多核平台上基于可声明并行性的程序优化框架

A Declarative Parallelism Based Program Parallel Optimization Framework on Multicore Platforms

针对多核体系平台上充分、有效地发掘目标程序中各种可用并行性的需求,通过引入"层次关系"、"等价关系"和"特性权重"的支持,提出了一种扩展的TStreams模型,并在此基础上实现了一个基于可声明并行性的程序并行优化框架(FAPOF).该框架支持用户对算法的并行特性进行多角度、多粒度的描述并指定适用的各类并行优化规则.基于用户描述,框架可以编译驱动的方式评估各种优化决策的组合,以半自动化的方式对目标程序进行并行优化.由此可将优化过程中程序员原本复杂、困难的并行优化的"决策"工作转化为可用并行性的"描述"工作.实验结果表明,此方法显著地降低了并行优化的难度,提高了并行优化的效率.

To satisfy the requirements for exploiting various types of available parallelism in destination programs on multicore platforms fully and efficiently,an extended TStreams model was developed by introducing the hierarchical relationship,equivalent relationship and characteristic weights as the support,so as to implement a declarative parallelism based feature aware parallel optimization framework（FAPOF） which is on the basis of the model.FAPOF supports multi-aspects and multi-grains description for the latent parallelism and appropriate optimization rules available in destination algorithms.With users＇ description,FAPOF can evaluate the different combinations of the optimization decisions in a compiler-driven method,thus the destination programs can be optimized in parallel semi-automatically.In this way,the programmers＇ complex and difficult ＂decision-making＂ jobs in parallel optimization will be transformed into an easier ＂description＂ job for available parallelism.Test results showed that the method mentioned above can reduce the difficulty in parallel optimization significantly with the optimization efficiency improved.