论文部分内容阅读
随着微处理器的系统结构日趋复杂以及多核技术的发展,其设计空间大小呈指数式增长,且软件模拟技术极为费时,导致设计空间探索成为微处理器设计的一个重要挑战。近年来,机器学习技术被广泛用来构建设计空间的预测模型,从而以较少次数的模拟有效地进行设计空间探索。然而,现有的方法大多是构建有监督学习的预测模型,要保证预测模型的精度,通常都需要大量有标记的训练样本,从而导致较大的模拟代价。为了减少模拟次数并提高模型的预测精度,本文提出了一种结合集成学习和半监督学习技术的高效设计空间探索方法。具体而言,该方法包括两个阶段:使用均匀随机采样方法从处理器设计空间中选择一小组具有代表性的设计点,通过模拟获得性能响应,从而组成训练样本集;提出结合半监督学习技术的Ada Boost预测模型SSLBoost预测未模拟的样本配置的响应,从而搜索最优的处理器设计配置。实验结果表明,与现有的基于人工神经网络和支持向量机的有监督预测模型相比,本文提出的SSLBoost能够使用更少的模拟样本构建出不差于现有方法性能的预测模型;而当模拟样本数量相同时,SSLBoost的预测精度更高。
As the system structure of microprocessors becomes more and more complicated and the development of multicore technology, the size of the design space increases exponentially, and the software simulation technology is extremely time-consuming, leading to the design space exploration becoming an important challenge in microprocessor design. In recent years, machine learning techniques have been widely used to build predictive models of design space to efficiently explore design space with less number of simulations. However, most of the existing methods construct a predictive model of supervised learning. To ensure the accuracy of the predictive model, a large number of labeled training samples are usually required, which leads to a large simulation cost. In order to reduce the number of simulations and improve the prediction accuracy of the model, this paper proposes an efficient design space exploration method that combines integrated learning with semi-supervised learning. Specifically, the method includes two stages: using a uniform random sampling method to select a small group of representative design points from the processor design space and obtaining performance responses through simulation, so as to form a training sample set; and combining the semi-supervised learning technology Ada Boost Prediction Model SSLBoost predicts the response of unmixed sample configurations to search for optimal processor design configurations. The experimental results show that compared with the existing supervised prediction model based on artificial neural network and support vector machine, the proposed SSLBoost can use fewer simulated samples to construct the prediction model which is not worse than the performance of the existing method. When With the same number of simulated samples, the prediction accuracy of SSLBoost is higher.