RIA: Tolerating Synchronization Delays Using Multithreadings

Saavedra多线程，沿着与软件预取、无锁定高速缓存和弱一致性模型一起，已经被提出作为一种有效的硬件机制，用于减轻预期存在于大规模并行机中的大延迟的有害影响。 本研究的重点是研究在何种程度上虚拟多线程可以用来隐藏同步事件所引起的长延迟。 虚拟多线程给人一种错觉，即处理器正在使用少量资源执行大量线程。 初步结果表明，有足够的硬件支持，虚拟线程的组合执行可以有效地用来隐藏非常大的同步延迟。 本研究的目标是：1）探索支持虚拟线程的新硬件机制并评估其成本-性能权衡：2）根据最相关的机器和程序参数开发分析公式来建模虚拟多线程的有效性; 3）研究细粒度并行程序在同步方面的动态行为，并使用这些信息来构建程序执行模型; 4）开发了一个详细的基于虚拟多线程并行机的仿真器，并进行了大量的实验，以验证分析模型和硬件机制。 ***

9308981 Saavedra Multithreading, along with software prefetching, lock-up free caches, and weak consistency models, has been proposed as an effective hardware mechanism for mitigating the pernicious effects of the large latencies expected to be present in massive parallel machines. This research focuses in studying the extent to which virtual multithreading can be used to hide the long delays incurred by synchronization events. Virtual multithreading gives the illusion that the processor is executing a large number of threads using only a small amount of resources. Preliminary results indicate that, with sufficient hardware support, the combined execution of the virtual threads can be used effectively to hide extremely large synchronization delays. The goals of this research are: 1) explore new hardware mechanisms for the support of virtual threads and evaluate their cost-performance tradeoffs; 2) develop analytical formulas to model the effectiveness of virtual multithreading in terms of the most relevant machine and program parameters; 3) investigate the dynamic behavior of fine-grain parallel programs with respect to synchronization and use this information to construct models of program execution; and 4) develop a detailed simulator of a virtual multithreaded-based parallel machine and run an extensive set of experiments in order to validate the analytical models and hardware mechanisms. ***