减少推测多线程回退的片上网络优化设计与分析-计算机系统结构专业论文.docxVIP

摘 摘 要 摘 要 推测多线程是一项旨在简化并行编程的技术。它创建多个未静态确定线程 间依赖关系的线程，然后推测地并行执行它们，并动态检测线程间数据依赖冲 突。一旦发生冲突，违反依赖的线程被回退并重启，以保证推测执行的正确性。 但是，回退会带来具有较大的性能开销，频繁的回退会严重影响线程执行的性 能。 为了减少线程回退，改进推测多线程的性能，本文提出了一种基于优先级 的片上网络，部署在使用推测多线程技术的片上多核处理器中，通过优先传送 前驱线程的消息，使推测线程的访存操作符合串行语义，从而减少依赖冲突和 回退的发生。 本文的主要研究工作和成果包括：（1）分析了推测多线程的线程回退行为， 指出前驱线程处于整个模型的关键执行路径上，其访存延迟显著影响推测多线 程的性能，因此应该在网络中为其赋予更高的优先级。（2）提出了基于优先级 的片上网络的设计，并讨论其硬件复杂度。基于优先级的片上网络的设计对系 统软件层是透明的，同时带来的硬件开销也是可以接受的。（3）选取了合适的 测试程序，对基于优先级的片上网络的性能进行量化评估和分析。在分析中我 们验证了其有效性和可扩展性，同时也得到了对推测多线程技术和基于优先级 的片上网络的进一步认识。 本文所提出的片上网络与上层应用协同设计的方案，具有硬件开销小、简 单有效和扩放性好等特点，从片上网络这个角度改进了推测多线程的性能，拓 展了其设计空间，同时对加快推测多线程的推广具有积极意义。 关键词： 片上多核处理器 片上网络 并行编程 推测多线程 回退 优先级 I ABSTRACT ABSTRACT ABSTRACT Thread Level Speculation (TLS) is a technique aims at simplifying parallel pro- gramming. It creates multiple threads without knowing they are independent, and then speculatively executes them in parallel, hoping not to violate the inter-thread data depen- dency. In case of violation happens, the offending thread will be squashed and restarted to guarantee the correctness of execution. However, squashing incurs considerable per- formance overhead, and frequent squashing significantly cripples TLS performance. To reduce the occurrence of squashing and improve TLS performance, we propose a priority-based Network-on-Chip (NoC), which is deployed in the Chip Multiproces- sor (CMP) supporting TLS. By preferentially transferring messages from predecessor threads, the priority-based NoC guides TLS threads to perform data accessing conform- ing with the sequential semantic, and therefore brings down the count of dependency violation and squashing. The focus and main contributions of this paper are as follows. First, by analyzing the TLS squashing behavior, we draw the conclusion that predecessor threads are on the critical executing path of the TLS model, and their memory access latency impose sig- nificant impact on the TLS performance. Therefore, they should be given higher priority in the NoC. Secon