However, optimization comes from "re-coupling" very often. see for example : monolithic kernels (monolithic, fast) vs micro kernel (decoupled, slow). If you really want something efficient, you'll have to think about it from the very beginning.