| Android system is currently the most mainstream operating system of mobile terminals.In order to optimize the system design,software and hardware developers need to obtain Android application runtimebehavior and micro-architecture characteristics of hardware.There are many tools based on the ARM architecture for example,ARM Streamline.However,for non-ARM architectures,such as the domestic Loongson CPU architecture,there still lacks of corresponding performance analysis tools and methods.Therefore,it is very meaningful to develop a set of Android application oriented performance testing methods for the Loongson CPU.PMU(Performance Monitoring Unit),a logical component of modern processors that can be configured to monitor(count)various types of performance-related hardware events,makes it possible to record runtime features of microarchitectural events under circumstances of minimal overhead and minimal interference to the target programs.After investigations of the realizations of the principle of existing performance analysis tools,such as VTnue,Perf and ARM Streamline and the hardware features of Godson 2H platform,this paper develops a performance testing tool which uses a full system sampling method based on PMU for Android applications on the Loongson 2H platform.There are three main modules: data analysis,data sampling control and data acquisition.The data analysis module is in charge of analyzing sampling data.The data sampling control is used to configure event type for sampling and data storage mode.The data acquisition moduleis responsible for controlling performance counters and collecting hardware event information.The main function is to accurately count the corresponding processor hardware events per threadduring Android application runtime,and the implementation details including thread starting time,cutting time,operation time etc.Since the Loongson 2H platform has the perf tool that does not distinguish threads with the same ID in thread statistics,not support and the operation is cumbersome.At the process level,we compared the performance test tools implemented in this paper with the accuracy of the existing tool perf on the Loongson 2H platform.Among them,0xbenchmark test results show that the maximum error of the cpu-clock number is 1.11%,the maximum error of branch instruction number is 1.41%,and the number of L1 instruction cache misses is 0.99%. |
