[FFmpeg-devel] Parallelizing MPEG encoding with MPI

Måns Rullgård mans
Sun Dec 7 13:56:15 CET 2008


Andreas Simbuerger <simi at osaft.eu> writes:

> On Sun, 2008-12-07 at 03:10 +0000, M?ns Rullg?rd wrote:
>
>> Search again.  I can assure you it's there.
>> 
>
> Before posting my question i should have slept over it again.
> Stood up today and just found it. *grml* :-)
>
> In theory the MPI approach leads to a high encoding performance
> gain especially on clusters. I can add more detail to this next
> week, when i heard a lecture from the doctoral candidate about
> the planned scheduling. My work should test this theory in a "dirty"
> environment.

When I was experimenting with this, I got close to linear speedup with
5-10 nodes.  These were mostly 1GHz P3 machines.  Faster nodes will
obviously increase the relative communication overhead.

> Concerning the slicing threads approach described in
> (http://akuvian.org/src/x264/sliceless_threads.txt ), i think this
> approach leads to a lot of synchronization work between those
> threads and limits you in the possible effect of parallelism because
> the slowest threads blocks the others.

Such blocking can sometimes be masked by have a few more threads than
nodes.

-- 
M?ns Rullg?rd
mans at mansr.com




More information about the ffmpeg-devel mailing list