Racing and pacing to idle: an evaluation of heuristics for energy-aware resource allocation

We examine the problem of assigning computing resources to an application to meet a performance goal while minimizing energy consumption. We present a general formulation of this problem as a linear program, discuss several potential heuristic solutions, and evaluate these heuristics on two real systems (one purchased in 2010, the other in 2013). We find that the well-known race-to-idle heuristic is close to the optimal solution on the older machine. On the newer machine, however, the optimal solution outperforms race-to-idle by over 35%. A generalization of race-to-idle, called pace-to-idle, is found to provide better results in a wider range of scenarios.