The main cost in the destructor is the delete[] for each array aiFace::mIndices. We already had an experimental bulk allocator in place that allocated linearly and never released the memory until the Importer instance was destroyed. This cut about 20% time on reading the scene and nearly all time on destroying the scene.

I think you might be able to re-enable this hack somewhere in the source, but I'm not sure what its current state is.