Racing and Pacing to Idle : Minimizing Energy Under Performance Constraints

The problem of minimizing energy for a realtime performance constraint has been widely studied, both in theory and in practice. Theoretical models have indicated large potential energy savings, but practical concerns have made these savings hard to realize. Instead, practitioners often rely on the race-to-idle heuristic, which makes all resources available to a task and then idles the system until the next task is released. While this heuristic has proven effective, recent results indicate that more sophisticated resource allocation schemes may now provide greater energy savings. This paper investigates resource allocation heuristics for real-time constraints using both analytical and empirical techniques. We formalize the problem as a linear program and develop a geometric interpretation, allowing derivation of the optimality conditions for various heuristics. We then demonstrate that the pace-to-idle heuristic is often better and never worse than race-to-idle. We confirm these analytical results by implementing a resource allocator based on the studied heuristics and measuring energy consumption for eight different applications on four different systems. The results confirm that pace-to-idle produces better energy savings than race-to-idle, by up to 20% on the newest platform in our study.