Thanks for the reply, yes I have been reading your posts, they have been very insightful. I want to clarify some aspects: the limitations on one submesh per model is just temporary. I have been experimenting with regular teapots and procedurally generated solid primitives, but of couse the engine needs to handle that situation.
Regarding the scene tree vs graph, from my understanding of data structures, a tree has to have only one root node. A graph does not have this restriction. Furthermore a tree can't have as children nodes that jump levels in the hierarchy. I.e. node A can't have as child node B who is a sibling of A's parent. There is also only one parent for each node. So for all logical purposes I'm using a scene tree and not a graph. Isn't this how it is supposed to work? What advantages would a graph have over a hierarchical tree? In my current implementation the absolute world matrix of any object belonging to the tree is computed by traversing the tree from the root to the leaf node.
A tree of meshes is an interesting concept. Yes it makes sense. I didn't think about that because as I said I was only dealing with very simple models but now that I think of it, it makes sense as that is the logical representation one would expect to find in a model. So would it be fine to have a method that allows this tree of meshes to convert itself to an appropriate branch to be inserted in the scene tree?
I'm using my tree just as a logical representation of the objects in the scene. It is not used for rendering. It is only needed to update the world matrices of the objects in the scene. I have a CommandManager class that is invoked to compute the most efficient way of rendering the scene. The first time it traverses through the scene tree putting each mesh in a bin according to its material. Then each (flattened) collection of meshes is attached to a RenderCommand (and specialized versions such as InstancingRenderCommand) which renders all objects having the exact same characteristics. The same class also checks for state changes and issues state change commands if needed. Each material defines what the collection of objects it will render should look like. I.e. their vertex layout, the blend/rasterizer/depthStencil state it expects, the shadow algorithm associated with it (if any) and so on.
To display the scene I call CommandManager.Run() and it goes through its linked list firing all commands. Every objects ultimately draws itself. My approach is to have all similar objects drawn together in order to minimize state changes, buffer rebindings and so on.