From what i see when rendering a single still image with 2 computers the LB makes a linear split on the x/y axis of the image. With uneven distributed complexity of a scene it would be more efficient to split the workload like a checkers/chess board so everyone get's a complex part without 'learning' from multiple runs.
Cheers
HI Martin, you're spot on, and that is happening with our next gen system we're currently building. There is an increased complexity in doing this, which is not actually related to the maths of splitting the screen, but more to do with the fact that to render a tile, we need to launch a background process of blender which has significant overhead. So to benefit from the speed, we're in research mode right now as to how to keep this overhead from eating too much of the gains of using smaller tiles.
We're hopefully going to be able to eventually find a way to launch a single render process that will persist and be able to accept commands to render tiles as needed which will negate this overhead entirely and give us back the lost speed from the process open/blend file load overhead.
Initially however, we'll experiment with a system that uses regions, so splitting the image up into larger regions and then using the current system within each region.
We'll keep you all posted with how development is coming along for that :)