基于GPU的并行计算性能分析模型

Parallel Computation Performance Analysis Model Based on GPU

针对GPU并行计算领域缺少精确的性能分析模型和有针对性的性能优化方法,提出一种基于GPU的并行计算性能定量分析模型,其通过对指令流水线、共享存储器访存、全局存储器访存的性能建模,来定量分析并行程序,帮助程序员找到程序运行瓶颈,进行有效的性能优化。实验部分通过3个具有代表性的实际应用(稠密矩阵乘法、三对角线性方程组求解、稀疏矩阵矢量乘法)的性能分析证明了该模型的实用性,并有效地实现了算法的优化。

In order to solve the problem of lacking accurate performance analysis model in parallel computation field based on GPU,we proposed a quantitative performance model which can simulate the performance of three major com- ponents of GPU including instruction pipeline, shared memory access time, and global memory access time. It is designed to build a performance model that helps programmer find the performance bottlenecks and improve the system＇s per-formance efficiently. To demonstrate the usefulness of the model and to optimize the algorithms performance, we ana- lyzed three representative real-world programs： dense matrix multiplication, tridiagonal systems solver, and sparse ma- trix vector multiplication.