基于支持向量机的移动Web浏览性能优化研究

Optimize Mobile Web Browsing Based on Support Vector Machine

随着网页内容和功能的丰富以及用户体验需求的提升,移动Web浏览中的计算能耗与日俱增.DVFS(Dynamic Voltage and Frequency Scaling)技术在同构多核架构中取得较好的节能效果,但在异构多核架构中,由于系统默认的调度策略没有充分利用低功耗核心,导致高性能核心的工作频率和时间往往高于实际需求,浪费大量电能.而面向异构平台的Linux HMP(Heterogeneous Multi-Processing)技术也没有充分利用异构平台的特性,无法对能效进行有效的提升.针对上述问题,该文面向移动异构平台,提出了一种基于支持向量机的CPU配置预测模型.首先选取500个热门网站主页,分析其主页面的架构(HTML)及样式(CSS)信息,进行特征选择;遍历不同CPU配置渲染网页,记录获得最优加载时间、能耗及EDP对应CPU配置;最后在线下利用支持向量机自主挖掘网页特征同最优配置的内在关系,以此构建移动异构平台的CPU资源调度预测模型.该模型通过分析网页特征,根据不同的优化目标,为渲染引擎分配合适的处理器资源.实验结果显示,同目前最先进的一种线性回归预测模型相比,该文提出的CPU资源调度模型在加载时间、能耗和EDP上的性能得到显著提升.

With the growing demand for high-performance mobile browsers and the increasing complexity of websites,the Web browsing consumes more and more energy to satisfy such requirements.The system default governor DVFS(Dynamic voltage and frequency scaling)technology performs well in homogeneous multi-core platform while it cannot make full use of low-power cores in heterogeneous architecture.For example,in order to provide a good user experience,the DVFS prefers to schedule the task on the high-performance cores(big core)with high frequency and takes extra time on activity state which causes much energy waste.Although the Linux HMP(Heterogeneous Multi-processing)scheduler is designed for the heterogeneous architecture,it cannot capture the characteristics of different workloads.The HMP performs better than DVFS on energy efficiency,but it still cannot take advantage of heterogeneous architecture therefore,leaving a huge space to optimize.In response to these problems,this paper proposes a predictive model for CPU configurations based on Support Vector Machine to optimize Web browsing on heterogeneous mobile platform.Firstly,it characterizes the web-page workload by conducting the feature selection of the structures(HTML)and styles(CSS)of top 500 popular websites’land-page.Then it schedules the rendering process on the big or little core with different clock frequencies.After that it records the best performing configuration for three different optimization goals(load-time,energy consumption and energy delay product-EDP).Finally,it builds up the predictive models off line by using the SVM(Support Vector Machine)to find the correlations between the selected features and the best configurations of different metrics.Meanwhile,it implements the predictive model as a web browser extension to schedule the rendering engine which is designed to run on the predicted core at specified frequency by analyzing features of web-page.The results show that the SVM-based models make significant improvement on load time,energy consumption and EDP as compared to the state-of-the-art linear regression predictive model.