可扩展、周期精确、快速多核模拟器研究

发布时间：2018-03-29 19:45

本文选题：多核模拟器　切入点：松耦合　出处：《复旦大学》2012年硕士论文

【摘要】：在计算机体系结构领域,体系结构的设计和评估需要科研人员反复对计算机体系结构进行修改和性能测试。由于硬件灵活性差且生产代价高,体系结构模拟器作为一种重要的设计和评估软件,已广泛应用于体系结构设计和评估的方方面面。随着多核技术的发展和成熟,全系统多核模拟器开始发挥越来越重要的作用。但是为了保证周期精确性,目前主流的多核模拟器的功能模拟模型和时序模拟模型采用紧耦合的连接方式,交互十分复杂,从而导致可扩展性差和性能低下。本文首先系统地分析了模拟器周期精确性的影响因素,主要包括分支预测错误、共享数据的访问顺序、中断和异常的处理以及共享页表的访问顺序四个因素。在此基础上,本文提出了一种松耦合结构的功能优先的多核全系统模拟器——Transformer。为了保证模拟器的周期精确性,Transformer设计了轻量级的错误路径模块、共享内存访问模块、共享页表访问模块和中断及异常处理模块。而为了提高模拟器的可扩展性,Transformer为功能模拟模型和时序模拟模型定义了体系结构无关的通用接口。最后,Transformer松耦合的设计去除了功能模拟模型与时序模拟模型之间复杂的交互,从而可以进一步并行功能和时序模拟模型以提高模拟器速度。实验数据表明,在相同精确性的情况下,与主流紧耦合多核模拟器相比,Transformer通过功能和时序模拟模型的简单交互取得了8.4%的加速。在实现了功能与时序模拟模型并行的情况下,Transformer可以达到35.3%的加速。为了验证Transformer的可扩展性,本文在其基础上扩展了新的功能模拟模型QEMU从而支持X86模拟。这项扩展工作由一名硕士生独立完成,只需要修改180行代码,总体耗时只需两个月。
[Abstract]:In the field of computer architecture, the design and evaluation of architecture require repeated modifications and performance tests by researchers. As an important design and evaluation software, architecture simulator has been widely used in all aspects of architecture design and evaluation. However, in order to ensure the accuracy of the cycle, the functional simulation model and the timing simulation model of the mainstream multi-core simulator are tightly coupled and the interaction is very complex. This leads to poor scalability and low performance. In this paper, the factors affecting the accuracy of the simulator cycle are analyzed systematically, including branch prediction errors, access order of shared data, handling of interrupts and exceptions, and access order of shared page tables. In this paper, a loosely coupled multi-core full-system simulator, Transformer-based, is proposed. In order to ensure the cycle accuracy of the simulator, Transformer has designed a lightweight error path module, a shared memory access module. The shared page table access module and interrupt and exception handling module. In order to improve the extensibility of the simulator Transformer defines the architecture independent general interface for the functional simulation model and the timing simulation model. Finally, the transformer loosely coupled design. The complex interaction between the functional simulation model and the time series simulation model is removed. Thus, the parallel function and the time series simulation model can be further improved to improve the speed of the simulator. Experimental data show that, with the same accuracy, Compared with the mainstream tightly coupled multi-core simulator, Transformer achieved an acceleration of 8.4% through simple interaction between function and time series simulation model. When the function was parallel with the timing simulation model, Transformer could achieve 35.3% acceleration. To prove the extensibility of Transformer, This paper extends the new functional simulation model QEMU to support X86 simulation, which is completed by a master student independently, and only needs to modify 180 lines of code. The total time is only two months.
【学位授予单位】：复旦大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP303

【相似文献】