A new framework for hierarchical cross-layer adaptation

Battery technology has not kept the pace with innovations in microprocessors and wireless networks, and as a result saving energy is increasingly important on modern wireless-equipped laptops. To save energy, one technique we turn to is reconfiguration, or adaptation, of system components to permit more efficient operation. Our adaptive GRACE system provides for coordinated adaptation across all system layers and applications through the use of an adaptation hierarchy. Resources are broadly allocated to applications by a global adaptation layer, which considers all applications and system layers but is relatively expensive and can therefore run only infrequently. These allocations are refined and converted into adaptation decisions on a job-by-job basis using a cross-layer "per-application" adaptor that considers only the current application's demands and allocations. Finally, each system layer optimizes itself independently within the application's allocation to minimize its energy consumption. To validate this approach, a simulation of the GRACE system was constructed around a custom-built adaptive video encoder that provides the system the ability to reduce its CPU utilization at the cost of increased network-bandwidth requirements. These simulations show that the GRACE architecture can save more than 50% of the total CPU and network energy by appropriately shifting demand from the CPU to the network. The GRACE system also allows utility to be allocated between applications. Conventionally, this is done using suboptimal heuristics. However, Lagrangian techniques can also be applied to the allocation problem, supplying a method whereby optimal allocations (to within convex-hull approximations) can be made without requiring exact foreknowledge of upcoming workloads or searches of the cross-product of present and future applications. Through the use of Lagrangian optimization, unbounded improvements in total utility can be achieved compared to the constant-workload heuristic; simulations demonstrate a factor of two improvement in total utility in advantageous circumstances.