It is an arrangement of cache servers in a tree structure usually remote from each other. So you could have a large server or array at head office, some smaller ones in the regional centres, and some babies in each branch for example. Each one uses the next higher one as a parent and requests it to fetch data when necessary on its behalf.

This maximizes effectiveness and we have examples where this structure improves cache hit rates by over 10% in comparison to single large installations. This also has a number of other benefits. Redundancy is good because you can go direct if the server above you is down or to another server in the same layer. If employed in a large WAN environment it reduces traffic by 30 to 40% across the expensive links.