kd-trees are used by default, and are created automatically.
I think this statement is clearly enough ;)
You don't need to use the Octree anymore, you can use it if you want but kd seems to be faster.
Thats the magic of VBOs
A small summary about VBOs :
Normally without VBOs you have to resend the vertexdata each time the model should be rendered. With VBOs the Vertexdata is stored onto the Gfx card and you only have to send a small int to say which data should be rendered. This saves a lot of bandwith and this is what makes this so fast. Btw: If you would make the spheres single surface instead of having 1000s of single objects your app will running again much faster than now ;)