基于模型树的多核设计空间探索技术

处理器设计面临的重要挑战是如何在庞大的设计空间中高效地找到满足约束的设计结构,而预测模型方法是探索复杂设计空间的重要方法.为了提高预测模型的实用性,提出一种基于模型树的多核设计空间探索技术.首先对整个设计空间中部分设计结构进行采样模拟,然后通过模型树算法构建设计参数与处理器响应之间的预测模型,最后通过该模型预测出其他设计结构的响应以找到满足约束的最优设计.实验结果表明,与现有的基于支持向量机和人工神经网络的预测模型技术相比,针对性能预测,采用文中技术能够提高74.87%和38.87%的准确度;针对能耗预测,能够提高2.66%和16.82%的准确度.

Model Tree Based Multi-core Design Space Exploration

During the design of microprocessors,it is a great challenge to efficiently determine an optimal design configuration to meet design specifications,and predictive modeling is a promising technique for efficient design space exploration.In order to improve the practicability of existing predictive modeling techniques,we propose to employ model tree based predictive modeling for multi-core design space exploration.First,a small fraction of design configurations are sampled and simulated.Then,such data are utilized to construct a model characterizing the relationship between design parameters and processor responses by model tree algorithm.Finally,such model could be used to predict the responses of other design configurations,and the optimal design configurations can be found.Experimental results show that,compared with SVM-based and ANN-based predictive models,our approach can improve the prediction accuracy by 74.87% and 38.87%,respectively,with respect to performance prediction,and the prediction accuracy by 2.66% and 16.82%,respectively,with respect to energy prediction.