A novel evaluation method for superscalar out-of-order ARM microprocessors targeting android applications
Yang Zhang, Zhi Qi, Xiaoxi Wu, Wenjie Fu · 2017
Nowadays, most of the mobile platforms are equipped with state-of-the-art ARM Cortex-A series superscalar Out-of-Order(OoO) mobile microprocessors. However, due to the increasing varieties and complexities of mobile applications supported by full-fledged Android Operation System(OS), the architects should revise and re-evaluate the microarchitecture design of the microprocessor consistently. This paper proposes a novel performance evaluation method based on a hybrid analytical model, specifically targeting the ARM Cortex-A series high performance OoO microprocessor running the latest Android applications. To generate more microarchitecture insights from the evaluation with higher accuracy, we modify the Gem5 simulator to trace and dump useful microarchitecture dependent characteristics of the application as well as some OS related daemon programs whenever thread switch happens. Then, with the consideration of generally neglected overheads introduced by serializing instructions and structural hazards that cannot be ignored under ARM Android environment, the proposed evaluation method integrates an improved hybrid analytical performance model with an Artificial Neural Network (ANN) empirical model. Additionally, all the performance predictions are visible through a friendly GUI in our method. The experimental results show that the average prediction error of the proposed evaluation method is between 3% and 5% when running the BBench Android benchmark compared with cycle accurate Gem5 simulation. The proposed model improves the quality of performance prediction. The comparison with the ground truth verifies the superior of our proposed model over other traditional evaluation models.