The Case for Complexity Prediction in Automatic Partitioning of Cloud-enabled Mobile Applications

As application demands out-pace the evolution of battery technology, many smartphone “app” developers will soon explore offloading compute-intensive tasks to the cloud. Such cloud-enabled mobile applications effectively partition application functionality between the phone and the cloud. Application partitioning must be dynamic, to successfully adapt to variability in resource availability. Dynamic partitioning systems rely on the ability to predict an application’s component’s resource usage, for which prior work has used simple approaches. In this paper we propose the use of complexity metrics that enhance these predictions by taking into account relevant properties of each component’s input, both generalpurpose (e.g., size) and type-specific (e.g., number of words in an audio sample). Our predictors improve the energy efficiency of partitioning a speech recognition library by 21% or more.