Re: Re: RV10 encoding speed

Karl, 

thanks very much for the response and result to share. I did
some test based on Steve's suggestion and come up with quite
similar result to yours, which we are glad that is something
at least we can have some common base, i.e. for talking head
sequence like CIF Aikyo, RV10 with complexity low setting
and keyframe setting at 15 can do 520+ frames per second to
encode bitstream of 225kbps  25fps (audience setting is
256K) on a P4 Dual core 3.0GHz machine. But I have some
troubles to understand why the above result doesn't scale
up:

We use the same machine with the above hw configuration to
encode 8 channels of the same clip with same setting (by
launching 8 different encoder processes), the CPU is very
much saturated and the encoded clips have some frame loss
(can only encode around 23fps). Intuitively, 8 channels = 8
* 25fps = 200fps which is only half of 500fps the machine
can achieve. Where is the CPU spin on? 

If what Karl's mentioning of file reading overhead really
matters, our test seems not echoing that statement because
we tried to read all YUV420 data into memory and keep using
it for every channel, it doesn't seem to matter that much
(there are slight improvement) to cover the 50% difference
we observed. 

we seems to run out of idea how to make sense of this
problem and would like to get some expert suggestion from
you or anyone in the Helix producer community who has some
experience on this issue. thanks,

Frank,

P.S., How can we configure the Helix producer to use RV9
instead of RV10 to see if there is any difference?

Karl Lillevold <karllreal.com> wrote: The
encoder also has several complexity levels that can vary
the
encoding speed up to 3-5X for real time adaptation on slower
systems.

For the system mentioned, the file reading overhead is going
to be large
part of the time it takes to encode an uncompressed test
sequence.

On my dual core 2 at 2.4 GHz, I get from 5 to 19 ms per
frame for
Foreman CIF, depending on chosen complexity, and from 2.5 to
14 ms per
frame for Akiyo, not counting file reading overhead.

Karl.


Steve McMillen wrote:
> I recommend you do a file to file encode of an input
clip of a fixed
> duration and your resulting duration will tell you how
many fps were
> processed.  I fully expect the result to be < actual
clip duration. 
> That is, given that frame size and bit rate, encoding
should easily be
> faster than realtime on the box you specified.
> 
> On 8/5/2007 6:56 PM, on focus wrote:
>> Hi, Steve,
>>
>> We have come to a point that we need to know more
exactly what is the
>> encoding performance of RV10. We checked the only
available Rv10 codec
>> review .pdf file from Helix community/Real website,
there is a brief
>> mentioning of its speed, but not quite exact (there
is actually a
>> conflicting and confusing info there with footnote)
to let us decide
>> what machine we can properly choose for our
encoding task. Basically,
>> are there any more thorough or exact doc that can
be shared in Helix
>> community that gives:
>>
>> What is the encoding speed or performance for
encoding a target of CIF
>> 352x288 25fps  256kbps on a  P4  Dual
3.0GHz machine or whatever
>> physical machine it could be targeted for a
standard test sequence,
>> like, Aikyo, or Foreman? The encoding speed can be
measured by how
>> many actual frames RV10 can encode per second?  or
How many mips it
>> takes for encoding a frame averagely?
>>
>> Many thanks,
>>
>> Frank,
>>
>>
------------------------------------------------------------
------------
>> Ready for the edge of your seat? Check out
tonight's top picks
>>  on Yahoo! TV. 
> 
> _______________________________________________
> Helix-producer-dev mailing list
> Helix-producer-devhelixcommunity.org
> http://lists.helixcommunity.org/mailman/listi
nfo/helix-producer-dev

_______________________________________________
Helix-producer-dev mailing list
Helix-producer-devhelixcommunity.org
http://lists.helixcommunity.org/mailman/listi
nfo/helix-producer-dev


       
---------------------------------
Building a website is a piece of cake. 
Yahoo! Small Business gives you all the tools to get
online.
_______________________________________________
Helix-producer-dev mailing list
Helix-producer-devhelixcommunity.org
http://lists.helixcommunity.org/mailman/listi
nfo/helix-producer-dev

Frank,

I am pretty sure what happens when you run multiple encoding
processes
in parallel, is contention over CPU resources, in particular
cache
memory. So I am not too surprised to see your numbers.  I do
not have
any good suggestions for speeding it up further than using
low complexity.

RV10  Low complexity is in fact the same as RV9,
algorithmically.

Karl.



on focus wrote:
> Karl,
> 
> thanks very much for the response and result to share.
I did some test
> based on Steve's suggestion and come up with quite
similar result to
> yours, which we are glad that is something at least we
can have some
> common base, i.e. for talking head sequence like CIF
Aikyo, RV10 with
> complexity low setting and keyframe setting at 15 can
do 520+ frames per
> second to encode bitstream of 225kbps  25fps
(audience setting is 256K)
> on a P4 Dual core 3.0GHz machine. But I have some
troubles to understand
> why the above result doesn't scale up:
> 
> We use the same machine with the above hw configuration
to encode 8
> channels of the same clip with same setting (by
launching 8 different
> encoder processes), the CPU is very much saturated and
the encoded clips
> have some frame loss (can only encode around 23fps).
Intuitively, 8
> channels = 8 * 25fps = 200fps which is only half of
500fps the machine
> can achieve. Where is the CPU spin on?
> 
> If what Karl's mentioning of file reading overhead
really matters, our
> test seems not echoing that statement because we tried
to read all
> YUV420 data into memory and keep using it for every
channel, it doesn't
> seem to matter that much (there are slight improvement)
to cover the 50%
> difference we observed.
> 
> we seems to run out of idea how to make sense of this
problem and would
> like to get some expert suggestion from you or anyone
in the Helix
> producer community who has some experience on this
issue. thanks,
> 
> Frank,
> 
> P.S., How can we configure the Helix producer to use
RV9 instead of RV10
> to see if there is any difference?
> 
> */Karl Lillevold <karllreal.com>/* wrote:
> 
>     The encoder also has several complexity levels that
can vary the
>     encoding speed up to 3-5X for real time adaptation
on slower systems.
> 
>     For the system mentioned, the file reading overhead
is going to be large
>     part of the time it takes to encode an uncompressed
test sequence.
> 
>     On my dual core 2 at 2.4 GHz, I get from 5 to 19 ms
per frame for
>     Foreman CIF, depending on chosen complexity, and
from 2.5 to 14 ms per
>     frame for Akiyo, not counting file reading
overhead.
> 
>     Karl.
> 
> 
>     Steve McMillen wrote:
>     > I recommend you do a file to file encode of an
input clip of a fixed
>     > duration and your resulting duration will tell
you how many fps were
>     > processed. I fully expect the result to be
< actual clip duration.
>     > That is, given that frame size and bit rate,
encoding should easily be
>     > faster than realtime on the box you
specified.
>     >
>     > On 8/5/2007 6:56 PM, on focus wrote:
>     >> Hi, Steve,
>     >>
>     >> We have come to a point that we need to
know more exactly what is the
>     >> encoding performance of RV10. We checked
the only available Rv10
>     codec
>     >> review .pdf file from Helix community/Real
website, there is a brief
>     >> mentioning of its speed, but not quite
exact (there is actually a
>     >> conflicting and confusing info there with
footnote) to let us decide
>     >> what machine we can properly choose for
our encoding task. Basically,
>     >> are there any more thorough or exact doc
that can be shared in Helix
>     >> community that gives:
>     >>
>     >> What is the encoding speed or performance
for encoding a target
>     of CIF
>     >> 352x288 25fps 256kbps
on a P4 Dual 3.0GHz machine or whatever
>     >> physical machine it could be targeted for
a standard test sequence,
>     >> like, Aikyo, or Foreman? The encoding
speed can be measured by how
>     >> many actual frames RV10 can encode per
second? or How many mips it
>     >> takes for encoding a frame averagely?
>     >>
>     >> Many thanks,
>     >>
>     >> Frank,
>     >>
>     >>
>    
------------------------------------------------------------
------------
>     >> Ready for the edge of your seat? Check out
tonight's top picks
>     >> on Yahoo! TV.
>     >
>     >
_______________________________________________
>     > Helix-producer-dev mailing list
>     > Helix-producer-devhelixcommunity.org
>     > http://lists.helixcommunity.org/mailman/listi
nfo/helix-producer-dev
> 
>     _______________________________________________
>     Helix-producer-dev mailing list
>     Helix-producer-devhelixcommunity.org
>     http://lists.helixcommunity.org/mailman/listi
nfo/helix-producer-dev
> 
> 
>
------------------------------------------------------------
------------
> Building a website is a piece of cake.
> Yahoo! Small Business gives you all the tools to get
online.
> <http://us.rd.yaho
o.com/evt=48251/*http://smallbusiness.yahoo.com/webhosting/?
p=PASSPORTPLUS>
> 

_______________________________________________
Helix-producer-dev mailing list
Helix-producer-devhelixcommunity.org
http://lists.helixcommunity.org/mailman/listi
nfo/helix-producer-dev