Folklore Confirmed: Compiling for Speed = Compiling for Energy

As we move towards exa-scale computing, energy is becoming increasingly important, even in the high performance computing arena. However, the simple equation, Energy = Power × Time, suggests that optimizing for speed already optimizes for energy, under the assumption that Power is constant. When power is not constant, a strategy that achieves energy savings at the cost of slower execution is Dynamic Voltage and Frequency Scaling (DVFS). However, DVFS is currently applicable only to the processor, and the entire system has many other sources of power dissipation. We show that there is little to gain in compilers by trying to trade off speed for energy using DVFS. It is best to produce code that runs full-throttle, completing as quickly as possible, an approach called “race to sleep.” Our result is based on analyses of a high-level energy model that characterizes energy consumption, related to survey of power consumption trends of recent processors for both desktop and server, as well as Cray supercomputers.