面向软硬件协同设计的性能优化框架

A performance optimization framework for hardware/software co-design

面对高性能计算机系统随着性能的提高其复杂性成倍增大的挑战,研究了复杂科学计算应用的优化,提出了一种面向软硬件特性设计的性能优化框架CPTF。该框架根据应用在运行时的剖析结果,结合应用的软件特性和平台的硬件特性,全局性地分析系统性能瓶颈及种类,并给出源码级的优化建议,并针对优化循环一类常见的问题,提出一种改进的循环合并算法。使用CPTF优化了一个物质点法粒子模拟应用,取得了近20%的性能提升。

The optimization of the application of complex scientific computation was studied to face the challenge that the complexity of high performance computers grows quickly with their performance increasing, and the CPTF, a hard- ware/software co-design performance tuning framework, was proposed. The CPTF can give an overall analysis of a computer system＇ s performance bottlenecks and their types according to the profile of the application at run-time combined with the application＇ s software characters and the platform＇ s hardware characters, and finally give opti- mization suggestions on source-level. And also it can propose an advanced loop fusion algorithm to solve the com- mon problems in loop optimizing. The CPTF was used to optimize the particles simulation application of MPM and achieved about 20% of performance improvement.