基于单程序多数据流并行软件的性能预测法

Performance prediction method for parallel software based on single program multiple data

为了提高预测并行软件性能的准确性和并行软件的开发效率,提出了一种基于单程序多数据流(SPMD)并行应用软件模块化技术的性能分析预测和并行软件辅助开发方法。通过量化计算开销、通信开销、通信与计算的耦合系数,阐述了在并行计算机系统中利用该方法开发并行计算软件和预测并行计算软件性能的过程。并就影响并行软件性能预测和开发效率的模块设计、模块性能数据建模、模块组合和计算与通信重叠等技术进行了研究。实验表明,该方法提高了预测并行计算软件性能的准确性,也提高了并行计算软件的性能和开发效率。

A performance prediction and development method for single program multiple data (SPMD) parallel software was developed based on parallel software modularization technology to enhance performance analyses of parallel software algorithms. The method addresses several problems, such as module designs, modeling, module combinations, and communication overlapps. Test results show that this method can improve the development efficiency of parallel software and enhance the parallel software performance.