基于QoS保证的2D-mesh片上网络延时评价与性能优化研究

发布时间：2018-04-04 14:14

本文选题：片上网络　切入点：服务质量　出处：《吉林大学》2015年博士论文

【摘要】：随着半导体工艺进入纳米时代，单芯片上集成数十亿支晶体管已经成为现实，从2015年开始的两年内，集成电路的工艺尺寸有望缩小到10nm。集成电路制作工艺的飞速发展推动了“片上系统”(System-on-Chip，SoC)从多核到众核(100-1000)、从总线到网络、从集中式存储到分布式存储、从二维到三维方向上的发展，片上多核间的复杂通信使得基于总线的传统片上通信结构成为主要的性能瓶颈。为了适应通信复杂度的需求，片上网络(Network-on-Chip，NoC)已经成为当前片上多核间标准的通信架构。面向不同的业务需求，NoC应提供不同的性能保证，比如：像电子邮件、文件传输等弹性业务需要系统提供足够的通信带宽，采取尽力服务的方式，提高网络资源的利用率，改善网络的平均性能；而对那些诸如多媒体业务等对时间敏感度高的刚性需求，则需要NoC提供确定性的端到端的延时保证。本文在保证网络服务质量(Quality of Service，QoS)的基础上，建立网络模型，评价网络通信延时，并采取相关技术，优化网络性能。主要针对以下四方面展开研究：第一，在以往的片上网络性能评价中，经常假设存储器在网络节点中均匀分布，且读取响应具有固定延时，没有考虑网络节点访问片外存储器的情况，导致访存执行过程过于简单，得到的评价结果过于乐观，不能准确反映网络中存贮器访问的真实情况。文中选用Open Core Protocol-International Partnership(OCP-IP)工作组的DRAM模型，采用C++和SystemC语言实现模型配置与接口，修改片上网络仿真平台Noxim，使其能够访问片外存储器，并运行SPLASH-2测试标准，获得真实有价值的片上网络访存延时数据。第二，在多核甚至众核系统中，处于不同路由节点的存储器读写延时，在整个网络延时中占有相当比重。本文将访存延时作为一项重要参数，对片上网络访存延时均衡性展开研究，建立片上网络往返访存延时预测模型，提出基于总延时预测的访存报文仲裁技术。首先，依据访存报文后续路径的拥塞信息预测访存报文未来等待延时，并计算出总延时。其次，基于预测的总延时对竞争同一链路的访存报文进行仲裁。本文在2D-mesh片上网络路由器中，对该技术进行了设计和实现，并对访存延时做了优化，有效地减小了片上访存的延时，获得了更好的访存延时均衡性，提升了系统性能。第三，减小系统的平均延时，对某些弹性业务的性能有一定程度的改善，但是对于一些实时性要求比较高的刚性业务，比如多媒体应用技术，如果网络服务不能保证视频流的最大延时，会导致某一点上视频播放效果急剧恶化，影响视频输出的整体质量，在这种情况下，应确保视频流端到端的有效传输。本文以网络演算为依据，首先分析片上网络GPS节点输入数据流的特征，建立紧致的到达曲线，然后根据网络的繁忙周期，分析每条数据流在不同时间段的积压，最后得到分段线性的GPS节点的准确服务曲线。与传统的GPS节点近似服务曲线相比，文中推导的服务曲线更紧致，服务性能更强，确保了数据流端到端的QoS。第四，建立片上网络流量控制演算模型，评价网络性能。通过在路由节点添加流量控制器，实现传输路径上相邻的下游节点到当前节点的反馈机制，根据反馈信息，控制注入到网络中的数据量，，从而降低网络拥塞，减小网络竞争。借助网络演算中的到达曲线和节点路由器的服务曲线，分析网络中的延时和积压，推导片上最优缓冲区大小。综上所述，本文面向不同业务，在确保网络QoS的前提下，构建片上网络通信模型，分析其延时函数，评价并优化系统的性能。
[Abstract]:In order to meet the demands of communication complexity , network - on - chip ( NoC ) has become the main performance bottleneck . In order to meet the needs of communication complexity , the network - on - chip ( NoC ) has become the main performance bottleneck . In order to meet the needs of communication complexity , the NoC should provide different performance guarantees . For example , the system can improve the utilization rate of network resources and improve the average performance of the network . In order to meet the requirements of communication complexity , the network - on - chip ( NoC ) has become the main performance bottleneck .
while there is a need for noc to provide deterministic end - to - end delay guarantees for those rigid requirements such as multimedia services for time sensitivity .

On the basis of ensuring the quality of service ( QoS ) , this paper establishes a network model , evaluates network communication delay , and takes the related technology to optimize the network performance .

First , in the previous network performance evaluation , it is often assumed that the memory is uniformly distributed in the network node , and the read response has a fixed delay , which leads to an excessively simple visit and execution process , which can not accurately reflect the real situation of memory access in the network . In this paper , an Open Core Protocol - International Partnership ( ocp - IP ) working group is adopted to realize the model configuration and interface , and the network simulation platform Noxim on the chip is modified , so that the network simulation platform Noxim can be accessed , and the SPLASH - 2 test standard is run to obtain the real valuable piece - on - chip network access time delay data .

Second , in the multi - core and even public - core system , the memory read - write delay of different routing nodes occupies a significant proportion in the whole network delay . In this paper , the memory delay is used as an important parameter to study the time delay of network visit and memory on the chip , and the total delay is calculated . Secondly , based on the predicted total delay , the time delay of the visiting message is predicted and the total delay is calculated . Secondly , based on the total delay of the forecast , the time delay of the access memory is optimized , the delay of the on - chip visit is optimized , the better memory delay balance is obtained , and the performance of the system is improved .

Third , reduce the average delay of the system , improve the performance of some elastic services to some extent , but for some real - time requirements , such as multimedia application technology , if the network service does not guarantee the maximum delay of the video stream , it can cause the video playing effect to deteriorate sharply at some point . In this case , the accurate service curve of each data stream is analyzed . The service curve derived in this paper is more compact and the service performance is stronger than the traditional GPS node approximate service curve .

fourthly , establishing an on - chip network flow control calculus model to evaluate the network performance , and realizing the feedback mechanism of the adjacent downstream nodes on the transmission path to the current node through adding a flow controller at the routing node , controlling the data quantity injected into the network according to the feedback information , reducing the network congestion and reducing the network competition .

In conclusion , in this paper , the network communication model is constructed on the premise of ensuring the QoS of the network , and the time delay function is analyzed , and the performance of the system is evaluated and optimized .

【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TN47

【参考文献】