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。
Traditional performance analysis models are designed only for a single processor. They do not consider the cost of data transfer between different processors and cannot be applied to performance analysis of heterogeneous systems effectively. A computation cost evaluation model based on CPU-GPU heterogeneous environment is presented. It estimates the computation time cost of the two execution modle by making models of hardware parameters and workload characteristics and combining LogGP and Roofline models. In addition,it calculates consumption energy by making a model with some levels. It applies LogGP model to transfer data between CPU and GPU. Besides,and uses EPCC benchmark to verify the accuracy. Experimental results show that the improved model can provide a higher accuracy on the performance evaluation of heterogeneous system when it deals with different kinds of data sets.
出处
《计算机工程》
CAS
CSCD
北大核心
2017年第9期12-16,共5页
Computer Engineering
基金
国家自然科学基金(U1510115)