QIHE: Quantifying the Importance of Hardware Events with Respect to Performance of Mobile Processors

Nowadays, an increasing number of applications run on mobile smartphones, making people's life much more convenient than ever before. In particular, the number of mobile phones running Android operating systems equipped with ARM processors is growing steadily, accounting for more than 50% of the global mobile phone market share. Therefore, the performance of these mobile phones still needs to be improved. Hardware (microarchitecture) events of the mobile processors contain the fundamental causes of their performance bottlenecks. However, it is challenging to clearly understand the details of the impact of the micro-architecture on the processor due to: 1) the difficulty of obtaining values of micro-architecture events, and 2) the large number (more than 200) of micro-architecture events. This paper proposes QIHE, a hybrid methodology which encompasses not only a way of collecting micro-architecture events, but also quantifying the importance of them with respect to performance. This method first collect 126 micro-architecture events for each of 70 applications on two mobile phones. Subsequently, it need quantify the importance of the events with respect to performance by using a machine learning algorithm — SGBRT (Stochastic Gradient Boosted Regression Tree). Finally, the 13 most important microarchitecture events are identified for all the applications running on the two mobile phones. These events can be used to optimize the processor microarchitecture as well as the performance of the applications.