基于高性能计算平台的并行RTI研究与实现

发布时间：2018-03-29 05:36

本文选题：RTI　切入点：高性能计算平台　出处：《国防科学技术大学》2012年硕士论文

【摘要】：高层体系结构HLA在军事仿真领域得到了广泛应用。随着应用的不断深入，仿真应用的规模越来越大，模型复杂度越来越高，仿真实体之间的信息交互愈加频繁，对仿真平台通信和计算资源的需求不断提高。传统的基于PC机网络的仿真平台已经不能满足应用需求。高性能计算平台可以为大规模仿真应用提供强大的计算和通信能力，然而传统的HLA运行支撑平台RTI主要面向分布式网络环境，在通信结构、消息处理方式、系统数据一致性维护等方面都难以有效发挥高性能计算平台的性能优势。因此，研发基于高性能计算平台的并行RTI，对于充分发挥平台高效的通信和计算能力，提高仿真系统运行效率，促进大规模仿真应用的发展等都具有十分重要的意义。论文针对传统RTI难以有效发挥高性能计算平台性能优势的现实问题，以提高大规模HLA仿真系统运行效率为根本目标，对RTI通信结构、消息处理技术、数据一致性维护方法等关键问题进行了深入研究，主要工作和创新点如下：（1）高性能计算平台往往提供了共享内存/IB/以太网等多种不同层次的硬件通信环境，然而，目前的RTI大多只提供TCP/IP通信支持，难以有效发挥高性能计算平台多层次并行通信优势。为此，论文提出了一种支持多层次并行通信的RTI通信结构，该通信结构采用多线程并行处理技术，把不同硬件通信环境数据的发送和接收分配到不同的线程处理，，从而可有效支持不同层次通信环境之间的并行通信，提高系统的通信效率和吞吐率。（2）RTI中消息的处理可分为接收解包和分类处理两个阶段。目前的RTI往往采用单线程消息处理方法，两个阶段只能串行进行，消息处理效率较低。为此，论文提出了一种基于消息缓冲队列的多线程流水处理方法，该方法通过在两个消息处理阶段中加入消息缓冲队列，数据接收解包后放入消息缓冲队列，再由专门的消息分类处理线程根据具体消息类型进行相应处理。测试结果表明：在消息交互较为频繁的HLA仿真中，与传统的单线程处理方法相比，采用该方法消息接收处理的效率可提高8%以上。（3）数据一致性是HLA仿真正确运行的前提。目前的RTI要么采用集中式策略，维护算法简单，但中心节点往往成为系统性能瓶颈；要么采用分布式策略，虽避免了系统性能瓶颈，但效率较低且算法复杂。综合两种维护策略的优缺点，论文提出了一种分布集中相结合的数据一致性维护方法，该方法根据数据修改模式的不同，把盟员和联盟数据分为两类：盟员可立即修改的数据（自主修改数据）和需要多个盟员共同决定是否修改的数据（协商修改数据），对于前者采用分布式维护策略，由各局部RTI部件（LRC）自主维护；对于后者采用集中式维护策略，由中心RTI部件（CRC）集中维护。该方法能够在保证正确性的基础上，有效减小数据同步对系统性能的影响。（4）在上述研究成果的基础上，设计实现了基于高性能计算平台的并行RTI软件shmRTI。测试结果表明：shmRTI较完整地实现了IEEE1516规范的管理服务功能，在多核PC环境中，其节点内数据吞吐率是pRTI1516的2.8倍以上，节点内通信延迟仅为pRTI1516的40%以下；在具有IB/以太网的高性能计算平台与PC机组成的混合网络环境中，shmRTI的综合时间推进性能比pRTI1516可提高30%以上。
[Abstract]:The high level architecture of HLA has been widely applied in the field of military simulation. With the deepening of the application, the application of simulation model is larger and larger, more complex, simulation entities and the information interaction between the more frequent, the simulation platform of communication and computing resources needs to be improved. The traditional simulation platform PC network has can not meet the application requirements. Based on high performance computing platform can provide powerful computing and communication capabilities for large-scale simulation applications, however, traditional HLA running platform of RTI is mainly for the distributed network environment, the communication structure, message processing method, system data consistency maintenance are difficult to effectively play the performance advantages of high performance computing platform. Therefore, development of parallel computing based on RTI platform with high performance, to make full use of communication and computing capabilities of the platform effectively, improve the simulation system. It is of great significance to carry out efficiency and promote the development of large-scale simulation applications.
The real problem in the traditional RTI can not effectively play the high performance computing platform performance advantage, in order to improve the efficiency of large-scale HLA simulation system as the basic goal of RTI communication structure, message processing technology, the deep study of the key problem of data consistency maintenance methods, the main work and innovation are as follows:
(1) high performance computing platform often provides different levels of hardware shared memory /IB/ Ethernet communication environment, etc. However, the current RTI only provides TCP/IP communication support, difficult to effectively play high performance computing platform multi-level parallel communication. Therefore, this paper proposed a RTI communication architecture supports multi-level parallel communication the communication structure, using multi-threaded parallel processing technology, send different hardware communication environment and receiving data assigned to different threads, which can effectively support the parallel communication between different levels of communication environment, improve communication efficiency and throughput rate of the system.
(2) RTI news can be divided into receiving two stages of unpacking and classification. At present RTI often use a single thread message processing method, two stage serial only, message processing efficiency is low. Therefore, this paper proposes a multi thread pipelining method based on message queue, the method the two message processing join message queue stage, data receiving and unpacking into the message queue, then special message classification processing thread according to the specific types of messages are processed. The test results show that the simulation in HLA message exchange is more frequent, compared with the traditional single thread processing method, the efficiency of the method of message reception processing can be increased by more than 8%.
(3) the data consistency is a prerequisite for the correct operation of HLA simulation. The current RTI or centralized maintenance strategy, the algorithm is simple, but the center node often become the bottleneck of the system; or a distributed strategy, is to avoid the bottleneck of the system, but the efficiency is low and complexity. The advantages and disadvantages of two kinds of comprehensive maintenance strategy. This paper presents a data consistency maintenance method of combining centralized distribution, according to the data modification mode, the alliance and alliance data are divided into two categories: members can immediately modify data (independently modify data) and the need for multiple members to decide whether to modify the data (negotiation data). The former adopts distributed maintenance strategy, by the local RTI component (LRC) self maintenance; for the latter centralized maintenance strategy, from the center of the RTI parts (CRC) centralized maintenance. The method can ensure On the basis of correctness, the effect of data synchronization on the performance of the system is effectively reduced.
(4) on the basis of the above research, the design and implementation of parallel RTI software shmRTI. test results show that the platform based on High Performance Computing: shmRTI is a complete implementation of the IEEE1516 specification management services in the multi-core PC environment, the node throughput is 2.8 times more than the pRTI1516, the intra node communication the delay is only less than 40% of pRTI1516; in high performance computing with IB/ Ethernet hybrid network environment platform and PC machine in the shmRTI integrated time performance than pRTI1516 can be increased by more than 30%.

【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】：TP38;TP391.9

【参考文献】