|
 |
Ken wrote:
>
> What is the basic consensus for render farming a Pov scene. Should the
> scene be split to render n/lines per machine in the cluster or should
> there be a sharing of resources used for a single frame render. Which
> will be faster ?
>
> --
> Ken Tyler
>
> See my 700+ Povray and 3D Rendering and Raytracing Links at:
> http://home.pacbell.net/tylereng/index.html
Well Ken,
I don't know of any consensus, honestly I didn't get the difference
between the options you gave.
But anyway, there are a few aspects of parallism that influence the way
to distribute rendering.
I) Parallism gets better with more computing time over communication
time.
II) Also tasks done multiple times are bad.
III) Another bad point would be idle processors.
III) gets important if the task is divided onto parts of different
duration. That is, you have 6 processes, do 6 parts, 5 of them take a
second, the 6th takes an hour.
II) gets done in pvmpov for example, each process does parsing and scene
creation.
I) gets into play if you parallise to fine, communicating every pixel at
once would produce a severe overhead.
The major conclusion: It depends. If your pps are equally distributed
over the image, you may do quite large parts. If they are very irregular
distributed, you need to use fine grids.
If you have a large animation, you probably don't have to parse the same
sceen in different processes (or distribute a parsed scene), so going
frame by frame is good. You may go for finer granularity for the last
frames, to reduce idle times.
But this heavily dipends on the ratio of parsing time vs. rendering
time.
My personal decision is, I don't like any of the attempts I have seen,
(having seen all mentioned in this thread), because they are not
flexible enough for such a tricky task as rendering. Handtune is the
only solution. At least as long as povray does not provide modularity,
which could shift some decisions (reduce parsing time to almost zero in
animations and the like).
Axel
Post a reply to this message
|
|
