I am not yet multi-threading. Atm i do the update every frame. It is not much work, just a view frustum check over the bounding boxes. Algorithm i use is cdlod. (Edit: i should say strongly based on ...) The quad tree is tile specific and is built while loading a tile. Only the selection of the nodes to draw is done every frame. CPU side is not the bottle neck.

I'm using kind of a chunk system. It's splits or merges chunks as you move around and also splits or merges all triangles in the those chunks to match.  A chunk is considered some point in the overall tree and everything under it. Each node in the tree has a bounding volume and as you go down the tree it calculates at what level a chunk should be, based on the square of the distance from the bounding volume to the camera.  Then there are rules......

For instance if you reach a point in the tree that should be the top of a chunk and it is in fact a chunk top, then you don't have to go further. If you reach a point where there is a chunk top but the chunk top should be further down in the tree based on the distance calculation, then you have to split the chunk into multiple chunks. Conversely if you reach a point where there should be a chunk top but you haven't encountered one yet on your way down the tree, you have to merge all chunks below that point into a new chunk that you create at that point in the tree. 

This way if you just move the camera just a little and nothing has changed, you only have to go down to the top of each chunk and not to each triangle within the chunk. You can see it in action in the first part of the video in my last blog entry. The system also works for voxels using an octree, but this one uses a quad-tree since it's dealing with height-mapped terrain albeit wrapped around a sphere. It's really just flat terrain here because it's a sun, but it would work for other terrain too.

A thing i like about doing selection each frame is that one can for example change depth or field (far plane - near plane) on the fly and recalculate (within limits) the depths and transition areas for each lod level in between frames. So that the view to the horizon from a mountain top has different lod levels (provided the tiles with height data are loaded) than one needs when driving through a narrow valley.

A moon doesn't have graceful fog ....