Improving Adaptive Offloading Using Distributed Abstract Class Graphs in Mobile Environments

Adaptive offloading dynamically distributes portions of a computationally heavy application to remote devices to achieve context specific optimisations. However, since existing state-of-the-art approaches incur significant overhead from storing, updating and partitioning application graphs this paper proposes a novel distributed approach to alleviate much of this overhead. Specifically, each device maintains a graph consisting only of components in its own memory space, while maintaining abstraction elements for components in remote devices. This approach removes the need to store and update complete application graphs on each device and reduces the cost of partitioning an application during adaptation. An evaluation involving computationally heavy open-source applications adapting in a heterogeneous collaboration showed that the new approach reduced graph update network cost by 100%, collaboration-wide memory cost by between 37% and 50%, battery usage by between 63% and 93%, and adaptation time by between 19% and 98%, while improving efficacy of adaptation by 12% and 34% for two of the considered applications.