Sorry, you need to enable JavaScript to visit this website.

Linux Kernel Performance

Linux development evolves rapidly. The performance and scalability of the OS kernel has been a key part of its success. However, discussions have appeared on LKML (Linux Kernel Mailing List) regarding large performance regression between kernel versions. These discussions underscore the need for a systematic and disciplined way to characterize, improve, and test Linux kernel performance. Our goal is to work with the Linux community to further enhance the Linux kernel with consistent performance increases (avoiding degradations) across releases. The information available on this site gives community members better information about what 0-Day and LKP (Linux Kernel Performance) are doing to preserve performance integrity of the kernel.

0-DAY CI LINUX KERNEL PERFORMANCE REPORT (V5.15)

BY Beibei Si ON Nov 04, 2021

 

1.            Introduction

0-Day CI is an automated Linux kernel test service that provides comprehensive test coverage of the Linux kernel. It covers kernel build, static analysis, boot, functional, performance and power tests. This report shows the recent observations of kernel performance status on IA platform based on the test results from 0-Day CI service. It is structured in the following manner:

        Section 2, test parameter description

        Section 3, merged regressions and improvements in v5.15 release candidates

        Section 4, captured regressions and improvements by shift-left testing during developers’ and maintainers’ tree during v5.15 release cycle

        Section 5, performance comparison among different kernel releases

        Section 6, test machine list

 

2.            Test parameters descriptions

Here are the descriptions for each parameter/field used in the tests.

 

Classification

Name

Description

General

runtime

Run the test case within a certain time period (seconds or minutes)

 

nr_task

If it is an integer, which means the number of processes/threads (to run the workload) of this job. Default is 1.

If it is a percentage, e.g. 200% means the number of processes/threads is double of cpu number

 

nr_threads

Alias of nr_task

 

iterations

Number to repeat this job

 

test_size

Test disk size or memory size

 

set_nic_irq_affinity

Set NIC interrupt affinity

 

disable_latency_stats

Latency_stats may introduce too much noise if there are too many context switches, allow to disable it

 

 

transparent_hugepage

Set transparent hugepage policy (/sys/kernel/mm/transparent_hugepage)

 

boot_params:bp1_memmap

Boot parameters of memmap

 

disk:nr_pmem

number of pmem partitions used by test

 

swap:priority

Priority means the  priority  of  the  swap device. priority is a value between -1 and 32767, the default is -1 and higher priority with higher value.

Test Machine

model

Name of Intel processor microarchitecture

 

brand

Brand name of cpu

 

cpu_number

Number of cpu

 

memory

Size of memory

 

3.            Linux Kernel v5.15 Release Test

Linus has released the 5.15 kernel after another nine-week development cycle, and mentioned “This release may have started out with some -Werror pain, but it calmed down fairly quickly and on the whole 5.15 was fair small and calm. Let's hope for more of the same - without Werror issues this time - for the upcoming merge window.” The code name for this release has been set to "Trick or Treat". Significant features in this release include: the realtime preemption locking code, descriptorless files for io_uring, BPF timers, the removal of mandatory file-locking support, the ksmbd SMB filesystem server (but see this article), printk() indexing, the process_mrelease() system call, The DAMON memory-management optimization system, the ntfs3 filesystem implementation, and much more. See the KernelNewbies 5.15 page for more information.

0-Day CI monitored the release closely to trace down the performance status on IA platform. 0-Day observed 5 regressions and 7 improvements during feature development phase for v5.15. We will share more detailed information together with correlated patches that led to the results. Note that the assessment is limited by the test coverage 0-Day has now. The list is summarized in the observation summary section.

3.1.            Observation Summary

0-Day CI observed 5 regressions and 7 improvements during the feature development phase for v5.15, which is in the time frame from v5.15-rc1 to v5.15 release.

Test Indicator

Mail

Test Scenario

Test Machine

Development Base

Status

aim7.jobs-per-min

[xfs] ab23a77687: -18.5% regression

disk: 1BRD_48G

fs: xfs

test: disk_rr

load: 3000

cpufreq_governor: performance

lkp-cpl-4sp1

v5.14-rc4

merged at v5.15-rc1, author doubt the regression and 0-Day CI team is following up

fxmark.hdd_ext4_no_jnl_DWTL_1_directio.works/sec

[loop] 2112f5c133: -49.6% regression

disk: 1HDD

media: hdd

test: DWTL

fstype: ext4_no_jnl

directio: directio

cpufreq_governor: performance

lkp-knm02

v5.14-rc4

merged at v5.15-rc1, no response from author yet

netperf.Throughput_tps

[sched/fair] 56498cfb04: -5.4% regression

ip: ipv4

runtime: 300s

nr_threads: 50%

cluster: cs-localhost

test: UDP_RR

cpufreq_governor: performance

lkp-icl-2sp2

v5.13-rc6

merged at v5.15-rc1, no response from author yet

netperf.Throughput_tps

[bpf] b89fbfbb85: -21.3% regression

ip: ipv4

runtime: 300s

nr_threads: 16

cluster: cs-localhost

test: TCP_CRR

cpufreq_governor: performance

lkp-csl-2ap3

v5.14-rc5

merged at v5.15-rc1, author doubt the regression and 0-Day CI team is following up

will-it-scale.per_process_ops

[memcg] 0f12156dff: -33.6% regression

nr_task: 50%

mode: process

test: lock1

cpufreq_governor: performance

lkp-skl-fpga01

v5.14

merged at v5.15-rc1, author accepted the regression and the patch was reverted on latest kernel tree

Improvement

 

 

 

 

 

aim7.jobs-per-min

[ext4] cc883236b7: 69.4% improvement

disk: 4BRD_12G

md: RAID0

fs: ext4

test: disk_rw

load: 3000

cpufreq_governor: performance

lkp-csl-2sp9

v5.14-rc2

merged at v5.15-rc4

fsmark.files_per_sec

[SUNRPC] e38b3f2005: 1857.1% improvement

iterations: 1x

nr_threads: 1t

disk: 1BRD_48G

fs: f2fs

fs2: nfsv4

filesize: 4M

test_size: 24G

sync_method: NoSync

cpufreq_governor: performance

lkp-csl-2ap2

v5.14-rc6

merged at v5.15-rc1

stress-ng.memhotplug.ops_per_sec

[mm/migrate] 295be91f7e: 151.7% improvement

nr_threads: 10%

disk: 1HDD

testtime: 60s

fs: ext4

class: os

test: memhotplug

cpufreq_governor: performance

lkp-csl-2sp5

v5.15-rc6

merged at v5.15-rc7

stress-ng.netdev.ops_per_sec

[net] b0e99d0377: 7.7% improvement

nr_threads: 100%

testtime: 60s

class: network

test: netdev

cpufreq_governor: performance

lkp-csl-2sp5

v5.14-rc1

merged at v5.15-rc1

will-it-scale.per_process_ops

[mm] 243418e392: 3.0% improvement

nr_task: 100%

mode: process

test: brk1

cpufreq_governor: performance

lkp-csl-2ap2

v5.15-rc2

merged at v5.15-rc3

will-it-scale.per_process_ops

[memcg] 1f828223b7: 37.3% improvement

nr_task: 100%

mode: process

test: page_fault3

thp_enabled: never

cpufreq_governor: performance

lkp-csl-2sp5

v5.15-rc2

merged at v5.15-rc3

will-it-scale.per_thread_ops

[fsnotify] e43de7f086: 10.2% improvement

nr_task: 100%

mode: thread

test: eventfd1

cpufreq_governor: performance

lkp-csl-2ap2

v5.14-rc4

merged at v5.15-rc1

3.2.            Will-it-scale.per_process_ops

 

Will-it-scale takes a test case and runs it from 1 through to n parallel copies to see if the test case will scale. It builds both process and threads based tests in order to see any differences between the two.

3.2.1   Scenario: process lock1

 

Commit 0f12156dff was reported to have -33.6% regression of will-it-scale.per_process_ops when comparing to v5.14. It was merged to mainline at v5.15-rc1.

 

Correlated commits

0f12156dff

memcg: enable accounting for file lock caches

branch

linus/master

report

[memcg] 0f12156dff: -33.6% regression

test scenario

nr_task: 50%

mode: process

test: lock1

cpufreq_governor: performance

test machine

lkp-skl-fpga01

status

merged at v5.15-rc1, author accepted the regression and the patch was reverted on latest kernel tree

 

3.3.            Stress-ng.memhotplug.ops_per_sec

stress-ng will stress test a computer system in various selectable ways. It was designed to exercise various physical subsystems of a computer as well as the various operating system kernel interfaces.

3.2.2   Scenario: memhotplug test on ext4

Commit 295be91f7e was reported to have 151.7% improvement of stress-ng.memhotplug.ops_per_sec when comparing to v5.15-rc6. It was merged to mainline at v5.15-rc7.

 

Correlated commits

295be91f7e

mm/migrate: optimize hotplug-time demotion order updates

branch

linus/master

report

[mm/migrate] 295be91f7e: 151.7% improvement

test scenario

nr_threads: 10%

disk: 1HDD

testtime: 60s

fs: ext4

class: os

test: memhotplug

cpufreq_governor: performance

test machine

lkp-csl-2sp5

status

merged at v5.15-rc7

3.4.            Will-it-scale.per_process_ops

Will-it-scale takes a test case and runs it from 1 through to n parallel copies to see if the test case will scale. It builds both process and threads based tests in order to see any differences between the two.

3.2.3   Scenario: process page_fault3

Commit 1f828223b7 was reported to have 37.3% improvement of will-it-scale.per_process_ops when comparing to v5.15-rc2. It was merged to mainline at v5.15-rc3.

 

Correlated commits

1f828223b7

memcg: flush lruvec stats in the refault

branch

linux/master

report

[memcg] 1f828223b7: 37.3% improvement

test scenario

nr_task: 100%

mode: process

test: page_fault3

thp_enabled: never

cpufreq_governor: performance

test machine

lkp-csl-2sp5

status

merged at v5.15-rc3

 

4.            Shift-Left Testing

Beyond testing trees in the upstream kernel, 0-Day CI also tests developers’ and maintainers’ trees, which can catch issues earlier and reduce wider impact. We call it “shift-left” testing. During the v5.15 release cycle, 0-Day CI had reported 5 major performance regressions and 5 major improvements by doing shift-left testing. We will share more detailed information together with possible code changes that led to this result for some of these, though the assessment is limited by the test coverage we have now. The whole list is summarized in the report summary section.

4.1         Report Summary

0-Day CI had reported 5 performance regressions and 5 improvements by doing shift-left testing on developer and maintainer repos.

 

Test Indicator

Mail

Test Scenario

Test Machine

Status

aim7.jobs-per-min

[memcg] 45208c9105: -14.0% regression

disk: 1BRD_48G

fs: xfs

test: disk_rr

load: 3000

cpufreq_governor: performance

lkp-icl-2sp2

currently not merged, author accepted the regression, fixed patch 1f828223b7 was merged at v5.15-rc3

stress-ng.loop.ops_per_sec

[watchdog] 34db6459c6: -36.4% regression

nr_threads: 100%

disk: 1HDD

testtime: 60s

class: device

test: loop

cpufreq_governor: performance

lkp-csl-2sp7

currently not merged, no response from author

stress-ng.memhotplug.ops_per_sec

[mm/migrate] 9eeb73028c: -53.8% regression

nr_threads: 10%

disk: 1HDD

testtime: 60s

fs: ext4

class: os

test: memhotplug

cpufreq_governor: performance

lkp-csl-2sp5

currently not merged, author accepted the regression, fixed patch 295be91f7e was merged at v5.15-rc7

will-it-scale.per_process_ops

[memcg] 059dd9003a: -39.8% regression

nr_task: 100%

mode: process

test: lock1

cpufreq_governor: performance

lkp-icl-2sp1

currently not merged, no response from author

will-it-scale.per_thread_ops

[memcg] fa4e6b1ad5: -15.4% regression

nr_task: 50%

mode: thread

test: poll2

cpufreq_governor: performance

lkp-hsw-4ex1

currently not merged, no response from author

Improvement

 

 

 

 

fio.write_iops

[block] dc5fc361d8: 9.8% improvement

disk: 1SSD

fs: xfs

runtime: 300s

nr_task: 32

rw: randwrite

bs: 4k

ioengine: sync

test_size: 256g

cpufreq_governor: performance

lkp-csl-2ap1

currently not merged

reaim.jobs_per_min

[mm] f11826446d: 3.9% improvement

runtime: 300s

nr_task: 100%

test: new_dbase

cpufreq_governor: performance

lkp-cpl-4sp1

currently not merged

stress-ng.vm-rw.ops_per_sec

29b51054ec: 81.4% improvement

nr_threads: 100%

testtime: 60s

class: memory

test: vm-rw

cpufreq_governor: performance

lkp-csl-2sp5

currently not merged

vm-scalability.median

[memory tiering] 76ff9ff49a: 5.0% improvement

runtime: 300s

size: 512G

test: anon-w-rand

cpufreq_governor: performance

lkp-csl-2ap4

currently not merged

will-it-scale.per_process_ops

[rcu] 925da92ba5: 2.5% improvement

nr_task: 16

mode: process

test: getppid1

cpufreq_governor: performance

lkp-csl-2sp9

currently not merged

4.2         Stress-ng.loop.ops_per_sec

stress-ng will stress test a computer system in various selectable ways. It was designed to exercise various physical subsystems of a computer as well as the various operating system kernel interfaces.

 

4.2.1   Scenario: loop test on localhost

Commit 34db6459c6 was reported to have -36.4% regression of stress-ng.loop.ops_per_sec when comparing to v5.15-rc4.

 

Correlated commits

34db6459c6

watchdog: iTCO: Drop vendor support

branch

groeck-staging/watchdog-next

report

[watchdog] 34db6459c6: -36.4% regression

test scenario

nr_threads: 100%

disk: 1HDD

testtime: 60s

class: device

test: loop

cpufreq_governor: performance

test machine

lkp-csl-2sp7

status

currently not merged, no response from author

4.3         Fio.write_iops

 

fio is a tool that will spawn a number of threads or processes doing a particular type of I/O action as specified by the user.

4.3.1   Scenario: randwrite on xfs

 

Commit dc5fc361d8 was reported to have 9.8% improvement of fio.write_iops when comparing to v5.15-rc6.

 

Correlated commits

dc5fc361d8

block: attempt direct issue of plug list

branch

axboe-block/for-5.16/block

report

[block] dc5fc361d8: 9.8% improvement

test scenario

disk: 1SSD

fs: xfs

runtime: 300s

nr_task: 32

rw: randwrite

bs: 4k

ioengine: sync

test_size: 256g

cpufreq_governor: performance

test machine

lkp-csl-2ap1

status

currently not merged

4.4         Stress-ng.vm-rw.ops_per_sec

Stress-ng will stress test a computer system in various selectable ways. It was designed to exercise various physical subsystems of a computer as well as the various operating system kernel interfaces.

4.4.1   Scenario: 1HDD-xfs-filesystem-dir

Commit 29b51054ec was reported to have 81.4% improvement of stress-ng.vm-rw.ops_per_sec when comparing to v5.15-rc3.

 

Correlated commits

29b51054ec

SIS_DEPTH

branch

mel/sched-sisdepth-v1r3

report

29b51054ec: 81.4% improvement

test scenario

nr_threads: 100%

testtime: 60s

class: memory

test: vm-rw

cpufreq_governor: performance

test machine

lkp-csl-2sp5

status

currently not merged

5        Latest Release Performance Comparing

 

This session gives some information about the performance difference among different kernel releases, especially between v5.15 and v5.14. There are 50+ performance benchmarks running in 0-Day CI, and we selected 9 benchmarks which historically showed the most regressions/improvements reported by 0-Day CI. Some typical configuration/parameters are used to run the test. For some of the regressions from the comparison, 0-Day did not successfully bisect it thus no related report sent out during the release development period, but it is still worth checking. The root cause to cause the regressions won’t be covered in this session.

 

In the following figures, the value on the Y-axis is the relative performance number. We used the v5.14 data as the base (performance number is 100).

5.1         test suite: vm-scalability

Vm-scalability exercises functions and regions of the mm subsystem of the Linux kernel. Below 2 tests show the typical test results.

vm-scalability Test 1

 

 

 

 

Here are the test configuration and performance test summary for above tests:                                                                                                                                                                                                          

 

vm-scalability Test 1

test machine

model: Cascade Lake

brand: Intel(R) Xeon(R) CPU @ 2.30GHz

cpu_number: 96

memory: 128G

runtime

No requirement

size

No requirement

vm-scalability test parameter

test: swap-w-seq

performance summary

vm-scalability.throughput on kernel v5.15 has -7.97% regression when comparing to v5.14

 

5.2         test suite: will-it-scale

Will-it-scale takes a test case and runs it from 1 through to n parallel copies to see if the test case will scale. It builds both process and threads based tests in order to see any differences between the two.

 

 

will-it-scale  Test 1

will-it-scale  Test 1

 

 

Here are the parameters and performance test summary for above tests:                                                                                                                                                                                                                         

 

will-it-scale Test 1

will-it-scale Test 1

test machine

model: Skylake

brand:

cpu_number: 104

memory: 192G

model: Ice Lake

brand:

cpu_number: 96

memory: 256G

nr_task

16

100%

will-it-scale test parameter

mode: thread

test: fallocate2

mode: process

test: page_fault1

summary

will-it-scale.per_process_ops on kernel v5.15 has -9.38% regression when comparing to v5.14

will-it-scale.per_process_ops on kernel v5.15 has 9.21% improvement when comparing to v5.14

5.3         test suite: unixbench

UnixBench is a system benchmark to provide a basic indicator of the performance of a Unix-like system.

Unixbench Test 1

 

 

Here are the test configuration and performance test summary for above tests:                                                                                                                                                                                                          

 

Unixbench Test 1

test machine

model: Ice Lake

brand:

nr_cpu: 96

memory: 256G

runtime

300s

nr_task

30%

unixbench test parameter

test: syscall

performance summary

unixbench.score on kernel v5.15 is almost the same as that in v5.14

5.4         test suite: reaim

Reaim updates and improves the existing Open Source AIM 7 benchmark. aim7 is a traditional UNIX system level benchmark suite which is used to test and measure the performance of a multiuser system.

                                  reaim Test 1

 

Here are the test configuration and performance test summary for above tests:                                                                                                                                                                                                                     

 

reaim Test 1

test machine

model: Cascade Lake

brand: Intel(R) Xeon(R) Platinum 9242 CPU @ 2.30GHz

cpu_number: 192

memory: 192G

runtime

300s

nr_task

100%

disk

No requirement

fs

No requirement

reaim test parameter

test: dbase

performance  summary

reaim.jobs_per_min on kernel v5.15 is almost the same as that in v5.14

 

5.5         test suite: pigz

Pigz, which stands for Parallel Implementation of GZip, is a fully functional replacement for gzip that exploits multiple processors and multiple cores to the hilt when compressing data.

pigz Test 1

 

 

Here are the test configuration and performance test summary for above tests:         

 

pigz Test 1

test machine

model: Knights Mill

brand: Intel(R) Xeon Phi(TM) CPU 7255 @ 1.10GHz

cpu_number: 272

memory: 112G

nr_threads

25%

pigz Test parameter

blocksize: 512K

performance  summary

pigz.throughput on kernel v5.15 is almost the same as that in v5.14

5.6         test suite: netperf

Netperf is a benchmark that can be used to measure the performance of many different types of networking. It provides tests for both unidirectional throughput, and end-to-end latency.

netperf Test 1

 

 

Here are the test configuration and performance test summary for above tests:                                                                                                                                                                                                                         

 

netperf Test 1

test machine

model: Cascade Lake

brand: Intel(R) Xeon(R) Platinum 9242 CPU @ 2.30GHz

cpu_number: 192

memory: 192G

disable_latency_stats

1

set_nic_irq_affinity

1

runtime

300s

nr_threads

50%

ip

ipv4

netperf test parameter

test: TCP_RR

performance  summary

netperf.Throughput_tps on kernel v5.15 has -11.07% regression when comparing to v5.14

5.7         test suite: hackbench

Hackbench is both a benchmark and a stress test for the Linux kernel scheduler. It's  main job  is  to  create a specified number of pairs of schedulable entities (either threads or traditional processes) which communicate via either sockets or pipes and time how long  it takes for each pair to send data back and forth.

hackbench Test 1

 

 

Here are the test configuration and performance test summary for above tests:                                                                                                                                                                                                          

 

hackbench Test 1

test machine

model: Cascade Lake

brand: Intel(R) Xeon(R) CPU @ 2.30GHz

cpu_number: 96

memory: 128G

disable_latency_stats

1

nr_threads

50%

unixbench test parameter

mode: process

ipc: pipe

performance summary

hackbench.throughput on kernel v5.15 has -7.33% regression when comparing to v5.14

5.8         test suite: fio

Fio was originally written to save me the hassle of writing special test case programs when I wanted to test a specific workload, either for performance reasons or to find/reproduce a bug.

fio Test 1

fio Test 2

Here are the test configuration and performance test summary for above tests:                                                                                                                                                                                                                         

 

fio Test 1

fio Test 2

test machine

model: Cascade Lake

brand: Intel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz

cpu_number: 96

memory: 256G

model: Cascade Lake

brand: Intel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz

cpu_number: 96

memory: 256G

runtime

200s

200s

file system

xfs

xfs

disk

2pmem

2pmem

boot_params

No requirement

No requirement

nr_task

50%

50%

time_based

tb

tb

fio test parameter

fio-setup-basic:

  rw: randwrite

  bs: 4k

  ioengine: sync

  test_size: 200G

fio-setup-basic:

  rw: write

  bs: 4k

  ioengine: libaio

  test_size: 200G

performance  summary

fio.write_bw_MBps on kernel v5.15 has 35% improvement when comparing to v5.14

fio.write_bw_MBps on kernel v5.15 has -26% regression when comparing to v5.14

5.9         test suite: ebizzy

Ebizzy is designed to generate a workload resembling common web application server workloads. It is highly threaded, has a large in-memory working set, and allocates and deallocates memory frequently.

 

                                       ebizzy Test 1

 

 

Here are the test configuration and performance test summary for above test:                                                                                                                                                                                                                         

 

ebizzy Test 1

test machine

model: Cascade Lake

brand: Intel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz

cpu_number: 96

memory: 192G

transparent_hugepage

No requirement

nr_threads

200%

iterations

100x

ebizzy test parameter

duration: 10s

performance  summary

ebizzy.throughput on kernel v5.15 is almost the same as that in v5.14

6        Test Machines

6.1         IVB Desktop

model

Ivy Bridge

brand

Intel(R) Core(TM) i3-3220 CPU @ 3.30GHz

cpu number

8

memory

16G

 

model

Ivy Bridge

brand

Intel(R) Core(TM) i3-3220 CPU @ 3.30GHz

cpu number

4

memory

8G

 

6.2         SKL SP

model

Skylake

brand

Intel(R) Xeon(R) CPU E5-2697 v2 @ 2.70GHz

cpu number

80

memory

64G

 

6.3         BDW EP

model

Broadwell-EP

brand

Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz

cpu number

88

memory

128G

 

6.4         HSW EP

model

Haswell-EP

brand

Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz

cpu number

72

memory

128G

 

6.5         IVB EP

model

Ivy Bridge-EP

brand

Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz

cpu number

40

memory

384G

 

model

Ivytown Ivy Bridge-EP

brand

Intel(R) Xeon(R) CPU E5-2697 v2 @ 2.70GHz

cpu number

48

memory

64G

 

6.6         HSX EX

model

Brickland Haswell-EX

brand

Intel(R) Xeon(R) CPU E7-8890 v3 @ 2.50GHz

cpu number

144

memory

512G