A Static Analytical Performance Model for GPU Kernel

摘要：

　　Graphics processing units(GPUs)have shown increased popularity play an important role as a kind of coprocessor in heterogeneous co-processing environment.Tens of thousands threads collaborative work in parallel to solve heavily dataparallel problems efficiently in GPUs architecture.The achieved performance,therefore,depends on the capability of multiple threads in parallel collaboration when processing algorithm,the effectiveness of latency of latency hiding,and the utilization of multiprocessors.In this paper,we propose a static analytical kernel performance(SAKP)model for GPU kernel.The model considers three important factors that affecting the performance of GPU kernel,which including the cost of computing instruction,memory accessing and synchronization.In the proposed model a set of kernel and device features for the target GPU is generated.In conjunction of kernel and device features we determine the performance limiting factor and we generate an estimation of kernels execution time.We performed experiments on matrix multiplication(MM)and histogram generation(HG)in NVIDIA GTX680 GPU card and showed an absolute error in predictions less than 6.8%.Meanwhile,we validated our proposed model is more accuracy and simple by comparing with other current kernel models.

关键词： GPUs co-processing collaborative work static analytical kernel performance (SAKP) model memory accessing synchronization GPU kernel

作者: Jinjing Li Qingkui Chen Baoping Liu Bocheng Liu

作者单位: University of Shanghai for science and Technology

会议类型: 国内会议

会议名称: 第10届全国计算机支持的协同工作学术会议暨中国计算机学会协同计算专委年度工作会议

会议地点: 太原

会议语种:英文

页码: 445-452

在线出版日期: 2015-08-28（万方平台首次上网日期，不代表论文的发表时间）