[gstreamer-bugs] [Bug 605594] relatively high cpu usage compared to other players

Sun Jan 31 13:19:54 PST 2010

https://bugzilla.gnome.org/show_bug.cgi?id=605594
  GStreamer | gstreamer (core) | 0.10.x

--- Comment #4 from Olaf Leidinger <leidola at web.de> 2010-01-31 21:19:48 UTC ---
> Please confirm that you use the same audiosinks?

They do, both use the alsa audio sink.

> Finally how is you gstreamer build. Is it a debug build or are you (or your distro) use e.g.
> CFLAGS="-DG_DISABLE_ASSERT -DG_DISABLE_CAST_CHECKS"

Using gentoo the package management compiles everything possible from source.
Therefore I used the following CFLAGS:

-march=k8 -msse -msse2 -msse3 -mssse3 -mfpmath=sse -ftree-vectorize -O2 -pipe 

This is a good set of flags for my ATOM netbook. march=k8 is intentional, as it
seems to be faster than optimizing for pentium-m.
Therefore the tests were sort of a debug build. I rebuilt gstreamer using 

CFLAGS=-march=k8 -msse -msse2 -msse3 -mssse3 -mfpmath=sse -ftree-vectorize -O2
-pipe -DG_DISABLE_ASSERT -DG_DISABLE_CAST_CHECKS 

And added --disable-trace to the configure flags. In total they where:

--prefix=/usr --build=i686-pc-linux-gnu --host=i686-pc-linux-gnu
--mandir=/usr/share/man --infodir=/usr/share/info --datadir=/usr/share
--sysconfdir=/etc --localstatedir=/var/lib --disable-static
--disable-dependency-tracking --enable-nls --disable-valgrind
--disable-examples --disable-trace --disable-tests
--with-package-name="GStreamer ebuild for Gentoo"
--with-package-origin=http://packages.gentoo.org/package/media-libs/gstreamer

###############

Here some benchmarks using this (and the old build):

Setting fakesink using playbin2 seems not to work. Therefore
I used the following:

#### gst (with asserts, async)
time gst-launch-0.10 filesrc location=tailspin.ogg ! oggdemux ! vorbisdec!
fakesink sync=false
[...]
real    0m9.168s
user    0m9.117s
sys    0m0.028s

#### gst (without asserts, async)
time gst-launch-0.10 filesrc location=tailspin.ogg ! oggdemux ! vorbisdec!
fakesink sync=false
[...]
real    0m8.963s
user    0m8.645s
sys    0m0.112s

###### As I couldn't find a switch in mplayer to disable sync output, again the
same with sync output

#### gst (with asserts, sync)
time gst-launch-0.10 filesrc location=tailspin.ogg ! oggdemux ! vorbisdec!
fakesink sync=true
[...]
real    5m17.516s
user    0m13.305s
sys    0m0.912s

#### gst (without asserts, sync)
time gst-launch-0.10 filesrc location=tailspin.ogg ! oggdemux ! vorbisdec!
fakesink sync=true
[...]
real    5m17.508s
user    0m12.185s
sys    0m0.768s

#### mplayer (null)
# And again with mplayer
time mplayer -ao null tailspin.ogg
[...]
real    5m27.494s
user    0m5.832s
sys    0m1.204s

########## And now with audio output, using playbin*

### gst (with asserts, playbin)
time gst-launch-0.10 playbin uri=file:///"`pwd`/tailspin.ogg" 
[...]
real    5m18.044s
user    0m20.837s
sys    0m1.976s

### gst (with asserts, playbin2)
time gst-launch-0.10 playbin2 uri=file:///"`pwd`/tailspin.ogg"
[...]
real    5m17.955s
user    0m20.441s
sys    0m2.112s

### gst (without asserts, playbin)
time gst-launch-0.10 playbin uri=file:///"`pwd`/tailspin.ogg" 
[...]
real    5m17.885s
user    0m19.021s
sys    0m1.752s

### gst (without asserts, playbin2)
time gst-launch-0.10 playbin2 uri=file:///"`pwd`/tailspin.ogg"
[...]
real    5m17.821s
user    0m21.669s  # error of measurement?
sys    0m2.044s

### mplayer (alsa)
real    5m18.415s
user    0m6.788s
sys    0m1.812s

### To conclude the tests I rebuilt gstreamer, the vorbis-plugin and libvorbis
with -O3 
### same options as above, but -O2 replaced (mplayer seems not to be linked
against libvorbis)

### gst (without asserts, playbin, -O3)
time gst-launch-0.10 playbin uri=file:///"`pwd`/tailspin.ogg" 
[...]
real    5m18.055s
user    0m19.541s
sys    0m2.024s

### gst (without asserts, playbin2, -O3)
time gst-launch-0.10 playbin2 uri=file:///"`pwd`/tailspin.ogg"
[...]
real    5m17.800s
user    0m18.617s
sys    0m1.996s

### To find out if playbin* are slow, I constructed the following pipeline
time gst-launch-0.10 filesrc location=tailspin.ogg ! oggdemux ! vorbisdec!
audioconvert ! alsasink
[...]
real    5m17.648s
user    0m14.501s
sys    0m1.012s

# Table of results:

 [seconds]      | w ass. |  wo ass.
--------------------------------------
fakesink, async |  9.117 |  8.645
fakesink, sync  | 13.305 | 12.185
playbin         | 20.837 | 19.021
playbin2        | 20.441 | 21.669
playbin  -O3    |  ---   | 19.541
playbin2 -O3    |  ---   | 18.617
alsasink -O3    |  ---   | 14.501
mplayer null    |      5.832
mplayer alsa    |      6.788

[%] cpu time/
runtime *100    | w ass. |  wo ass.
--------------------------------------
fakesink, async | 99.4   | 96.5  
fakesink, sync  |  4.2   |  3.8  
playbin         |  6.6   |  6.0
playbin2        |  6.4   |  6.8
playbin   -O3   |  ---   |  6.1
playbin2  -O3   |  ---   |  5.9    
alsasink  -O3   |  ---   |  4.6
mplayer null    |       1.8  
mplayer alsa    |       2.1

Conclusion:

 - Using -O3 doesn't bring much more performance
 - Playbin has overhead!
 - The asserts have also a noticeable overhead

-- 
Configure bugmail: https://bugzilla.gnome.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the QA contact for the bug.
You are the assignee for the bug.