Arch Linux

gedgon

Member

From: Poland

Registered: 2011-01-27

Posts: 95

[Solved] System with Linux 3.14 (intel_pstate) noticeable less...

... responsive and significantly slower (some numbers below) under light/medium workload, even when compared with ondemand (up_threshold 95). What happened to that famous race to idle concept? Relevant bug report https://bugzilla.kernel.org/show_bug.cgi?id=70941 (apparently "fixed")

"Tests" on 2500k (4 cores SB @ stock freq.):

1) in tmpfs

for num in {1..50000}; do
    dd if=/dev/urandom of=file$num.log bs=10k count=1
done

2) Sunspider 1.0.2 https://www.webkit.org/perf/sunspider-1 … river.html

3) rsync in tmpfs

Results:

Linux 3.14 intel_pstate /w powersave governor (default)

1) 
real	2m7.609s
user	0m5.060s
sys	0m12.797s

2)
run 1 - 200.9ms +/- 1.9%
run 2 - 199.1ms +/- 1.9%
run 3 - 202.3ms +/- 1.0%

3)
first run: $ time rsync -aAX /tmp/modules /tmp/test
real	0m1.692s
user	0m2.033s
sys	0m0.887s

second run: $ time rsync -aAX /tmp/modules /tmp/test
real	0m0.578s
user	0m0.267s
sys	0m0.270s

Linux 3.14 ondemand up threshold 95 (intel_pstate=disable)

1)
real	1m35.773s
user	0m4.947s
sys	0m12.890s

2)
run 1 - 154.3ms +/- 0.8%
run 2 - 155.2ms +/- 0.8%
run 3 - 154.0ms +/- 0.7%

3)
first run: $ time rsync -aAX /tmp/modules /tmp/test
real	0m1.097s
user	0m1.210s
sys	0m0.620s

second run: $ time rsync -aAX /tmp/modules /tmp/test
real	0m0.473s
user	0m0.200s
sys	0m0.230s

Linux 3.13.9 intel_pstate /w powersave governor (default)

1)
real	1m3.553s
user	0m15.353s
sys	0m42.444s

2)
run 1 - 155.4ms +/- 0.9%
run 2 - 154.6ms +/- 0.5%
run 3 - 153.9ms +/- 1.1%


3)
first run: $ time rsync -aAX /tmp/modules /tmp/test
real	0m1.014s
user	0m1.351s
sys	0m0.353s

second run: $ time rsync -aAX /tmp/modules /tmp/test
real	0m0.344s
user	0m0.154s
sys	0m0.149s

Last edited by gedgon (2014-06-24 09:29:23)

kostas213

Member

From: Zürich

Registered: 2012-04-11

Posts: 53

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

I have the same issue with a Sandy Bridge mobile processor. I have not run any tests, but I see that the frequency is almost everytime at 800 MHz with kernel 3.14. When I had been using kernel 3.13, the frequency was more or less fluctuating from ~1.3 to ~2.3 GHz (max frequency is 2.9 GHz / i7-2630QM).

ziro360

Member

Registered: 2012-04-27

Posts: 30

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

I have this issue as well, downgraded to kernel 3.13 for now

MarcvdM

Member

Registered: 2014-05-08

Posts: 1

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

same here (Fedora 20). i7-4800MQ on 3.13 in-rest at 2.7 GHz, on 3.14 800MHz. Boot times 30% up. A wrapper around commands shows elapsed times for "systemd-analyze blame". Both kernel boots in "powersave" governor:
4 times command execution on kernel 3.13:
.121303
.124094
.122310
.125066
and the same in kernel 3.14:
.442166
.442842
.447207
.430436
upping the --min through cpupower to 2.7 GHz I stell don't get the perf of 3.13:
.139518
.140489
.139541
.136105

graysky

Wiki Maintainer

From: :wq

Registered: 2008-12-01

Posts: 10,606

Website

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

... And the same tests using the linux-ck package in the AUR?

---

CPU-optimized Linux-ck packages @ Repo-ck  • AUR packages • Zsh and other configs

gedgon

Member

From: Poland

Registered: 2011-01-27

Posts: 95

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

... And the same tests using the linux-ck package in the AUR?

IIRC that was 3.13.9-ck from your repo, but the difference between -arch and -ck was minimal in this case, close to the margin of error.

ooo

Member

Registered: 2013-04-10

Posts: 1,638

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

http://lkml.iu.edu/hypermail/linux/kern … 03087.html
looks like intel_pstate mainainer is on to it.

helge

Member

Registered: 2012-10-22

Posts: 4

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

I've had this problem on 3.14 too.. It seems to fixed in linux-mainline 3.15 (using RC8 here).
Using the powersave pstate-governor the cpu perfectly scales up to its maximum and lowers the freq when idling.

gedgon

Member

From: Poland

Registered: 2011-01-27

Posts: 95

Re: [Solved] System with Linux 3.14 (intel_pstate) noticeable less...

It seems to fixed in linux-mainline 3.15 (using RC8 here).

Thanks for the heads-up, Helge. Marking as solved, despite it's still 10-15% slower than 3.13. At least we known, that it was a bug, not a feature