一种新型的基于猜测共享变量管理并行编程模型.pdfVIP

一种新型的基于猜测共享变量管理并行编程模型* 1+ 1 1 1 郭辉 ，沈立 ，徐帆 ，王志英 1( 国防科学技术大学 计算机科学与技术系，湖南 410073) A Novel Parallel Programming Model Based on Management of Speculative Shared Data* 1+ 1 1 1 GUO Hui , SHEN Li , XU Fan , WANG Zhiying 1(Department of Computer, National University of Defense Technology, Hunan 410073, China) + Corresponding author: E-mail: samgh1230@ Abstract ：How to exploit parallelism in sequential applications and utilize ample computing resources to speed up efficiently is always a primary problem of parallel programming. The proposal of Thread-Level Speculation (TLS) has been regarded as a promising method to simplify parallel programming. But, at present, poor speedup and low energy efficiency are two major drawbacks, when TLS is applied to applications where complex data dependencies exist. In this paper, we propose a hardware supported parallel programming model, according to the analysis of shared data dependences between speculative threads. This model realizes a high-efficient management of shared data depended on a simple hardware structure. Besides, it schedules speculative threads and verifies the correctness using a runtime system. A prototype is built on SESC simulator and accelerates our benchmarks by an average of 5.69x using 8 speculative threads. Key words: Parallelism; Programming Model; Thread-Level Speculation; Loop-carried data dependence; 摘要：如何充分地挖掘串行程序的并行性并且利用现有丰富的硬件计算资源，一直以来都是困扰程序员并 行编程的首要问题。线程级前瞻执行（Thread-Level Speculation TLS）模型被认为是一种能在一定程度上简 化并行编程的新兴技术。但是当处理具有复杂数据依赖关系的并行段时，现有的TLS 存在性能加速不明显 以及性能功耗比不高等问题，这使得其在实际应用中受到很多限制。通过对前瞻线程间共享数据依赖的分 析，本文提出了一种具有底层硬件结构支持的并行编程模型。该模型使用简单的硬件结构实现了共享数据 的高效管理。另外，该模型使用运行时系统对前瞻线程进行调度管理和正确性检验。本文在模拟器SESC 上 实现了原型系统并且对测试程序进行了测试，在使用 8 个并发前瞻线程时，所有测试程序总体平均加速比 达到了5.69x。 关键词：并行化；编程模型；线程级前瞻执行；循环携带数据相关 中图法分类号: TP316 文献标识码: A 1 引言 随着多核、众核体系结构的快速发展，商用处理器已经具有了非常高的并行计算潜力。与之相矛盾的 是，并行编程工具