摘要
[目的]采用处理器和通信信道共享策略的异构系统正逐渐成为研究主流,但是算法部署过于复杂,而采用工作负载分配的通信计算模型又会导致通信的不平衡,导致系统均衡性存在偏差。[方法]为解决该问题提出一种内核成本测量可伸缩通用矩阵乘异构通信τ-LOP模型分析方式。首先,在混合异构平台上,采用τ-LOP模型估计并行应用数据的通信开销扩展。其次,使用两个具有相反特征的数据构建并行核的扩展模型,建立具有高通信与计算比SUMMA矩阵乘法和二维波动方程的有限差分求解器,有效降低通信复杂度。[结果]最后,通过仿真实验,验证了所提算法模型的有效性。
Heterogeneous systems using processor and communication channel sharing strategies are becoming the mainstream of research, but the algorithm deployment is too complex, and the communication computing model using workload distribution will lead to communication imbalance, resulting in system balance bias. To solve this problem, a kernel cost measurement can be proposed. The generalized universal matrix multiplying heterogeneous communication τ-LOP model analysis method. Firstly, the τ-LOP model is used to estimate the communication overhead expansion of parallel application data on hybrid heterogeneous platforms. Secondly, an extended model of parallel kernel is constructed by using two data with opposite characteristics, and a finite difference solver with high communication and computational ratio SUMMA matrix multiplication and two-dimensional wave equation is established, which can effectively reduce the communication complexity. Finally, the effectiveness of the proposed algorithm is verified by simulation experiments.
作者
赵彤
何涛
ZHAO Tong;HE Tao(School of Automation & Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《电视技术》
2019年第2期73-80,共8页
Video Engineering
