|
|
Steven Pigeon wrote:
> I think to minimize computation, tiles would be preferable.
> I think of radiosity and photons which are not local at all in
> nature: they use large parts of the scene.
>
>
> For example, having a bunch of render-slaves computers, one
> could be master and dispatch tasks as the CPUs become
> available: once a CPU is done with is tile, it requests another
> region of the picture (without flushing its own data, that'd be
> a shame to start from scratch with photons and radiosity.) The
> master collects tiles and builds the final picture.
>
> I would much rather see that master/slave thing implemented
> in POV than using a script-based dispatcher.
>
> Best,
>
> S.
Wow, what timing :) I've been working feverishly over the past two
weeks on MegapovXRS, which is similar to what you describe.
I've essentially retrofitted MegaPOV 1.0 with an XML-RPC server. This
provides a multiplatform way for a remote client (the master) to send
work requests via the internet or a LAN using http/(and possibly https).
Ideally, the authors of SMPOV, POV-Anywhere, Rendview and others could
add a module that supports rendering with a MegapovXRS slave.
The way it works is the master opens a new session, delivering
"command-line" arguments over RPC. Then, the MegapovXRS slave parses the
scene, shoots any photons and then waits for tile requests. When tiles
come in they are rendered *without* having to parse or shoot photons
again. This should provide a substantial boost (compared to a pure
script based approach) when it comes to network rendering scenes with
long parse times.
I have a highly functional beta here and I'm working as we speak on
preparing it for release to my beloved POV-Ray community for review.
Best Regards,
George Pantazopoulos
http://www.gammaburst.net
Post a reply to this message
|
|