避免无关依赖的众核线程划分机制

How to convert the massive transistor resources in manycore architecture into effective computing power, furthermore extend the parallelism of the mainstream software transparently, is a hot issue in current study. For now the block is the smallest unit in manycore thread partition mechanism, it would bring in unnecessary dependences in parallel execution. This project proposes a fine-grained thread partition method, which can minimize the unnecessary dependences among speculative threads. It divides the hard-to-parallel program into lean and separate speculative threads so that can solve the traditional parallel programming model and environment problems. The topics of the research include: fine-grained thread partition method based on dependence separation in manycore, hardware support of lightweight & implicit thread synchronization mechanism, dynamic resource allocation mechanism in manycore. The target of this project is to propose a reasonable thread partition mechnism in manycore, which can balance the easy-to-parallel and efficient-to-parallel. It also would improve the resource utilization, synchronization overhead and load imbalance problems in manycore architecture. It has important meanings for the high performance general-purpose chip architecture research in future 2013 - 2020.

如何将众核结构提供的丰富晶体管资源转化为有效的计算能力，实现主流应用软件可并行性的透明扩展，是当前研究中的热点问题。针对目前众核线程划分机制均以基本块为最小单位而引入无关依赖的现状，本项目提出一种基于依赖分离的细粒度线程划分方法，通过最小化线程间的无关数据依赖，将计算和访存依赖不规则、传统上难以手工和自动并行化的串行程序划分成精简而独立的线程并行执行，解决传统并行编程模型和环境存在的问题。主要研究内容包括：众核结构上基于依赖分离的细粒度线程划分方法、轻量级线程间隐式同步机制的硬件支持、符合线程并行执行特征的动态资源分配机制。本项目的预期目标是提出一种符合众核体系结构特征的线程划分方法,平衡并行编程的易编程性和并行程序的执行效率，改善众核结构研究当前面临的片上资源利用率不高、同步开销过大和负载不平衡等问题。对于探索2013-2020前后高效能通用微处理芯片体系结构的发展道路具有重要意义。

利用众核加速串行程序的最大难点在于如何将计算和访存依赖不规则的串行程序合理并行化，解决传统并行编程模型和环境存在的问题。本项目从线程划分、硬件支持和动态资源分配机制开展研究。该项目提出了一种众核结构上避免无关依赖的细粒度线程划分方法，进行了将众核线程划分进一步深入到基本块内部从而避免线程间无关依赖的探索；设计了一种无锁的共享存储编程模型，实现了一种轻量级线程间隐式同步机制的硬件支持，平衡了并行编程的易编程性和并行程序的执行效率；研究了符合线程并行执行特征的动态资源分配在线剖析指导机制，并结合云计算开展了根据资源特征进行任务优化分配的相关研究。实验结果表明该方法对于计算和访存依赖不规则的应用取得了较好的加速效果。本项目共资助出版学术专著1部（科学出版社），申请国家发明专利1项，在计算机学会CCF推荐重要国际学术会议和计算机学报等国内外核心期刊发表科技论文共11篇。