An Analytical Model for Energy Minimization

Abstract

Energy has emerged as a critical constraint in mobile computing because the power availability in most of these systems is limited by the battery power of the device. In this paper, we focus on the memory energy dissipation. This is motivated by the fact that, for data intensive applications, a significant amount of energy is dissipated in the memory. Advanced memory architectures like the Mobile SDRAM and the RDRAM support multiple power states of memory banks, which can be exploited to reduce energy dissipation in the system. Therefore, it is important to design efficient controller policies that transition among power states. Since the addressed memory chip must be in the active state in order to perform a read/write operation, the key point is the tradeoff between the energy reduction due to the use of low power modes and the energy overheads of the resulting activations. The lack of rigorous models for energy analysis is the main motivation of this work. Assuming regular transitions, we derive a formal model that captures the relation between the energy complexity and the memory activities. Given a predetermined number of activations, we approximate the optimal repartition among available power modes. We evaluate our model on the RDRAM and analyze the behavior of each parameter together with the energy that can be saved or lost.