基于优化决策树的时延敏感流智能感知调度

doi:10.11959/j.issn.1000-0801.2023095

摘要/Abstract

摘要：

目前流量调度策略无法做到智能按需化，尤其对于网络突发故障造成的拥塞以及高价值业务的护航场景，无法按需保障时延敏感的业务体验。通过分析研究不同网络业务流量时延敏感性属性需求，探索挖掘不同网络业务流量的行为特征与其时延敏感性需求之间的内在关联关系。然后利用AI技术对这种内在的关联关系进行学习，构建其映射关系，实现了时延敏感流智能感知调度。同时，考虑AI模型的可解释性及可部署性实际问题，采用强化学习剪枝优化可解释性决策树模型，提高模型的鲁棒性同时使模型更轻量化，易于设备部署实现。通过真实网络流量实验，强化学习优化后的决策树模型在单次推理情形下感知正确率提高1.75%，推理速度提升约 30%；同时，实验也证明了使用局部微观统计特征多次推理有助于提高模型感知正确率。在所有实验中，强化学习优化的决策树模型规模缩小了60.0%～87.2%，并且Saras比Q-learning具有更好的优化表现。

关键词: 流量分析, 流量调度, 时延敏感属性, 强化学习, 决策树

Abstract:

Currently, network traffic scheduling strategy cannot be intelligent and on-demand, especially in the congestion caused by sudden network failures and escort scenarios of high-value services.They cannot guarantee latency-sensitive service experience on demand.The delay-sensitive attribute requirements of different network traffic were analyzed and studied, and the internal correlation between the behavior characteristics of varying network traffic and its delay sensitivity requirements was explored.Then, AI technology was used to learn this inherent relationship and construct its mapping relationship, realizing a traffic scheduling technical solution based on the intelligent awareness of delay-sensitive traffic.At the same time, considering the practical issues of interpretability and deploy-ability of AI models, reinforcement learning (RL) technology was used to prune and optimize the interpretable decision tree model, which improved the robustness of the model and made model lighter and easier to implement in equipment deployment.Through experiments by the collected real network traffic, the decision tree model optimized by reinforcement learning could improve the awareness accuracy by 1.75% in a single inference case, and the inference performance was improved by about 30%.The experiment also proved that using micro-statistical features for multiple inferences could help improve the model accuracy; in all experiments, the scale of the decision tree model optimized by RL was reduced by about 60.0%～87.2%, and the Saras had better optimization performance than Q-learning.

Key words: traffic analysis, traffic scheduling, delay-sensitive attribute, reinforcement learning, decision tree

中图分类号:

TP393

王雪荣, 唐政治, 李银川, 齐美玉, 朱建波, 张亮. 基于优化决策树的时延敏感流智能感知调度[J]. 电信科学, 2023, 39(4): 120-132.

Xuerong WANG, Zhengzhi TANG, Yinchuan LI, Meiyu QI, Jianbo ZHU, Liang ZHANG. Delay-sensitive traffic intellisense scheduling based on optimal decision tree[J]. Telecommunications Science, 2023, 39(4): 120-132.

图/表 15

图1

表1

图2

表2

表3

图3

图4

图5

图6

表5

图7

图8

表6

图9

表7

参考文献 17

[1]	张彤, 冯佳琦, 马延滢 ,等. 时间敏感网络流量调度综述[J]. 计算机研究与发展, 2022,59(4): 747-764.
	ZHANG T , FENG J Q , MA Y Y ,et al. Survey on traffic scheduling in time-sensitive networking[J]. Journal of Computer Research and Development, 2022,59(4): 747-764.
[2]	李文信, 齐恒, 徐仁海 ,等. 数据中心网络流量调度的研究进展与趋势[J]. 计算机学报, 2020,43(4): 600-617.
	LI W X , QI H , XU R H ,et al. Data center network flow scheduling progress and trends[J]. Chinese Journal of Computers, 2020,43(4): 600-617.
[3]	AURELI D , CIANFRANI A , DIAMANTI A ,et al. Going beyond DiffServ in IP traffic classification[C]// Proceedings of NOMS 2020 - 2020 IEEE/IFIP Network Operations and Management Symposium. New York:ACM Press, 2020: 1-6.
[4]	ALIZADEH M , YANG S , SHARIF M ,et al. pFabric:minimal near-optimal datacenter transport[J]. ACM SIGCOMM Computer Communication Review, 2013,43(4): 435-446.
[5]	BAI W , CHEN L , CHEN K ,et al. PIAS:practical information-agnostic flow scheduling for commodity data centers[C]// Proceedings of IEEE/ACM Transactions on Networking. Piscataway:IEEE Press, 2017: 1954-1967.
[6]	CHEN L , LINGYS J , CHEN K ,et al. AuTO:scaling deep reinforcement learning for datacenter-scale automatic traffic optimization[C]// Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication. New York:ACM Press, 2018: 191-205.
[7]	Intelligent buffer management on Cisco Nexus 9000 series switches white paper[EB]. 2017.
[8]	Application-aware routing[EB]. 2022.
[9]	WU M K , HUANG H , SUN Y ,et al. Active Keeper:an accurate and efficient algorithm for finding top-k elephant flows[J]. IEEE Communications Letters, 2021,25(8): 2545-2549.
[10]	XIE K , TIAN J Z , WANG X ,et al. Efficiently inferring top-k elephant flows based on discrete tensor completion[C]// Proceedings of IEEE INFOCOM 2019 - IEEE Conference on Computer Communications. Piscataway:IEEE Press, 2019: 2170-2178.
[11]	LOU K H , YANG Y J , WANG C C . An elephant flow detection method based on machine learning[C]// Proceedings of 2019 International Conference on Smart Computing and Communication. Cham:Springer, 2019: 212-220.
[12]	AL-SAADI M , KHAN A , KELEFOURAS V ,et al. Unsupervised machine learning-based elephant and mice flow identification[C]// Proceedings of the 2021 Computing Conference. Cham:Springer, 2021: 357-370.
[13]	ZUO P , SHU Y A . An elephant flows scheduling method based on feedforward neural network[C]// Proceedings of 2021 World Conference on Computing and Communication Technologies. Piscataway:IEEE Press, 2021: 16-20.
[14]	Configuration guidelines for DiffServ service classes[EB]. 2006.
[15]	WANG W , ZHU M , ZENG X W ,et al. Malware traffic classification using convolutional neural network for representation learning[C]// Proceedings of 2017 International Conference on Information Networking. Piscataway:IEEE Press, 2017: 712-717.
[16]	MOORE A W , ZUEV D . Discriminators for use in flow-based classification[R]. 2005.
[17]	SUTHAHARAN S . Decision tree learning[M]// Machine learning models and algorithms for big data classification. Boston: Springer, 2016: 237-269.

方案	局限性
ACL	维护复杂，表资源消耗大，手工配置困难
DSCP	可能出现误配错配，存在不信任问题
DPI	存在隐私安全问题，资源消耗大，加密流识别率低
基于流size	区分粒度粗，时敏大流无保障，小流误识别

服务类别	流量特征	网络性能			DSCP	应用举例
服务类别	流量特征	丢包率	时延	抖动	DSCP	应用举例
网络控制类	报文大小可变，主要为非弹性短消息，但也存在突发	低	低	是	CS6	网络路由协议
IP语音类	固定大小的小报文，发包率固定，非弹性和低速率	非常低	非常低	非常低	EF	IP语音承载
信令类	报文大小可变，一些是突发的且存活时间短的流量	低	低	是	CS5	IP语音信令
多媒体会议	报文大小可变，传输间隔固定，速率自适应，丢包敏感	低─中	非常低	低	AF4	H.323/V2 视频会议（自适应）
实时交互	实时传输协议/用户数据报协议流，非弹性，大部分是变速率	低	非常低	是	CS4	视频会议和交互性游戏
多媒体数据流	报文大小可变，弹性且变速率	低─中	中	是	AF3	点播流媒体视频和音频
广播视频	固定或可变速率，非弹性且无突发	非常低	中	低	CS3	广播电视和直播
低时延数据	可变速率、弹性的且存活时间短的弹性流量	低	低─中	是	AF2	电子商务及金融交易
操作维护管理	报文大小可变，弹性或者非弹性的流量	低	中	是	CS2	操作维护管理保障
高吞吐数据	可变速率、突发的且存活时间长的弹性流量	低	中─高	是	AF1	存储和转发应用
标准	都涉及	无具体要求			DF（CS0）	未分化应用
低优先级数据	非实时性和弹性的流量	高	高	是	CS1	任何不需要带宽保障的流量

服务类别	网络性能			DSCP	SLA 等级
服务类别	丢包率丢包率	时延时延	抖动抖动	DSCP	SLA 等级
网络控制类	低	低	是	CS6	1
IP语音类	非常低	非常低	非常低	EF	0
信令类	低	低	是	CS5	1
多媒体会议	低─中	非常低	低	AF4	0
实时交互	低	非常低	是	CS4	0
多媒体数据流	低─中	中	是	AF3	2
广播视频	非常低	中	低	CS3	2
低时延数据	低	低─中	是	AF2	1
操作维护管理	低	中	是	CS2	2
高吞吐数据	低	中─高	是	AF1	3
标准	无具体要求			DF（CS0）	3
低优先级数据	高	高	是	CS1	3

模型	叶子节点数/个	推理耗时/s
原始决策树模型	69 694	0.333 1
优化决策树模型	8 904	0.231 3

模型		原始决策树			优化决策树
模型	训练集正确率	测试集正确率	叶子节点数/个	训练集正确率	测试集正确率	叶子节点数/个
下行流模型	99.99%	86.39%	56 169	93.64%	86.88%	22 259
上行流模型	99.94%	98.05%	5 683	98.97%	97.94%	2 493
综合	90.57%			90.85%