基于CPU-GPU异构环境的运算代价评估模型

发布时间：2018-06-09 05:35

  本文选题：运算代价模型 + 性能分析　； 参考：《计算机工程》2017年09期

【摘要】：传统性能分析模型仅针对单个处理器,未考虑异构系统中处理器之间数据的传输开销,不能有效地评估异构系统的性能。为此,提出一种运算代价评估模型。通过对计算平台硬件参数和工作负载特征属性的分层建模,结合LogGP模型和Roofline模型,估算不同执行方式的运算时间成本。依照建模的粒度粗细,通过多层建模计算消耗的能量,采用LogGP模型实现CPU与GPU之间的数据传输,并运用EPCC测试集对运算代价模型进行验证。实验结果表明,该模型对处理不同大小数据集时的性能评估具有较高的准确性。
[Abstract]:The traditional performance analysis model is only for a single processor and does not take into account the data transmission overhead between processors in heterogeneous systems, so it can not effectively evaluate the performance of heterogeneous systems. Therefore, a computational cost evaluation model is proposed. Through the hierarchical modeling of hardware parameters and workload characteristic attributes of the computing platform, combined with LogGP model and Roofline model, the computational time cost of different execution modes is estimated. According to the granularity of modeling, the energy consumed is calculated by multi-layer modeling, and the data transmission between CPU and GPU is realized by LogGP model, and the EPCC test set is used to verify the cost model. The experimental results show that the model is accurate for the performance evaluation of data sets of different sizes.
【作者单位】： 中国矿业大学计算机科学与技术学院;
【基金】：国家自然科学基金(U1510115)
【分类号】：TP332
，

本文编号：1999181