Extending the capabilities of mobile platforms through remote offloading over social device networks

Mobile computing is becoming the preferred method of personal computing for millions of users. In order to meet the increasing demands of computationally-intensive applications, recent mobile platforms have been augmented with multi-core CPUs, powerful GPUs, and special types of hardware accelerators. Despite these enhancements to the hardware of mobile platforms, their limited battery capacities and small form factor remain a bottleneck, hindering mobile platforms from utilizing their computing capabilities.;To address this restriction, there have been research efforts on remote offloading systems which seek intelligent ways to enable mobile platforms to leverage computing capabilities of more powerful resources over the network. Even though existing approaches provide core mechanisms to transform typical mobile applications to offloading-enabled applications, they still lack service discovery mechanisms while assuming the availability of remote computing nodes with static endpoints. Moreover, data privacy and secure communication between the mobile client and remote resources are of increasing importance for secure computing on mobile environments.;This dissertation presents a novel framework which enables remote workload offloading to external resources within a social virtual private network defined by the mobile user, in which trusted remote computing resources are aggregated in a virtual network regardless of user mobility. The proposed system accomplishes this by utilizing a peer-to-peer virtual private networking technique as a substrate for the discovery, configuration of trusted remote resources, and secure communication between the mobile device and remote resources. Based upon the evaluation on the performance of the proposed offloading framework, various mobile workloads are characterized for the suitability of offloading from the perspective of computation to communication ratio, which is a comprehensive measurement consolidating network conditions and workload characteristics.;In addition, this dissertation proposes applying machine learning techniques to runtime schedulers for mobile offloading frameworks. By adopting machine learning techniques to remote offloading scheduling problems, decisions on offloading do not rely on application-dependent parameters or predefined static scheduling policies. Instead, the scheduler can automatically learn the offloading effectiveness from previous offloading experiences and dynamically make decisions on whether mobile workloads should be offloaded or executed locally according to the current conditions at runtime.