第**卷 第*期
201*年*月
天津大学学报（自然科学与工程技术版）
Journal of Tianjin University（Science and Technology）
Vol. ** No. *
*.
201*
收稿日期：201*-**-**；修回日期：201*-**-**.
作者简介：刘开华（1956— ），男，博士，教授，liukaihua@tju.edu.cn．
通讯作者：李 卓，zli@tju.edu.cn．
基金项目：国家自然科学基金资助项目（61602346）.
Supported by the National Natural Science Foundation of China（No. 61602346）.
基于神经网络的命名数据网学习型 FIB 研究
刘开华，闫 柳，李 卓，宫霄霖，彭 鹏，王彬志
(天津大学微电子学院，天津 300072)
摘 要：针对命名数据网转发信息库（Forwarding Information Base，FIB）快速检索差异化名称数据、高效存储转
发信息和有效支持最长名称前缀匹配机制的需求和挑战，首次提出了基于神经网络的命名数据网学习型FIB整体方
案，称L-FIB。首先，介绍了L-FIB的索引结构Learning Tree，通过使用塔式两级神经网络模型学习索引内容在存储
器中的分布情况，实现更均匀的数据映射，降低映射冲突，提高存储效率。其次，研究了L-FIB的存储结构和名称
数据检索算法，片内高速存储器部署多个与不同名称前缀组件数相对应的索引结构Learning Tree，片外低速存储器
部署多个与索引结构Learning Tree对应的FIB存储池，并通过相应的名称数据检索算法实现对兴趣包的转发信息检索
和转发信息更新操作，有效支持了命名数据网的最长名称前缀匹配机制，提高了名称数据检索速度。实验结果表
明，L-FIB在误判概率、存储消耗和吞吐量方面的综合性能明显优于其他对比方案。在误判概率低于1%的条件下，
L-FIB的索引结构存储消耗仅为58.258MB，能够部署于高速存储器SRAM上。L-FIB的实际吞吐量约为1,164万数据包
/秒，可以满足当前命名数据网对数据包快速处理的要求。
关键词：命名数据网；转发信息库；神经网络；名称数据检索
中图分类号：TP393.0 文献标志码：A 文章编号：0493-2137(201*)00-0000-00
A Learning Forwarding Information Base for Named Data Networking
with Neural Networks
Liu Kaihua，Yan L iu ，Li Zhuo，Gong Xiaolin，Peng Peng，Wang Binzhi
(School of Microelectronics，Tianjin University，Tianjin 300072，China)
Abstract：Designing an effective Forwarding Information Base (FIB) for Named Data Networking (NDN) is a major
challenge within the overall NDN research area, since FIB has to perform fast lookup for complex names, provide high
capacity and accurately support the mechanism of Longest Name Prefix Matching (LNPM). Therefore, a learning FIB
based on neural networks, called L-FIB, was proposed. First, the index of L-FIB, named Learning Tree, uses a two-
level neural network model to learn the distribution characteristic of data indexed in static memory, which achieves
more uniform mapping, reduces the false positive probability and improves the memory utilization. Second, the stor-
age structure and name lookup algorithms of L-FIB were put forward. T
he on-chip memory using SRAMs deploys
multiple Learning Trees corresponding to the name prefixes with different numbers of components, while the
off-chip memory using DRAMs deploys multiple FIB stores corresponding to the Learning Trees. The name
lookup algorithms were also described to implement the retrieval of forwarding information for the Interes
t
packets and the update of forwarding information, which
can well support the LNPM mechanism and realize fast
name lookup. Experimental results show that the overall performance of L-FIB is superior to the comparison schemes
in terms of the false positive probability, the memory consumption and the throughput. The index of L-FIB can ex-
tremely reduce the memory consumption to 58.258MB with the probability of false positive under 1%, which means i
t
剩余7页未读，继续阅读
资源评论
weixin_38609693
- 粉丝: 8
- 资源: 961
