GStreamer-devel

strange CPU loading for audiotestsrc and filesrc

Classic

List

Threaded

5 messages Options

Viraj Karandikar

strange CPU loading for audiotestsrc and filesrc

Hi,

I am seeing some strange behaviour while doing CPU load measurement when running gstreamer.
I am relying on uptime and idle time give by "cat /proc/uptime" to calculate CPU load.
We have implemented custom audio source (dalsrc) and sink (dalsink) plugins for our board (ARM9E on TI DM6446).
I am using gstreamer-0.10.24.

I have following two cases:

Case 1:

gst-launch audiotestsrc wave=silence is_live=true blocksize=160 ! audio/x-raw-int,rate=8000,depth=16,channels=2,width=16,signed=\(boolean\)TRUE,endianness=\(int\)1234 ! dalsink
Average CPU load over 5 minutes : 18%

Case 2:

gst-launch filesrc location="wa12fs1_8k_stereo.wav" blocksize=160 ! wavparse ! dalsink
Average CPU load over 5 minutes : 2%

I am not able to explain difference in CPU loading in above 2 cases.
Does anyone have any clue?

regards,
Viraj

------------------------------------------------------------------------------
Come build with us! The BlackBerry(R) Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay
ahead of the curve. Join us from November 9 - 12, 2009. Register now!
http://p.sf.net/sfu/devconference
_______________________________________________
gstreamer-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/gstreamer-devel

michael smith-6-3

Re: strange CPU loading for audiotestsrc and filesrc

On Wed, Oct 7, 2009 at 10:49 PM, Viraj Karandikar <[hidden email]> wrote:

> Hi,
>
> I am seeing some strange behaviour while doing CPU load measurement when
> running gstreamer.
> I am relying on uptime and idle time give by "cat /proc/uptime" to calculate
> CPU load.
> We have implemented custom audio source (dalsrc) and sink (dalsink) plugins
> for our board (ARM9E on TI DM6446).
> I am using gstreamer-0.10.24.
>
> I have following two cases:
>
> Case 1:
> gst-launch audiotestsrc wave=silence is_live=true blocksize=160 !
> audio/x-raw-int,rate=8000,depth=16,channels=2,width=16,signed=\(boolean\)TRUE,endianness=\(int\)1234
> ! dalsink
> Average CPU load over 5 minutes : 18%

As well as your cpu usage measurement technique being pretty
inaccurate, it's not surprising that audiotestsrc is slow - it's a
test element, and has has absolutely no optimisation done, or even any
attempt to make it vaguely ok. It's probably just doing something
stupid for the silence case.

Additionally, the first case is a live source - your second isn't.
Those aren't very directly comparable. There might be something that's
slower because of that.

Mike

------------------------------------------------------------------------------
Come build with us! The BlackBerry(R) Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay
ahead of the curve. Join us from November 9 - 12, 2009. Register now!
http://p.sf.net/sfu/devconference
_______________________________________________
gstreamer-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/gstreamer-devel

Viraj Karandikar

Re: strange CPU loading for audiotestsrc and filesrc

Hi,

I had tried audiotestsrc without live option too. But loading wasnt any different from live case.

I looked into audiotestsrc source. Its doing simple memset() to zero on the buffer. Didnt see any complex code there at least for silence case.

About profiling, I tried to get oprofile working on the board. But its not working. Might have some dependency issue with the kernel.

regards,

Viraj

On Thu, Oct 8, 2009 at 11:27 AM, Michael Smith <[hidden email]> wrote:

On Wed, Oct 7, 2009 at 10:49 PM, Viraj Karandikar <[hidden email]> wrote:
> Hi,
>
> I am seeing some strange behaviour while doing CPU load measurement when
> running gstreamer.
> I am relying on uptime and idle time give by "cat /proc/uptime" to calculate
> CPU load.
> We have implemented custom audio source (dalsrc) and sink (dalsink) plugins
> for our board (ARM9E on TI DM6446).
> I am using gstreamer-0.10.24.
>
> I have following two cases:
>
> Case 1:
> gst-launch audiotestsrc wave=silence is_live=true blocksize=160 !
> audio/x-raw-int,rate=8000,depth=16,channels=2,width=16,signed=\(boolean\)TRUE,endianness=\(int\)1234
> ! dalsink
> Average CPU load over 5 minutes : 18%

As well as your cpu usage measurement technique being pretty
inaccurate, it's not surprising that audiotestsrc is slow - it's a
test element, and has has absolutely no optimisation done, or even any
attempt to make it vaguely ok. It's probably just doing something
stupid for the silence case.

Additionally, the first case is a live source - your second isn't.
Those aren't very directly comparable. There might be something that's
slower because of that.

Mike

------------------------------------------------------------------------------
Come build with us! The BlackBerry(R) Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay
ahead of the curve. Join us from November 9 - 12, 2009. Register now!
http://p.sf.net/sfu/devconference
_______________________________________________
gstreamer-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/gstreamer-devel

--
- Viraj
Reality is merely an illusion, albeit a very persistent one.

------------------------------------------------------------------------------
Come build with us! The BlackBerry(R) Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay
ahead of the curve. Join us from November 9 - 12, 2009. Register now!
http://p.sf.net/sfu/devconference
_______________________________________________
gstreamer-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/gstreamer-devel

Stefan Sauer

Re: strange CPU loading for audiotestsrc and filesrc

In reply to this post by michael smith-6-3

Michael Smith schrieb:

> On Wed, Oct 7, 2009 at 10:49 PM, Viraj Karandikar <[hidden email]> wrote:
>> Hi,
>>
>> I am seeing some strange behaviour while doing CPU load measurement when
>> running gstreamer.
>> I am relying on uptime and idle time give by "cat /proc/uptime" to calculate
>> CPU load.
>> We have implemented custom audio source (dalsrc) and sink (dalsink) plugins
>> for our board (ARM9E on TI DM6446).
>> I am using gstreamer-0.10.24.
>>
>> I have following two cases:
>>
>> Case 1:
>> gst-launch audiotestsrc wave=silence is_live=true blocksize=160 !
>> audio/x-raw-int,rate=8000,depth=16,channels=2,width=16,signed=\(boolean\)TRUE,endianness=\(int\)1234
>> ! dalsink
>> Average CPU load over 5 minutes : 18%
>
> As well as your cpu usage measurement technique being pretty
> inaccurate, it's not surprising that audiotestsrc is slow - it's a
> test element, and has has absolutely no optimisation done, or even any
> attempt to make it vaguely ok. It's probably just doing something
> stupid for the silence case.

Well, it actually has bben optimized somewhat. Especially for silence it
*shouldn't* do much. But anyway the only way to find out would be to run e.g.
oprofile on it (its available on arm too).

A slightly better version to check the load would imho be to use sync=false on
the sink and run the example with time. Then see which version is finished earlier.

Stefan

>
> Additionally, the first case is a live source - your second isn't.
> Those aren't very directly comparable. There might be something that's
> slower because of that.
>
> Mike
>
> ------------------------------------------------------------------------------
> Come build with us! The BlackBerry(R) Developer Conference in SF, CA
> is the only developer event you need to attend this year. Jumpstart your
> developing skills, take BlackBerry mobile applications to market and stay
> ahead of the curve. Join us from November 9 - 12, 2009. Register now!
> http://p.sf.net/sfu/devconference
> _______________________________________________
> gstreamer-devel mailing list
> [hidden email]
> https://lists.sourceforge.net/lists/listinfo/gstreamer-devel

Viraj Karandikar

Re: strange CPU loading for audiotestsrc and filesrc

Hi,

I was able to build oprofile and get it working on DM6446.

But I am not getting consistent numbers with oprofile. For case 1, I am getting CPU load around 12-14%, i.e 2-3% variation across runs.

Is this expected?

Also in case 1, most of the CPU is spent in gstreamer libraries (libgstreamer-0.10.so.0.21.0, libgstbase-0.10.so.0.21.0, libpthread-0.10.so, libgobject-2.0.so.0.2105.0, libglib-2.0.so.0.2105.0) and audiotestsrc is taking around 0.2% of total (for silence case).

I also tried --disable-debug, --disable-gst-debug, --disable-trace --disable-alloc-trace options while configuring the gstreamer.

With this build I got reduction of around 50% in gstreamer loading (out of 8% CPU load in one of the test case). This should help in reducing overall system load in final application.

Thanks,

Viraj

On Tue, Oct 13, 2009 at 1:56 AM, Stefan Kost <[hidden email]> wrote:

Michael Smith schrieb:

> On Wed, Oct 7, 2009 at 10:49 PM, Viraj Karandikar <[hidden email]> wrote:
>> Hi,
>>
>> I am seeing some strange behaviour while doing CPU load measurement when
>> running gstreamer.
>> I am relying on uptime and idle time give by "cat /proc/uptime" to calculate
>> CPU load.
>> We have implemented custom audio source (dalsrc) and sink (dalsink) plugins
>> for our board (ARM9E on TI DM6446).
>> I am using gstreamer-0.10.24.
>>
>> I have following two cases:
>>
>> Case 1:
>> gst-launch audiotestsrc wave=silence is_live=true blocksize=160 !
>> audio/x-raw-int,rate=8000,depth=16,channels=2,width=16,signed=\(boolean\)TRUE,endianness=\(int\)1234
>> ! dalsink
>> Average CPU load over 5 minutes : 18%
>
> As well as your cpu usage measurement technique being pretty
> inaccurate, it's not surprising that audiotestsrc is slow - it's a
> test element, and has has absolutely no optimisation done, or even any
> attempt to make it vaguely ok. It's probably just doing something
> stupid for the silence case.

Well, it actually has bben optimized somewhat. Especially for silence it
*shouldn't* do much. But anyway the only way to find out would be to run e.g.
oprofile on it (its available on arm too).

A slightly better version to check the load would imho be to use sync=false on
the sink and run the example with time. Then see which version is finished earlier.

Stefan

>
> Additionally, the first case is a live source - your second isn't.
> Those aren't very directly comparable. There might be something that's
> slower because of that.
>
> Mike
>
> ------------------------------------------------------------------------------
> Come build with us! The BlackBerry(R) Developer Conference in SF, CA
> is the only developer event you need to attend this year. Jumpstart your
> developing skills, take BlackBerry mobile applications to market and stay
> ahead of the curve. Join us from November 9 - 12, 2009. Register now!
> http://p.sf.net/sfu/devconference
> _______________________________________________
> gstreamer-devel mailing list
> [hidden email]
> https://lists.sourceforge.net/lists/listinfo/gstreamer-devel

------------------------------------------------------------------------------
Come build with us! The BlackBerry(R) Developer Conference in SF, CA
is the only developer event you need to attend this year. Jumpstart your
developing skills, take BlackBerry mobile applications to market and stay
ahead of the curve. Join us from November 9 - 12, 2009. Register now!
http://p.sf.net/sfu/devconference
_______________________________________________
gstreamer-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/gstreamer-devel