Rendering entity Management

If you have a few hundred entities, running the bounds-vs-view-frustum test per entity isn't that wasteful. It's certainly faster than actually drawing each entity!

When you get into the thousands, you want some spatial subdivision, such as a loose octree (google for that) where your entities go. You then traverse the octree with a frustum culler, which only visits the nodes that are within the frustum. This culls a lot of entities by just culling their parent nodes in the octree.

For the entities that move, they have to update their position in the tree before the render pass.

Another alternative is to arrange bounding volumes into a hierarchy of nested bounding volumes or what is known as a bounding volume hierarchy (BVH), it's a spatial data structure but it's not a spatial partitioning type like octrees etc. For visibility determination it is unnecessary to do exact determination so it's best to keep all geometry at leaves only and bounding volumes in both internal & leaf nodes. Large static geometry can be further subdivided with what ever spatial data structure you want.

Generally an octree is the best structure, but you can use a containing heirarchy as a fallback, organising your entity list into a cohesive scenegraph.

How this is done is a matter of preference, and will depend to some extent on the types of objects you're using (large terrain chunks / particle systems / standard object meshes):

You can ensure all geometry is only in leaf nodes (as snk_kid suggests).

Or you can allow entities to attach to any hierarchical point, and adjust bounding volumes accordingly.

Cool, thanks guys.