Paper: A Decentralized SDN Architecture for the WAN (SIGCOMM’24)
Authors: Alexander Krentse, Nitika Saran, Bikash Koley, Subhasree Mandal, Ashok Narayanan, Sylvia Ratnasamy, Ali Al-Shabibi, Anees Shaikh, Rob Shakir, Ankit Singla, Hakim Weatherspoon
Presenter: Chengjing Zhou, Nankai University, China
Guest of Honor : Alex Krentsel
After the paper sharing, the participants had a lively discussion. The following is a partial Q&A record:
Q1: What is the cost and frequency of each node updating messages in the flooding algorithm? Will it consume a lot of bandwidth?
A1: It appears in one of the graphs where we compare the propagation time of cSDN to the propagation time of dsdn. Google actually has two wide area networks, one based on SDN, B4, and one based on traditional protocols, B2, and they are both global wide area networks. Flooding node update messages are similar to the IS-IS protocol link state update messages running in the B4 network, so we equivalent it by the propagation of the IS-IS protocol link state update messages in the B4 network, that is, extracting the logs from the routers in the B4 network to get the time required for one router to decide to flood the node update message until all other routers receive the message. It turns out that global flooding only takes a few hundred milliseconds from the first router sending the message to the last router confirming receipt of the flood message. In terms of bandwidth, the bandwidth requirements are very low, so this is actually not an issue we need to worry about.
Q2: Before running the TE algorithm, dSDN needs to know the traffic demand. Does this mean that dSDN is only applicable to stable networks? Is it applicable in more dynamic networks?
A2: Traditional network flooding protocols, such as link-state protocols like OSPF, flood topology information, but not demand information. The messages that our proposed dSDN floods include not only topology information, but also demand information. Therefore, it is possible to run some applications on the routers that sample packets and build an approximate view of demand, especially packets coming from external interfaces. The router can get the traffic rate entering the network at each node, which is the information that dSDN includes in the node update message, so it also works for dynamic networks. In fact, we have a fairly dynamic network at Google, and dSDN can also be used in dynamic networks.
Q3: Does dSDN only support connection-based data transmission? Can it support connectionless data transmission?
A3: I think dSDN is completely independent of whether a connection is established or not. It’s really a question of how to estimate your traffic demand. Currently, we collect demand from hosts, asking them how much data they sent and then summing up the information to get the send rate of each application. Another way to estimate demand is to sample packets from routers, and the sampling is independent of whether there is a connection or not. However, this estimate may be less accurate if it is not based on the connection demand estimate, but I don’t think it will make a big difference.
For more information about the discussion session, please watch our meeting recording, which has been uploaded to YouTube. Thank you for your support!