Kconfig option for max batch scale factor in [4/9] per Mel's
comments. Thanks!
- Collected Acked-by. Thanks Mel!
v2:
- Fix the kbuild test configuration and results. Thanks Andrew for
reminding on test results!
- Add document for sysctl behavior extension in [06/10] per Andrew's comments.
Best Regards,
Huang, Ying
