GRACE: A Hierarchical Cross-Layer Adaptation Framework for Saving Energy

Mobile devices running soft real-time multimedia applications are becoming an increasingly important computing platform. Such systems are often limited by their battery life; therefore, minimizing their energy usage has become a primary design goal. A widely used technique to save energy is to adapt the system in response to changes in application demands and resource availability. Researchers have proposed such adaptations in all layers of the system; e.g., hardware, application, operating system, and network. A system with multiple adaptive layers requires careful coordination of these adaptations to reap their full benefit. The Illinois GRACE project (Global Resource Adaptation through CoopEration) has proposed such a coordinated cross-layer adaptation framework [4, 2]. A cross-layer adaptive system must balance the conflicting demands of adaptation scope and time scale. Ideally, it should invoke both global and frequent adaptation that coordinates all layers in response to all changes in the system. Unfortunately, global adaptation can be expensive and so must be infrequent, but long intervals between adaptation risks inadequate response to intervening changes. To balance this conflict, the GRACE framework has proposed a hierarchical approach, performing expensive global adaptations occasionally, and inexpensive limited-scope adaptations frequently [4, 2]. Specifically, the GRACE framework proposes three levels of adaptation, exploiting the natural frame boundaries in periodic real-time multimedia applications (Figure 1 [2]). Global adaptation considers all applications and system layers together, but only occurs at large system changes (e.g., application entry or exit). Per-application adaptation considers one application at a time and is invoked every frame, adapting all system layers to that application’s current demands. Internal adaptation adapts only a single system layer (possibly considering several applications) and may be invoked several times per application frame. All adaptation levels are tightly coupled, ensuring that the resource allocation decisions made through global coordination are respected by the limited-scope adaptations. We have implemented GRACE on a Pentium M based laptop system running Linux 2.6.8-1, with a real adaptive CPU, fully implemented adaptive applications and soft real-time CPU scheduler, but a simulated network layer [3]. For CPU adaptation, we use dynamic voltage and frequency scaling on the Pentium M processor. For applications, we use H.263 video encoders and perform adaptations first proposed in [1]. These adaptations vary the CPU time and network bandwidth requirement of the encoder, to minimize the sum of the CPU and network transmission energy, without perceptibly changing the video quality. Our implementation incorporates global and per-application adaptations in the CPU and application, along with global and internal adaptations in the soft real-time scheduler. It respects the constraints of CPU utilization and network bandwidth, while minimizing CPU and network transmission energy. GRACE’s global optimizer chooses, for each application, an application configuratime la ye r