Efficient behavior-driven runtime dynamic voltage scaling policies

Power consumption has long been a limiting factor in microprocessor design. In seeking energy efficiency solutions, dynamic voltage/frequency scaling (DVFS), a technique to vary voltage/frequency on the fly, has emerged as a powerful and practical power/energy reduction technique that exploits computation slack due to relaxed deadlines and memory accesses. DVFS has been implemented in some modern processors such as Intel XScale and Transmeta Crusoe. Hence the bulk of research efforts have been devoted to developing policies to detect slack and pick appropriate V/f assignments such that the energy is minimized while meeting performance requirements. Since slack is a product of memory accesses and relaxed deadlines, the number of instances and the duration of available slack are highly dependent on the runtime program behavior. Runtime DVFS policies must take into consideration program characteristics in order to achieve significant energy savings. In this paper, we characterize program behavior and classify programs in terms of the memory access behavior. We propose a runtime DVFS policy that takes into consideration the characteristics of program behavior for each category. Then we examine the efficiency of the proposed DVFS policies by comparing with previously derived upper bounds of energy savings. Results show that the proposed runtime DVFS policies approach the upper bounds of energy savings in most cases.