clipka <ano### [at] anonymousorg> wrote:
> The interface used for message passing does not conform to the MPI
> standard defined by the Argonne National Laboratory, but uses a
> proprietary API: The thing called POVMS, which - as Thorsten has already
> noted - comes in both a C and a C++ version (the latter being
> implemented as a wrapper around the C API).
>
> If your goal is not to just throw together a quick hack to run POV-Ray
> on multiple nodes, but to create some decent solution that future
> development can build on, maybe you should actually work on the POVMS
> implementation itself to get it flying across node boundaries. (Or, as a
> potential alternative, rip out POVMS entirely and replace it with some
> widespread and well-maintained MPI library.)

One reason to create POVMS was that MPI at the time (1997/1998) was not really
good at handling multiple threads (especially cooperative threading like Macs
had at the time) and implementations existed mostly for Unix, yet multithreading
is a huge benefit for ray-tracing. Seems like this has not changed a lot since
then, i.e.
http://www.open-mpi.org/~jsquyres/www.open-mpi.org/doc/v1.8/man3/MPI_Init_thread.3.php
still says it is no good for heavily multithreaded applications :-( So
effectively one would need to have a three layer communication instead of a two
layer communication in POV 3.7 to use MPI: MPI processes of POV with a central
server process controlling rendering and collecting results - POV processes each
with single MPI communication thread that bundles local POVMS messages - POV
threads that use POVMS for in-process communication.

As for POVMS message data exchange, it will work by simply putting POVMS message
streams into MPI byte arrays. POVMS streams already handle byte ordering, float
type conversion etc., so it should be straight forward.

Post a reply to this message