2. Develop a "driver" to test the cluster
3. Combine clusters and move upward
4. Finally, interface with top control module
This approach has the advantage that "worker" modules are tested early, and there is no need for stubs and/or dummy simulation of data
If module A exists but B and C do not exist, B and C are replaced by stubs (a dummy program that does nothing exceept announcing it is being invoked) so that you can test module A without having modules B and C.