SOSR 17

概要

网络流量中有一部分是用于网络管理&＃xff0c;&＃xff08;根据packet process survey&＃xff0c;该部分流量属于包转发的slow path部分&＃xff09;由于sdn的数控分离&＃xff0c;交换机需要向控制器发送大量控制信息&＃xff0c;数据平面和控制平面的控制流量比传统网络的要多。本文提出了一个算法用于测量sdn环境下的控制流量&＃xff0c;该算法基于以下三点来测量&＃xff1a;数据平面和控制平面周期性产生控制流量、通过生成树协议的信息来估计控制流量、通过全网的主机和交换机数目来估计控制流量。通过仿真实验验证&＃xff0c;错误率可在3%以下。

Abstract

• The data packet statistics sent by OpenFlow compliant switches cumulatively includes statistics about control traffic which is used for network control and management. This reduces the accuracy of calculation of QoS metrics and thus hampers network monitoring. We present here a novel algorithm to accurately measure the fraction of control packets in SDN within 3% error rate.

Bckground

• A fraction of traffic in the network is responsible for network control & manangement like monitoring, enforcing security, calculating QoS metrics etc.
• such as: MDNS, NDP, MLD, DHCP etc.

In SDN

• Many services run in the SDN controller continuously in the background that generate a lot of control packets, e.g. topology discovery, network monitoring via packet injection, pushing configurations etc.
• Therefore, more amount of control packets are generated in SDN than traditional networks.
• Considerable number of LLDP packets are injected by the controller for discovering links.

relative work

• PayLess proposes a frequency adaptive
  statistics collection scheduling** algorithm & Pakzad et al.

• propose a new approach to reduce processing cost due to topology discovery in the controller with a minimum reduction of 67% in terms of messages.

• For auto configuration in SDN, extensions to current protocols such as DHCP-SDN have been proposed that will lead to even greater fraction
  of control traffic.

Solution

• we propose an algorithm to measure the
  fraction of the control traffic in Software Defined Networks.

How to measure control traffic

• The message exchanges are periodic in nature, e.g.the router send control packet every 30 sec to discover network.
• The number of control packets in a subnet being proportional to the number of switches & hosts in the network can help to estimate the total number of control packets.
• Using the spanning tree information, the number of control packets through each link can be estimated

Algorithm

CONCLUSION

• While emulating various topologies (ring, tree, mesh, star etc.) with different number of switches (4{16), links (8{32) and switch degrees, we measured the control packets transferred and compared it with qdisc linux utility for accuracy. We successfully report the control packets within 3% error rate approximately for the networks,increasing accuracy of calculation of QoS metrics like packet loss.