基于SMP的离散事件执行驱动性能模拟器设计与实现

发布时间：2018-01-05 04:02

本文关键词：基于SMP的离散事件执行驱动性能模拟器设计与实现　出处：《国防科学技术大学》2012年硕士论文　论文类型：学位论文

【摘要】：近十几年来，生物分子模拟，航空宇宙计算，飓风预测等科学、工程和社会领域的超高计算能力的大型应用的进展，推动着高性能计算系统的持续发展。为了降低成本，系统开发和设计人员需要对系统性能进行全面深入的性能分析，评测应用在他们计划研制的系统上的性能能否达到预期，以指导软硬件选型和系统平衡设计，为系统的设计、实现以及优化提供参考。由于相对建模分析等其它方法，在预测精度、时间开销以及灵活性等方面的优势，模拟方法已成为当前并行系统性能分析与预测的主流技术。本文针对性能模拟技术展开研究，主要研究内容包括离散事件执行驱动的性能模拟器设计、实现以及实验验证。本文首先分析了基于对称多处理机的离散事件模拟器设计中面临的关键问题，然后提出对应的解决办法。基于SMP统一编址的存储结构，SMP-SIM模拟器事件处理的核心思想包括两个方面，一是以CPU为粒度对事件进行全局管理，在共享段存放全局数据结构，统一管理所有绑定在该CPU上运行的进程所产生的事件。其基本思想是每次选择产生时间最小的事件进行处理。二是采用虚实结合的方法处理被选择的事件：选中某个事件进行处理时，如果该事件是通信开始事件，则采用虚处理，通信开销通过通信模型计算；如果事件是串行计算开始事件，则采用实处理，串行计算的时间通过直接运行获得。然后基于对关键问题的分析，提出了基于SMP的离散事件模拟器SMP-SIM的设计方案，并对其中关键功能模块的原理进行了详细分析。SMP-SIM模拟器的功能是通过在MPI库层次结构中引入模拟API层来实现的，模拟API层包含3个功能模块：通信原语分解模块，通信模型模块和事件管理模块。在通信原语分解模块中，所有的通信原语通过核心点对点通信原语重新构造。当目标程序在宿主机上执行到某个通信原语时，该通信原语会被通信原语分解模块分解为核心原语，从而调用核心点对点通信原语。核心点对点通信原语的调用和返回产生相应的事件。事件管理模块统一调度事件，，对其进行处理，并更新对应进程的模拟时间。在处理通信类事件时，事件管理模块将会和通信模型模块进行交互，负责对通信类事件的相关时间进行计算。最后，基于MPICH2通信库实现了SMP-SIM原型系统，并从模拟精度、执行效率和可扩展性三个方面对模拟器进行了全面的测试与分析。实验结果表明该模拟器具有较高的精度和良好的可扩展性，初步验证了设计方案的正确性和有效性。
[Abstract]:In the last decade, biomolecular simulation, aerospace computing, hurricane prediction and other scientific, engineering and social fields of ultra-high computing power of large-scale applications. In order to reduce the cost, system developers and designers need to conduct a comprehensive and in-depth performance analysis of the system performance. To evaluate the performance of the system they plan to develop can meet the expectations to guide the hardware and software selection and system balance design for the design of the system. Compared with other methods, such as modeling and analysis, it has advantages in prediction accuracy, time cost and flexibility. Simulation method has become the mainstream technology of parallel system performance analysis and prediction. In this paper, the performance simulation technology is studied, including discrete event execution driven performance simulator design, implementation and experimental verification. In this paper, the key problems in the design of discrete event simulator based on multiprocessor are analyzed, and the corresponding solutions are proposed. The storage structure is based on SMP unified addressing. The core idea of event handling in SMP-SIM simulator includes two aspects. One is to manage the event globally with the granularity of CPU and store the global data structure in the shared segment. Unifies the management of events generated by all processes that are bound to run on the CPU. The basic idea is to select the event that has the minimum generation time to handle the event at a time. The second is to process the selected event using a combination of virtual and real methods:. When an event is selected for processing. If the event is a communication start event, it uses virtual processing, and the communication overhead is calculated by the communication model. If the event is a serial computing start event, it uses real processing, and the time of serial calculation is obtained by running directly. Then, based on the analysis of the key problems, the design scheme of discrete event simulator SMP-SIM based on SMP is proposed. And the principle of the key function modules is analyzed in detail. The function of SMP-SIM simulator is realized by introducing the simulated API layer into the hierarchy of MPI library. The simulation API layer consists of three functional modules: communication primitive decomposition module, communication model module and event management module. All communication primitives are reconstructed by core point-to-point communication primitives. When the target program executes a communication primitive on the host, the communication primitive is decomposed into the core primitive by the communication primitive decomposition module. Thus the core point-to-point communication primitive is called. The call and return of the core point-to-point communication primitive produce the corresponding events. The event management module schedules the events uniformly and processes them. When dealing with communication class events, the event management module will interact with the communication model module and be responsible for calculating the relative time of communication class events. Finally, the prototype system of SMP-SIM is implemented based on MPICH2 communication library, and the simulation precision is obtained. The performance efficiency and extensibility of the simulator are tested and analyzed. The experimental results show that the simulator has high accuracy and good scalability. The correctness and validity of the design scheme are preliminarily verified.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.9;TP302.7

【参考文献】