| Besides light-weight Internet applications, computation-and/or energy-heavy appli-cations are increasingly deployed to mobile devices. However, the limited computation power, battery lifetime and network connectivity make it difficult for mobile devices to become the center of mobile computing. Mobile offloading, or remote execution, between a mobile device and a capable server has proven effective in improving mobile applica-tion’s performance.A mobile collaborative application consists of several tasks, some of which are "of-floadable" that can be offloaded to the collaboration server for execution. Lying at the core of a mobile offloading system is the offload engine that is responsible for determin-ing when and where to execute each task—i.e., locally on the mobile device or remotely on the server—based on resource availability and application’s performance requirements. For example, in order for the mobile device to offload a computation-and/or energy-heavy task to the server, its execution state must first be sent to the server, and the execution re-sult should later be returned to the mobile device. Fluctuations in network condition and server workload may affect data transmission time and server response time. The offload engine must thus carefully weigh the expected gain in the device’s energy consumption and the task completion time against the cost of network communication.In the thesis2, collaborative application’s execution is modeled with a graph, where the execution is decomposed into different stages which are connected via transaction-s. The execution of an application may take different paths, each of which consumes different amount of time and device energy. The offload engine must select an optimal execution path, balancing between resource availability and performance requirements.Existing offload decision-making strategies usually assume that resource availability—e.g., network bandwidth and server workload—remains unchanged throughout path exe-cution. This assumption yields non-adaptive offload decisions, which, once determined, remain unchanged until the execution is completed or a failure is detected. However, ac- tual resource availability often exhibits high variability on the time-scale of seconds and also varies with location. The highly variable condition of the underlying resources dic-tates computation and communication costs, possibly causing long application execution delays (even failures) or missing opportunities to improve application performance.Realizing the impact of resource-availability variations on the application execution performance, the thesis proposes a set of novel adaptive offload decision-making strate-gies, trying to minimize the risk of application execution delays or failures as well as to improve application execution performance and device energy performances by oppor-tunistically exploiting resource-availability improvements, if any. To be more specific, the thesis proposes following contributions.·Dynamic decision making strategy is proposed to enable offload decision-making at runtime to (re)-evaluate the impact of resource-availability changes on appli-cation performance and then change the execution path accordingly on-the-fly. This enables mobile application’s responsive and flexible adaptation to resource-availability variations.·Redundant execution is proposed to enable application execution using multiple execution paths so as to better explore performance improvement opportunities. Redundant execution enables mobile collaborative application to be executed with two execution paths simultaneously, whichever result comes first is used. Offload engine can also introduce probability-based decision-making approach, account-ing for the risk of incorrect decisions to minimize its impact on application execu-tion.·Joint decision-making strategy is proposed to enable mobile client to share its execution performance requirements with the server, thus permitting the server to advise on an offload decision.·Server may also enable incremental delivery of results, so as to reduce the execu-tion delay/failure due to the fluctuating wireless network condition.The thesis studies the design, implementation of adaptive-decision making strate-gies based mobile offloading systems, Wing and Mind. The thesis evaluates the decision quality of adaptive decision making strategies in coping with server workload variance and network connectivity variance. Wing is based on dynamic decision making strategy and redundant path execution while Mind is based on joint decision making strategy and incremental data delivery mechanism. Two mobile applications, mobile face detection and mobile panorama, are created as target applications for evaluation.A controlled experiment is designed to evaluate the quality of adaptive decision making strategies in adaptation to collaboration server workload variance. This synthetic experiment is intended to help to understand system’s behavior in controlled scenarios rather than precisely emulate actual server workload variance. Our experimental results indicate that Wing can achieve30%(50%) reduction of application execution time and30%(50%) reduction of device energy consumption. The overhead incurred by Wing is trivial, which is within10ms when performing per runtime decision making.A trace-based emulation experiment is designed to evaluate the quality of adaptive decision making strategies in adaptation to end-to-end network bandwidth fluctuations between mobile device and collaboration server. The use of real-world traces provides experimental repeatability and allows a careful comparison among different strategies. The network trace is collected in the north and central campus of University of Michigan and in Ann Arbor city, covering both stationary scenarios, e.g., offices, public cafeteria and walking scenarios3. The total length of WiFi network trace is around2,400minutes, including over10,000records and decision making points for evaluation. Furthermore, the experiment evaluates the impact of4basic types of most commonly used network bandwidth estimators (Spot, Ody, Avg, Medium) on decision quality. Experiment results show that comparing with non-adaptive decision making strategy, dynamic decision mak-ing strategy with single path execution can reduce incorrect decision rate by an average of40-50%while achieving a30%(40-50%) reduction of application execution time (device energy consumption) over the non-adaptive offload decision-making strategy. With redun-dant path execution, dynamic decision making strategy is able to achieve30%reduction of application execution time and30%reduction of device energy consumption. Evaluation shows that joint decision making strategy with incremental data delivery can reduce in-correct decisions by an average of10-20%over the non-adaptive offload decision-making strategy.In conclusion, the thesis proposes a set of adaptive decision making strategies to cope with resource dynamics fluctuations in mobile offloading between a mobile device and a collaboration server. As far as is known, this is the first effort trying to address the impact of highly-variable resource availability on the application execution performance. The synthetic experiment and in-depth evaluation based on real-world network traces indicate that comparing with non-adaptive decision making strategy, adaptive decision making strategies can significantly improve application execution performance in terms of execution time and device energy consumption while reducing execution delays and/or failures. The proposed strategies and underlying approaches are applicable not only to mobile offloading but also to other research work where adaptation to resource changes is required to ensure performance. |
PDF Full Text Request