Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!watmath!clyde!rutgers!mit-eddie!husc6!ut-sally!utah-cs!utah-gr!peterson
From: peterson@utah-gr.UUCP
Newsgroups: comp.graphics
Subject: Re: Ray Tracing & filter references
Message-ID: <1908@utah-gr.UUCP>
Date: Tue, 3-Feb-87 15:58:52 EST
Article-I.D.: utah-gr.1908
Posted: Tue Feb  3 15:58:52 1987
Date-Received: Wed, 4-Feb-87 07:18:53 EST
References: <179@ndmath.UUCP>
Organization: University of Utah CS Dept
Lines: 31
Keywords: references, ray tracing, box filters
Summary: Making pictures on multiple machines

> One other thing...I would appreciate comments from any graphics hackers on
> the workability of splitting up complex Ray Tracing or other types of hidden
> surface rendering algorithims between a number of CPU's for speed in process!
> I expect to have access to an ethernet with a large number (20 or more)
> of identical Unix machines and I would like to experiment with breaking a
> picture up in pieces and giving each machine a chunk to work on.
>

If your making movies this is easy - just dole out frames to individual
processors.

For making a single complex still, the issue becomes much more complex.
First, you will usually lose if you just hand out equal portions to
various processors.  In most images, some complex scanlines may take 5-10
times as long as those in simpler areas of the image, particularly if
you're doing adaptive anti-aliasing.

One approach I've used is to render a small (say 128x128 pixel) copy of
the image, and keep track of the times required for each scanline.
Then use these scanline times as guide for splitting up the work on the
big picture.

Another approach - if you're really into network hacking - is to make
the ray tracer work on small chunks (say 1-10 scanlines) as a client
process.  A single server on the network then hands out chunks of the
rendering work to the ray-tracing clients until the picture is finished
(I think Apollo may make their SIGGRAPH movies this way).

If you use the first techinique, I strongly recommend generating an
alpha (matte) mask for indicating which pixels are finished.  Then all
you have to do to put pieces together is feed them to your compositor.