The Implementation And Optimization Of GPU Architecture Simulator

Simulator is an important tool for architecture research and design,and widely used in academia and industry.There are many advantages of system-wide simulator,like fast developed,low cost,easy to modify and other characteristics.Simulator plays an important role in system design.Due to the development of computer architecture simulator is usually very difficult,running some benchmark took lots of time and plenty of deviation,limiting the effectiveness of the system architecture software simulation technology in the design of computer systems.Many researchers have proposed a variety of methods and techniques to solve these problems,but so far,these problems have not been fundamentally resolved.In recent years,GPU(Graphic Processing Unit)has been rapid development,GPU is well suited for high efficiency and low cost,high-performance parallel numerical computation.The GPGPU(General-purpose computing on graphics processing units)is a task of processing graphics professional graphics processor to engage in common computing tasks originally handled by the central processor,these general-purpose computing is often nothing to do with graphics.Modern graphics processors powerful parallel processing capabilities and the programmable pipeline,so processor with a stream of non-graphical data as possible.Especially in the face of a single-instruction scheduling and transmission of data is much larger than the amount of computation required stream multiple data stream(SIMD)and data processing,general graphics processor performance far beyond the traditional central processor applications.On the other hand,CPU due to restrictions by Moore’s Law,CPU use to improve the process and speed of approach encountered barriers workmanship,temporarily unable sudden;and from 1993,the performance of the GPU to 2.8 times the annual rate of growth.So for the GPU in the current academic research it is very popular,but NVIDIA does not disclose its source code and the mainstream GPU simulator upgrading slow,leading to more difficult to research in this area.The GPU simulator in academic is GPGPU-SIM(General Purpose GPU-Sim)by UBC(University of British Columbia),which was developed in 2010,because of its good performance,and the experimental data and the results highly consistent physical GPU(after correction data consistency up to 98%)is widely used in academia.But there are two disadvantages,(1):Does not support preemptive multitasking operation,(2):Code obsolete and does not apply to flexible application scenariosThis thesis is a work made up of re-build a new GPU simulator that is Z-GPU-Simulator from the bottom,the main work includes the following four parts:1:Simple debugger implementations,the debugger can print register status,scan memory,implement the evaluation of complex expressions.2:The memory module implementations,this step for the memory chips to simulate and to achieve access to the memory read and write operations,while giving the GPU to return.3:Implementing an on-chip network module,built on a piece of the entire network simulator to simulate the process information simultaneously on the network interaction.4:Analyze and use the Cache module to achieve a Cached module and discusses a new memory access policy.