The Internetworking Problem

Massachusetts Institute of Technology Department of Electrical Engineering and Computer Science Lecture 2: The Internetworking Problem Overview. This lecture looks at the issues that arise in internetworking different networks together. We will study the Internet design principles, IP, the Internet's best-effort service model, its design and protocol standardization philosophy, and IP. We then discuss the split between IP and TCP, and study how TCP achieves reliable data transport. We will look at TCP's cumulative acknowledgment feedback, and two forms of retransmission upon encountering a loss: timer-driven and data-driven retransmissions. The former relies on estimating the connection round-trip time (RTT) to set a timeout, whereas the latter relies on the successful receipt of later packets to initiate a packet recovery without waiting for a timeout. Examples of data-driven retransmissions include the fast retransmission mechanism and selective acknowledgment (SACK) mechanism. Readings: CK74, Cla88. 1 The Internetworking Problem: Many Different Networks In the previous lecture, we discussed the principles underlying packet switching, and developed a scheme for connecting local-area networks (LANs) together. The key idea was a device called a switch, which forwarded packets between different LANs. While such networks work and are common, they are not enough to build a global data networking infrastructure. They have two main failings: 1. They don't accommodate heterogeneity. 2. They don't scale well to large networks. Each point bears some discussion. Although commonplace, Ethernets are by no means the only network link technology in the world. Indeed, even before Ethernets, packets were being sent over other kinds of links including wireless packet radio, satellites, and telephone wires. Ethernets are also usually ill-suited for long-haul links that have to traverse hundreds of miles. Advances in link technologies of various kinds has always been occurring, and will continue to occur, and our goal is to ensure that any new network link technology can readily be accommodated into the global data networking infrastructure. This is the goal of link heterogeneity. The LAN switching solution we studied in Lecture 1 requires switches to store per-host (actually, per-interface) state in the network; in the absence of such state for a destination, switches revert to broadcasting the packet to all the links in the network. This does not scale to more than a moderate number of hosts and links. The internetworking problem may be stated as follows: Design a scalable network infrastructure that interconnects different smaller networks, to enable packets to be sent between hosts across networks of very different performance characteristics. Some examples of the differing characteristics include: Massachusetts Institute of Technology Department of Electrical Engineering and Computer Science Lecture 2: The Internetworking Problem Overview: This lecture looks at the principles that arise in internetworking different networks together. We will study the Internet's design issues. The Internet's best-effort service model, its design, and protocol standardization philosophy, and TCP and IP. We will discuss the cumulative acknowledgment and selective acknowledgment schemes for TCP. We will then look at TCP's split between IP and TCP, and study how TCP achieves reliable data transport upon encountering a packet loss: timer-driven packet recovery without waiting for a time-out. Examples of data receipt mechanisms include the fast retransmission mechanism and the selective acknowledgment (SACK) mechanism. Readings: CK74, Class8. 1 The Internetworking Problem: Many Different Networks In the previous lecture, we discussed the principles underlying packet switching, a device called a switch, for connecting local-area networks (LANs) together. The key idea was and developed a scheme in which forwarded packets between different LANs.   While such networks have two main failings, they are not enough to build a global data networking infrastructure. They work and are common, They don't scale well to large networks. They don't accommodate heterogeneity. Each network joins some low-haul link. Although commonplace, Ethernets are by no means the only network link technology in the world. Indeed, even before Ethernets, packets were being sent over telephone lines, radio, and other media. The Internet is built on the idea that packets can also be sent over point-to-point links for long-haul, that packets have to traverse hundreds of miles. Advances in link technologies of various kinds has always been occurring, and will continue to occur, and our goal is to ensure that any new network link technology can link readily be accommodated into the global data networking infrastructure. This is the goal of link heterogeneity.   The LAN switching solution we studied in the lecture 1 requires switches to store packets (actually, per-interface) state in all the links in the absence of such states for a destination, such switches revert to broadcasting the packet to all networks; in the absence of this state does not scale to more than a moderate number of hosts and links.   The internetworking problem may be stated as follows: Design a scalable network infrastructure that interconnects different smaller networks, to enable packets to be sent between hosts across networks of very differing performance characteristics.   Some examples of the differing characteristics include: Addressing. Each network might have its own way of addressing nodes; for example, Ethernet uses a 6-byte identifier while telephones use a 10-digit number.   Bandwidth & latency. Heterogeneity in bandwidths range from a small number of bits/s (e.g., power lines) to many Gigabits/s, spanning many orders of magnitudes. Similarly, latencies can range from microseconds to several seconds.   Packet size. The maximum packet sizes in general will vary between networks. Loss rates. Networks differ widely in the loss rates and loss patterns of their constituent links.   Packet routing. Packet routing could be handled differently by each constituent network. Our scalability goal is to reduce the state maintained by the switches and the bandwidth consumed by control traffic (e.g., routing messages, periodic broadcasts of packets/messages, etc.) as much as we can.   1.1 Gateways Communication between different networks in an internetwork is facilitated by entities called gateways. Gateways interface between two networks with potentially differing characteristics and move packets between them. There are at least two ways of interfacing between different networks: translation and a unified network layer.   Translation-based gateways work by translating packet headers and protocols over one network to the other, trying to be as seamless as possible. Examples of this include the OSI X.25/X.75 protocol family, where X.75 is a translation-based "exterior" protocol between gateways. There are some major problems with translation.   Feature deficiency. As network technologies change and mature and new technologies arise, there are features (or bugs!) in them that aren't preserved when translated to a different network. As a result, some assumption made by a user or an application breaks, because a certain expected feature is rendered unusable since it wasn't supported by the translation.   Poor scalability. Translation often works in relatively small networks and in internetworks. where the number of different network types is small. However, it scales poorly because the amount of information needed by gateways to perform correct translations scales proportional to be square of the number of networks being inter-connected. This approach is believed to be untenable for large internetworks. (As an aside, it's worth noting that the widespread deployment of NATs is making portions of the Internet look a lot like translation gateways!)   1.2 Design Principles The designers of the Internet (the "big I" Internet!) protocols recognized the drawbacks of translation early on and decided that the right approach was to standardize some key properties across all networks, and define a small number of features that all hosts and networks connected to the Internet must implement.