Abstract
The cache subsystem is a major contributor to energy consumption in commercial microprocessors used in embedded systems. To reduce energy, designers can perform design space exploration (DSE) to determine a suitable cache configuration that matches system constraints and goals while minimizing energy consumption. Traditionally, this cache tuning step has been a static process where heuristics or analytical models are used to determine an optimal or near-optimal cache configuration prior to runtime given a known application, application set, or application domain. Even though the configuration may change during runtime for different phases of execution, the specific configuration for each phase remains fixed. This static nature is too restrictive for modern, complex embedded systems that are expected to operate under diverse, unknown operating environments, run unknown applications, and with vastly different user quality of experience (QoE) expectations (e.g., smart phones). Therefore, cache tuning must change from a static optimization process to a dynamic optimization process that adapts online during runtime transparently to the user/system needs. The key challenge is determining the configuration that adheres to QoE expectations while minimizing energy consumption without degrading the user experience during DSE. Despite the wealth of progress that has been made, the realization of a closed-loop, fully adaptive, online-tunable cache subsystem still faces many challenges. In this paper, we review the progress made in the area of static and dynamic cache tuning, discuss the challenges that still exist in this area, and propose a predictionassisted control-theoretic framework to address these challenges.
