Research On Performance Optimization To Interactive Mobile Applications In Mobile Cloud Computing

In recent years, smart mobile devices become more and more popular, which leads to the booming of mobile applications. These mobile applications with powerful func-tions demand more and more capacity on mobile devices, which forces users to fre-quently upgrade their devices. However, the restricted resources on mobile devices are still the bottleneck of the enhancing the performance of these applications. Along with the development of mobile networks and cloud computing technology, researchers try to offload some compute-intensive part of mobile applications onto the cloud, in order to harness the unlimited resources on the cloud for enhancing the user experience of mobile applications. With the accumulated of these research works, people call the new research area as Mobile Cloud Computing (MCC). Researchers achieve exiting results from theory analysis, framework design, key technics design & validation, practical application, etc.Based on the former researchers, this thesis focus on the interactive mobile appli-cations, which are the dominant part of mobile applications. By analyzing the charac-teristic of interactive mobile applications, we specially design and implement an MCC platform AppBooster and its specially designed performance optimization framework for these applications. As these interactive applications are sensitive to the response time and likely involve complicated calculation and analysis, some application-specific performance metrics (e.g., recognition rate, accuracy) may largely restricted to the ca-pacity of mobile devices while developers designing the applications. In this thesis, we propose that dynamically adjusting these application-specific quality together with the offloading of computation, by tuning some tunable parameters (e.g., number of itera-tions, size of dataset) on the fly. By evaluating the platform with a demo application, which is an imaged-based object recognition, we found that the demo application can achieve much higher performance than ordinary strategy.In the optimization framework of AppBooster, the dependency of application-specific and general performance metrics on the tunable parameters and offloading scheme are automatically learned by machine learning technique, which is a key point as well as a difficult point. The main challenge comes from the inevitable sparsity of historical execution records in a realistic scenario. In response to this challenge, we di- vide an application’s performance model into several sub-models, which are separately trained offline and assembled as a whole model for performance prediction online. In this way, the performance model can be more robust to sparse data. Moreover, we adopt auto-learned feature vectors to reflect the similarity between different modules as well as different devices, which can reduce the negative effect of the lack of exe-cution records from new published applications and new emerged devices. A series of experiments show that our proposed framework of performance prediction in MCC can handle the sparsity of history records in a realistic scenario.Besides the prediction of application performance, the decision of offloading schemes is another key point of the optimization framework of AppBooster. Based on the pre-dicted performance, we designed a general genetic algorithm for deciding offloading schemes, which is integrated with the decision of several distributed speedup technics (e.g., data parallel, pipeline). The experiment results show that this general algorithm is able to provide good decision of offloading scheme for the majority of interactive mobile applications and fulfill their real-time requirements as well. For those interac-tive mobile applications based on streaming data, we propose that those users harness multiple cloudlets at one time, in order to maximize the use of their bandwidth. We specially designed a two-phase unified decision algorithm for multiple users whom are competing resources on multiple cloudlets. The two-phase algorithm is consists of a distributed offline preprocess and a centralized online heuristic, which achieves a good real-time capability and scalability. The experiment results shows that this algorithm can provide near-optimal decisions for all users in a multi-cloudlet scenario.By analyzing the special requirements and characteristic of interactive mobile ap-plications, we implement an MCC platform and proposes a special designed optimiza-tion framework for them. After a series research on the prediction of performance and the decision of offloading scheme, which are the two key points of the framework, we designed effective frameworks and algorithms for the tasks, and evaluated them on a real-world MCC platform with a real application. By summing all the experiments re-sult, we believe that the research works in this thesis can help and guide the industry to optimize the performance of interactive mobile applications in the future.