Embedded systems often operate in so variable conditions that design can only be carried out for some worst-case scenario. This leads to over-provisioned resources, and undue power consumption. Feedback control is an effective (and not yet fully explored) way to tailor resource usage online, thereby making the system behave and consume as if it was optimized for each specific utilization case. A control-theoretical methodology is here proposed to complement architecture design in a view to said tailoring. Experimental results show that a so addressed architecture meets performance requirements, while consuming less power than any fixed (i.e., uncontrolled) one capable of attaining the same goals. Also, the methodology naturally induces computationally lightweight control laws.