The modern internet has revolutionized how we communicate with one another. However, it had humble beginnings in the Victorian era. One of the earliest precursors to the internet was telegraph networks which were operational as early as 1850. Back then, it used to take 10 days to send a message from Europe to North America by sea. Telegraph networks reduced that to 17 hours. By the late 19th century, the telegraph was a fully successful communication technology that was used widely in the two world wars. Around that time, people started building computers to help in cracking enemy codes. Unlike our modern mobile phones and laptops, those computing machines were often huge and needed specialized environments to be able to operate smoothly. Thus, it was necessary to put those in special locations while the operators would sit on a terminal. The terminal needed to be able to communicate with the computer over short distances. A number of local area networking technologies enabled this, the most prominent one being Ethernet. Over time, these networks grew and by the 1960s, some of these networks were being connected with one another to form a larger network of networks. The Advanced Research Projects Agency Network (ARPANET) was established in 1969 and it became the first internetwork that resembles the modern internet. Around 1973, there were a number of such internetworks all around the world, each using their own protocols and methods for communication. Eventually, the protocols were standardized so that the networks could communicate with each other seamlessly. All of these networks were later merged to form what is the internet today.

Since networks evolved in silos all around the world, they were often organized according to geographical proximity. A Local Area Network (LAN) is a collection of host machines in small proximity like a building or a small neighborhood. A Wide Area Network (WAN) is one that connects multiple neighborhoods; the global internet is at the top of the hierarchy. The next picture shows a map of the ARPANET in 1977. Each node in this map is a computer (a server, in today's terms). Most of these were located in large universities like Stanford or at national laboratories like Lawrence Berkeley (source: https://commons.wikimedia.org/wiki/File:Arpanet_logical_map,_march_1977.png).

Computer science often focuses on subdividing a problem into smaller, hopefully independent components that can be solved in isolation. Once that is done, all that is needed is a set of rules on how those components should communicate to have a solution to the larger problem. This set of rules, along with a pre-agreed data format, is called a protocol. A network is composed of a number of layers, each of which has a fixed purpose. Thus, each of these layers run one or many protocols, forming a stack of protocols. In the early days of networking, different people implemented their networks in different ways. When the internet was conceived, there was a need to make these networks communicate seamlessly. Since they were constructed differently, this turned out to be difficult.

There was a clear need to agree on standard protocols and interfaces to make the internet work. The first attempt at standardizing networking protocols was in 1977, which led to the OSI model. This model has the following layers:

• Physical layer: It defines how data is transmitted in the physical medium in terms of its electrical and physical characteristics. This can either be by wire, fiber optic, or a wireless medium.
• Data link layer: It defines how data is transmitted between two nodes connected by a physical medium. This layer deals with prioritization between multiple parties trying to access the wire simultaneously. Another important function of this layer is to include some redundancy in the transmitted bits to minimize errors during transmission. This is referred to as coding.
• Network layer: It defines how packets (made up of multiple units of data) are transmitted between networks. Thus, this layer needs to define how to identify hosts and networks uniquely.
• Transport layer: It defines mechanisms to reliably deliver variable length messages to hosts (in the same or different networks). This layer defines a stream of packets that the receiver can then listen to.
• Session layer: It defines how applications running on hosts should communicate. This layer needs to differentiate between applications running on the same host and deliver packets to them.
• Presentation layer: It defines common formats for data representation so that different applications can interlink seamlessly. In some cases, this layer also takes care of security.
• Application layer: It defines how user-centric applications should send and receive data. An example is the web browser (a user-centric application) using HTTP (an application layer protocol) to talk to a web server.

The following figure shows a visual representation of this model (source: https://commons.wikimedia.org/wiki/File:Osi-model-jb.svg). This also shows two vertical classifications, the host running the network stack and the physical media (including the wire and the network device). Each layer has its own data unit, the representation of the information it works on, and since each layer encapsulates the one below it, the data units encapsulate too. A number of bits form a frame, a number of frames form a packet, and so on, to the top:

While OSI was working on standardizing this model, Defense Advanced Research Projects Agency (DARPA) came up with a full implementation of the much simpler TCP/IP model (also known as the IP (Internet Protocol) suite). This model has the following layers, from closest to the physical medium to the farthest:

• Hardware interface layer: This is a combination of layers one and two of the OSI model. This layer is responsible for managing media access control, handling transmission and reception of bits, retransmission, and coding (some texts on networking differentiate between the hardware interface layer and the link layer. This results in a five layer model instead of four. This hardly matters in practice, though.)
• IP layer: This layer corresponds to layer three of the OSI stack. Thus, this layer is responsible for two major tasks: addressing hosts and networks so that they can be uniquely identified and given a source and a destination address, and computing the path between those given a bunch of constraints (routing).
• Transport layer: This layer corresponds to layer four of the OSI stack. This layer converts raw packets to a stream of packets with some guarantees: in-order delivery (for TCP) and randomly ordered delivery (for UDP).
• Application layer: This layer combines layers five to seven of the OSI stack and is responsible for identifying the process, data formatting, and interfacing with all user level applications.

Note that the definition of what a particular layer handles changes as we move from one layer to another. The hardware interface layer handles collection of bits and bytes transmitted by hosts, the IP layer handles packets (the collection of a number of bytes sent by a host in a specific format), the transport layer bunches together packets from a given process on a host to another process on another host to form a segment (for TCP) or datagram (for UDP), and the application layer constructs application specific representations from the underlying stream. For each of these layers, the representation of data that they deal with is called a Protocol Data Unit (PDU) for that layer. As a consequence of this layering, when a process running on a host wants to send data to another host, the data must be broken into individual chunks. As the chunk travels from one layer to another, each layer adds a header (sometimes a trailer) to the chunk, forming the PDU for that layer. This process is called encapsulation. Thus, each layer provides a set of services to layers above it, specified in the form of a protocol.

The modern internet exhibits a form of geographical hierarchy. Imagine a number of homes which are served by a number of Internet Service Providers (ISPs). Each of these homes is in a LAN (either via Ethernet, or more commonly, Wi-Fi). The ISP connects many such LANs in a WAN. Each ISP has one or many WANs that they connect to form their own network. These larger networks, spanning cities, which are controlled by a single business entity, are called Administrative Systems (AS). Routing between multiple ISPs is often more complex than regular IP routing since they have to take into account things like trading agreements and so on. This is handled by specialized protocols like the Border Gateway Protocol (BGP).

As mentioned before, one of the earliest and most successful networking technologies is Ethernet. First introduced in 1974, it quickly became the predominant technology for LAN and WAN due to its low cost and relative ease of maintenance. Ethernet is a shared media protocol where all the hosts must use the same physical medium to send and receive frames. Frames are delivered to all hosts, which will check if the destination MAC address (these addresses will be described in the next section) matches its own address. If it does, the frame is accepted, otherwise, it is discarded. Since the physical medium can only carry one signal at any given moment, there is a probability that frames might collide in transit. If that does occur, the sender can sense the collision by sensing transmission from other hosts while it is transmitting its frame. It then aborts the transmission and sends a jam signal to let other hosts know of the collision. Then, it waits for an exponentially backed off amount of time and retries the transmission. After a fixed number of attempts, it gives up if the transmission does not succeed.

This scheme is called carrier-sense multiple access with collision detection (CSMA/CD). One problem with Ethernet is its relatively short range. Depending on the physical wiring technology used, the maximum length of an Ethernet segment varies between 100 m to 500 m. Thus, multiple segments must be connected to form a larger network. The most common way of doing that is using layer two switches between two adjacent Ethernet segments. Each port of these switches forms different collision domains, reducing the overall probability of collisions. These switches can also monitor traffic to learn which MAC addresses are on which ports so that eventually, they will send out frames for that MAC address only on that port (referred to as a learning switch). In modern homes, Wi-Fi is often the dominant LAN technology compared to Ethernet.

We have seen why it is important to identify hosts and networks uniquely to be able to deliver packets reliably. Depending on the scale, there are three major ways of doing this; we will discuss each of those in this section. The end to end process of IP routing will be discussed in the next section. One interesting fact to note is that for each of these addressing modes, one or more addresses are reserved for special use. Often, these are marked by a known set of bits being on or off in a known pattern:

• Ethernet address: This is also known as a Media Access Control (MAC) address. It is a 48-bit long unique identifier assigned to a network device (usually stored on the card) that is used to identify it in a network segment. Usually, these are programmed by the network card manufacturer, but all modern OS's allow one to modify it. The standard way of writing Ethernet addresses are in six groups of two hexadecimal digits (01-23-45-67-89-ab-cd-ef). Another common way is to use a colon to separate the digits (01:23:45:67:89:ab:cd:ef). A few special sequences of bits are reserved for addressing special cases: the sender can request that an Ethernet frame should be received by all hosts in that segment by setting the least significant bit of the first octet to 1; this is called multicasting. If that particular bit is set to 0, the frame should be delivered to only one receiver. Today, these are used widely with Ethernet and Wi-Fi.
• IP address: This is an address assigned to each device in an IP network. The original IP address standard (IPv4) defined 32-bit addresses in 1980. However, by 1995, it was obvious that the total number of available addresses on the internet is not enough to cover all devices. This led to the development of IPv6, which expanded the address space to 128 bits. The standard way of dealing with a group of IP addresses is using the CIDR notation, for example, 192.168.100.1/26 (IPv4). The decimal number after the slash counts the number of leading 1s in the network mask. Thus, in this particular case, there are 2^(32-26) = 64 addresses in the network starting from 192.168.100.0 to 192.168.100.63. The Internet Assigned Numbers Authority (IANA) assigns blocks of publicly routable IP addresses to organizations. A number of IPv4 and v6 addresses are reserved for various purposes like addressing in private networks and so on. In a home network (which will always use special private range addresses), these are assigned by the Dynamic Host Configuration Protocol (DHCP) by the Wi-Fi router.
• Autonomous system number: This is a 32-bit number used to uniquely identify autonomous systems. Like IP addresses, these are assigned and maintained by the IANA.

Apart from these, communication between hosts often uses a port number to distinguish between processes. When the OS allocates a specific port to a process, it updates its database of the mapping between process identifier and port number. Thus, when it receives incoming packets on that port, it knows what process to deliver those packets to. In case the process has exited by that time, the OS will drop the packets and in the case of TCP, initiate closing of the connection. In the subsequent sections, we will see how TCP works in practice.

To understand how IP routing works, we must first begin with the structure of IPv4 addresses. As described in the last section, these are 32 bits in length. They are written in a dotted decimal notation in groups of 4 bytes (for example, 192.168.122.5). A given number of bits in that network prefix is used to identify the network where the packet should be delivered, and the rest of the bits identify the particular host. Thus, all hosts in the same network must have the same prefix. Conventionally, the prefix is described in the CIDR notation with the starting address and the number of bits in the network portion of the address separated by a slash (192.168.122.0/30). The number can then be used to find out how many addresses are available for hosts in the network (in this case, 2^(32-30) = 4). Given an IP address and a prefix, the network address can be extracted by bitwise-ANDing the address with a mask of all 1s in the network portion. Calculating the host address is just the reverse; we will need to AND with the network mask's logical negation (the host mask), which has all 0s in the network portion and all 1s in the host portion. Given an address and a prefix like 192.168.122.5/27, we will compute these as shown in the following figure. Thus, for the given CIDR, the network address is 192.168.122.0 and the host address is 0.0.0.5:

There are two broad classes of IP address; some blocks of addresses can be routed in the public internet, these are called public IP addresses. Some other blocks can only be used in private networks that do not directly interface with the internet, these are called private addresses. If a router on the internet receives a packet that is destined for a private IP address, it will have to drop that packet. Other than these two, IP addresses are also classified on various parameters: some are reserved for documentation only (192.0.2.0/24), some are reserved for point to point communication between two hosts (169.254.0.0/16), and so on. The Rust standard library has convenience methods to classify IP addresses according to their types.

All routers maintain a routing table which maps prefixes to the outgoing interface of the router (while a router administrator might decide to store individual addresses instead of prefixes, this will quickly lead to a large routing table in a busy router). An entry in the table basically says If a packet needs to go to this network, it should be sent on this interface. The next host that receives the packet might be another router or the destination host. How do routers figure out this table? Multiple routers run routing protocols between those which compute those tables. Some common examples are OSPF, RIP, and BGP. Given these primitives, the actual routing mechanism is fairly simple, as shown in the next diagram.

An interesting aspect of IP is the use of the Time To Live (TTL) field, this is also known as hop limit. The host sends out packets with a fixed value of TTL (usually 64). Each router the packet crossed decreases the TTL. When it reaches 0, the packet is discarded. This mechanism ensures that packets are not stuck in an infinite loop between routers:

Note that while trying to match the prefix to routes in the routing table, multiple routes might match. If that happens, the router must select the most specific match and use that for forwarding. Since the most specific routes will have the maximum number of leading 1s, and hence the largest prefix, this is called the longest prefix match. Say our router has the following routing table, as shown in the diagram. eth1, eth2, and eth3 are three network interfaces attached to our router, each having a different IP address in different networks:

At this point, if our device gets a packet that has a destination address set to 192.168.1.33, all three prefixes have this address but the last one is the largest of the three. So, the packet will go out through eth3.

A lot of what we described so far about IPv4 addresses does not change for IPv6, except, of course, it has a larger address space of 128 bits. In this case, the length of the network mask and the host mask depends on the address type.

One might be wondering, how do routers construct the routing table? As always, there are protocols to help with that. Routing protocols are of two major types: interior gateway protocols which are used for routing inside an autonomous system, and exterior gateway protocols which are used in routing between autonomous systems; an example of the latter is BGP. Interior gateway protocols can again be of two types, depending on how they look at the whole network. In link state routing, each router participating in the protocol maintains a view of the whole network topology. In distance vector routing, each router only knows about its one hop neighbors. An example of the former is the Routing Information Protocol (RIP) and of the latter is Open Shortest Path First (OSPF). Details about these are beyond the scope of this book. However, we can note that the common theme among all the routing protocols is that they work by exchanging information between routers. Thus, they have their own packet formats for encapsulating that information.