[MPlayer-users] Distributed Processing

[Rich Felker]

On Fri, May 06, 2005 at 10:30:46PM +0200, Reimar Döffinger wrote:
> Hi,
> On Fri, May 06, 2005 at 03:06:50PM -0500, Kichigai Mentat wrote:
> > On May 6, 2005, at 10.33, Rich Felker wrote:
> > >Still splitting it across different machines will make it slow as
> > >hell.
> > Do you mean that due to the process of forcing lavc to parallelize  
> > over multiple machines, it will be as slow as hell, or that in  
> > theory, parallelizing a mencoder job period would be slow as hell?
> 
> Well, there is lots of data to shuffle around, so you would at least
> need an extremely fast network - since I doubt that you have something
> like Myrinet or infiniband lying around I'd say there are more effective
> ways of speeding things up.

No, the network needs to be infinite speed in order not to slow it
down. The entire encode process is memory-io bound; processing takes
essentially 0 time. Even if the network is as fast as main memory (in
practice it's at least 10-100 times slower) it would still make
encoding take twice as long.

If you want to distribute encoding, there are some decent ways to do
it, like encoding audio on a separate machine then muxing later. No
special distributed-computing crap is needed for this. Just run
separate processes on each computer.

Rich