An evaluation infrastructure for power and energy optimisations

This paper presents an infrastructure that uses program phase behaviour and brings together the advantages of simulation and physical measurement to capture detailed program power behaviour. Based on the direct correlation between instructions-per-cycle and power dissipation for the same piece of code, we refine the phase classification used in SimPoint to improve the similarity among the intervals in the same phase in terms of time-dependent power behaviour. Experimental results show that this classification method significantly decreases the average relative standard deviation across the intervals within each phase by 68% with respect to the average power dissipation of the intervals.