Where should distributed load balancing occur in a horizontally scalable architecture?