Extraction For Characteristic Fragments Of ARM Android Multimedia Applications

With the advent of the era of mobile Internet, people prefer to browse multimedia information on the Internet by smart mobile terminal. Multimedia applications have become a key factor in users experience. For every phase has different workload characteristics when Android multimedia applications are running, in this thesis, we extract the microarchitecture-independent representative characteristic fragments, which can not only significantly reduce the existing Android multimedia applications, but also effectively represent the workload characteristics of the application.This thesis analyzes the microarchitecture-independent workload characteristics of ARM Android multimedia applications, including instructions mix, instruction-level parallelism, register traffic, temporal locality of memory access, spatial locality of instruction pre-fetch, temporal locality of instruction pre-fetch and branch predictability. For a lot of thread switches when Android multimedia applications are running, different workload characteristics of different threads, and different workload characteristics of different phases in the same thread, a benchmark reduction standard based on the thread switch is used, which can extract representative characteristic fragments of hot threads when Android multimedia applications are running. Due to the adoption of the benchmark reduction standard based on the thread switch, the characteristic fragments have non-uniform sizes. Thus, we extract representative characteristic fragments based on self-organizing map clustering algorithm, which reduces the effect of non-uniform characteristic fragment size on extracting representative characteristic fragments.In this thesis, Gem5 simulator is used to simulate the ARM Android platform, which is widely used in educational circles, add the function of tracing microarchitecture-independent workload characteristics based on the Gem5 official source code, and VLC multimedia player is used to experiment. The results show that comparing with K-means clustering algorithm, the relative error of the average CPI decreases by 25.98%.