Overview of the Network Layer · Computer Networking Interactive

The network layer moves packets from one host to another, it runs in every router as well as in every host, and it promises nothing about whether they arrive.

Words you will meet

Forwarding — moving a packet from an input link to an output link, inside one router. Nanoseconds. Hardware.
Routing — working out the whole path across the network. Seconds. Software.
Data plane — the per-router machinery that does the forwarding.
Control plane — the network-wide logic that decides what the forwarding machinery should do.
Forwarding table — the table in a router that maps a header value to an outgoing interface.
Best-effort service — the Internet’s network-layer service model. It promises nothing.
Network service model — the list of what a network layer does and does not guarantee.

Why this matters

Every earlier chapter ran in end systems. Open a browser, run a server, write to a socket — all of it happened on machines at the edge. The network in the middle was a cloud that either delivered your packet or did not.

This chapter opens that cloud. And the first thing to know is structural: unlike the application and transport layers, there is a piece of the network layer in every host and in every router. That is why the book calls network-layer protocols the most challenging in the stack, and why it needs two chapters rather than one.

The second thing to know is the answer to a question chapter 3 left hanging. Section 3.1 said a transport protocol cannot offer a service the network layer underneath it does not support, and then never said what the network layer does support. This section says: almost nothing. Everything TCP (Transmission Control Protocol) achieves, it achieves on top of a layer that will not even promise to deliver the packet.

The network layer is in the middle, not at the edge

Figure 4.1 is one picture making one point. Two end systems, several routers between them, and each device drawn with the layers it actually implements.

Figure 4.1 — the network layer runs in every host and every router

1 / 25

End system H1 · Application — The application has a message to send. At this point it is just data — nothing has been added to it yet.

Scrub through it. The two end systems have all five layers. Every router in between has only three — and that truncated stack is the whole point of the figure.

Read all 25 steps as text

End system H1 · Application [a segment from the transport layer] — The application has a message to send. At this point it is just data — nothing has been added to it yet.
End system H1 · Transport [Ht | a segment from the transport layer] — The transport layer puts its header Ht in front. Everything below now treats the message as an opaque payload. This is a segment.
End system H1 · Network [Hn | Ht | a segment from the transport layer] — The network layer adds Hn, carrying the addresses of the two hosts. This is a datagram.
End system H1 · Link [Hl | Hn | Ht | a segment from the transport layer] — The link layer adds Hl, which describes how to cross this one link. This is a frame.
End system H1 · Physical [Hl | Hn | Ht | a segment from the transport layer] — The physical layer puts the individual bits of the frame onto the medium.
Router R1 · Physical [Hl | Hn | Ht | a segment from the transport layer] — The bits arrive and are reassembled into a frame.
Router R1 · Link [Hn | Ht | a segment from the transport layer] — The old link header is removed and discarded. It described the link the frame just crossed, and that link is behind us now.
Router R1 · Network [Ht | a segment from the transport layer] — The router reads the destination address in Hn and chooses the outgoing link. It stops here: it does not implement the transport layer, so Ht means nothing to it.
Router R1 · Link [Hl | Ht | a segment from the transport layer] — A brand-new link header is added for the next link. It may be a completely different link technology — Ethernet here, WiFi next — and the header changes to match.
Router R1 · Physical [Hl | Ht | a segment from the transport layer] — The frame goes out onto the next link as bits.
a router in the core · Physical [Hl | Hn | Ht | a segment from the transport layer] — The bits arrive and are reassembled into a frame.
a router in the core · Link [Hn | Ht | a segment from the transport layer] — The old link header is removed and discarded. It described the link the frame just crossed, and that link is behind us now.
a router in the core · Network [Ht | a segment from the transport layer] — The router reads the destination address in Hn and chooses the outgoing link. It stops here: it does not implement the transport layer, so Ht means nothing to it.
a router in the core · Link [Hl | Ht | a segment from the transport layer] — A brand-new link header is added for the next link. It may be a completely different link technology — Ethernet here, WiFi next — and the header changes to match.
a router in the core · Physical [Hl | Ht | a segment from the transport layer] — The frame goes out onto the next link as bits.
Router R2 · Physical [Hl | Hn | Ht | a segment from the transport layer] — The bits arrive and are reassembled into a frame.
Router R2 · Link [Hn | Ht | a segment from the transport layer] — The old link header is removed and discarded. It described the link the frame just crossed, and that link is behind us now.
Router R2 · Network [Ht | a segment from the transport layer] — The router reads the destination address in Hn and chooses the outgoing link. It stops here: it does not implement the transport layer, so Ht means nothing to it.
Router R2 · Link [Hl | Ht | a segment from the transport layer] — A brand-new link header is added for the next link. It may be a completely different link technology — Ethernet here, WiFi next — and the header changes to match.
Router R2 · Physical [Hl | Ht | a segment from the transport layer] — The frame goes out onto the next link as bits.
End system H2 · Physical [Hl | Hn | Ht | a segment from the transport layer] — The bits arrive and are reassembled into a frame.
End system H2 · Link [Hn | Ht | a segment from the transport layer] — The old link header is removed and discarded. It described the link the frame just crossed, and that link is behind us now.
End system H2 · Network [Ht | a segment from the transport layer] — The network layer removes Hn and passes the segment up.
End system H2 · Transport [a segment from the transport layer] — The transport layer removes Ht, checks it, and works out which application the message belongs to.
End system H2 · Application [a segment from the transport layer] — The message arrives at the application, byte for byte what was sent.

The end systems H1 and H2 have all five layers. Every router has three: network, link, physical. The book draws them with a truncated stack on purpose — routers do not run application- and transport-layer protocols.

What the network layer does at each place:

At H1, it takes a segment from the transport layer, wraps it in a datagram , and hands the datagram to the nearby router R1.
At every router, it forwards the datagram from an input link to an output link.
At H2, it receives datagrams, extracts the segments, and passes them up.

Compare with the transport layer, one chapter ago

Figure 3.1 is the same shape of picture with the opposite lesson. There, every router showed an empty dashed box where the transport layer would be, because a segment passes through the network untouched.

Here, the network layer box is filled in at every single hop. That is the difference between a service implemented only at the two ends and one implemented everywhere along the path. It is also why changing TCP means updating end systems, while changing IP (Internet Protocol) means updating the entire Internet.

Forwarding and routing are not the same word

The book is blunt about this: many authors use forwarding and routing interchangeably, and this book will not.

	Forwarding	Routing
What	move a packet from an input link interface to the right output link interface	determine the end-to-end path packets take from source to destination
Where	inside one router, local	across the whole network
How long	a few nanoseconds	typically seconds
Built from	hardware	software
Which plane	data plane	control plane

Everyday picture — the drive to Florida

Chapter 1.3.1 sent a traveller from Pennsylvania to Florida.

Forwarding is getting through one interchange. The car arrives on one road, and something decides which road it leaves by. It happens in seconds, over and over, and it needs no knowledge of Florida.

Routing is planning the trip. Before setting off, the driver consults a map and picks one of many possible paths, each a series of road segments joined at interchanges.

Where the picture stops. A driver plans their own route and then executes it. In a network the two jobs are done by different machinery, on different timescales. The router doing the forwarding never sees the plan — only the one table entry that the plan produced. No driver forgets where they are going between interchanges; a router never knew.

The forwarding table is the whole decision

A router forwards a packet in three moves. It examines one or more fields in the arriving packet’s header. It uses those values to index into its forwarding table. It sends the packet out of the interface the table names.

Figure 4.2 — a routing algorithm inside every router fills its own forwarding table

1 — A routing algorithm runs inside this routerstep 1 of 5

It talks to the routing algorithms in the neighbouring routers, exchanging routing messages according to a routing protocol. This is the traditional arrangement, and it is what chapter 5 is about.

The book’s own example. A packet arrives with 0110 in the header field the router looks at. Step forward to watch the lookup, then compare with Figure 4.3 below — the bottom half of the picture never changes.

Read all steps as text

1 — A routing algorithm runs inside this router — It talks to the routing algorithms in the neighbouring routers, exchanging routing messages according to a routing protocol. This is the traditional arrangement, and it is what chapter 5 is about.
2 — It computes this router’s forwarding table — The table maps a header value to an outgoing interface. Nothing else in the router decides where packets go; the table is the decision, already made.
3 — A packet arrives with header value 0110 — The router does not care what the packet contains, where it came from, or what it is for. It reads one field.
4 — The value indexes into the table — Row two matches: header 0110, output 2. This lookup is the whole of forwarding, and on a real router it takes a few nanoseconds.
5 — The packet goes out on interface 2 — And that is the data plane, complete. Notice what never happened: this router did not consider the destination’s distance, the state of the wider network, or any alternative route. All of that thinking happened earlier, in the control plane.

In plain words

The router does not think. The thinking already happened, somewhere else, and its entire result is one row of a table.

This is worth holding on to, because it is what makes forwarding fast enough to be useful. A packet arriving on a 100 Gbps link leaves a few nanoseconds to decide what to do with it. There is no time to consult a map, ask a neighbour, or reconsider. There is time to read a field and look up a row.

Two ways to fill the table in

So how does the table get its contents? The book gives two answers, and the difference between them is the shape of the modern network layer.

The traditional approach

A routing algorithm runs in every router. It communicates with the routing algorithms in other routers, exchanging routing messages according to a routing protocol, and from those messages it computes its own forwarding table. Forwarding and routing both live inside the router. Sections 5.2 to 5.4 cover the algorithms and the protocols.

The software-defined approach

Figure 4.3 — a remote controller computes the tables instead

1 — The routing computation has left the routerstep 1 of 3

A physically separate, remote controller computes the forwarding tables for every router it manages. There is no routing algorithm inside this router at all.

Compare it with Figure 4.2 step 5. The data plane below the dashed line is identical, byte for byte. Only the thing above the line has moved.

Read all steps as text

1 — The routing computation has left the router — A physically separate, remote controller computes the forwarding tables for every router it manages. There is no routing algorithm inside this router at all.
2 — It distributes a table to each router — The controller has a view of the whole network, so it can compute tables that no single router could work out alone. Chapter 5 covers how, and what it costs.
3 — The packet arrives, and nothing below the line has changed — Same header value, same table, same lookup, same output interface. This is the point of drawing the two figures the same way: separating the control plane changes who fills the table in, and changes nothing whatsoever about how a packet is forwarded.

A physically separate, remote controller computes the forwarding tables and distributes them to every router. The routers still forward exactly as before — they simply no longer work out the tables themselves. This is SDN (Software-Defined Networking) , and chapter 5 covers the controllers.

The book’s own thought experiment

Before either approach, the book asks you to imagine something deliberately absurd: a network where every forwarding table is configured by hand, by human operators standing at the routers.

It would work. No routing protocol would be needed at all. The operators would have to agree with each other so that packets actually reached their destinations — but nothing about forwarding requires an algorithm.

The point of the thought experiment is to separate two things that usually arrive together. Forwarding needs a table. It does not care where the table came from. Once you see that, a remote controller filling the table in is not a strange idea at all. It is one of the obvious options, and humans with keyboards was always the other one.

The book adds the honest caveat: hand configuration would be more error-prone and much slower to respond to a change in the network than a routing protocol.

What the network layer promises

Now the deflating part. What service does a network layer give the transport layer above it?

The book lists what a network layer could offer, then says what the Internet does offer.

What a network layer could promise, and what the Internet promises

	Could be offeredthe book’s list of candidates	The Internetbest-effort service	ATMa network that did offer more
Guaranteed deliverythe packet will eventually arrive
Bounded delayand it will arrive within, say, 100 ms
In-order deliverypackets arrive in the order sent
Minimum bandwidthbehaves like a link of a stated bit rate
Confidentialitydatagrams encrypted at the source, decrypted at the destination

Cells marked ⓘ have a reason behind them — click to read it.

Five services the book lists as possible, against the one the Internet actually offers. Click any cell for the reasoning.

The Internet’s network layer provides a single service, known as best-effort service . Packets are not guaranteed to arrive in the order sent, nor guaranteed to arrive at all. There is no delay bound and no minimum bandwidth.

”Best effort” is a euphemism, and the book says so

Read the definition again and notice what it excludes: nothing.

The book makes the point itself — a network that delivered no packets at all would satisfy the definition of best-effort delivery service. The name promises effort; the specification promises nothing.

This is not a criticism, and it is not an accident. It is the design decision that everything else in this book rests on. Section 4.5 comes back to it as an architectural principle. The goal is connectivity, the tool is the Internet Protocol, and the intelligence is end to end rather than hidden in the network.

So why did best effort win?

Alternatives existed and were deployed. The ATM (Asynchronous Transfer Mode) architecture provided guaranteed in-order delivery, bounded delay and guaranteed minimum bandwidth. The Intserv architecture [RFC 1633] proposed end-to-end delay guarantees for the Internet itself.

The book’s answer is not that the alternatives were wrong. It is that best effort turned out to be good enough, when combined with two other things:

adequate bandwidth provisioning — build enough capacity and congestion becomes rare;
bandwidth-adaptive application-level protocols — such as the DASH (Dynamic Adaptive Streaming over HTTP) streaming protocol of section 2.6.2. It changes what it asks for based on what it is getting.

Between them, those two allow Netflix, video over IP (Internet Protocol) , Skype and FaceTime to run over a network layer that guarantees them nothing at all. The book calls this “interesting”, and it is. The guarantees moved out of the network and into the applications, and the applications turned out to be a better place for them.

One piece of vocabulary, used carefully

The book warns about two pairs of words before going further.

Forwarding and switching are used interchangeably by researchers and practitioners, and the book says it will use both. This site will not. Here, forwarding is always the action, and switching fabric is the piece of hardware inside a router that carries packets across. That keeps one word for one idea, as everywhere else on this site.

Packet switch, router and link-layer switch are used carefully by the book and will be here too:

A packet switch is any device that transfers a packet from an input link interface to an output link interface, based on header field values.
A router decides using fields in the network-layer datagram. Routers are layer-3 devices.
A link-layer switch decides using fields in the link-layer frame. Switches are layer-2 devices, and chapter 6 covers them.

Because this chapter is about the network layer, it mostly says router. Section 4.4 deliberately switches to packet switch, because the devices there match on fields from three layers at once and are honestly neither.

Check yourself

0 of 6 answered

1.In Figure 4.1, why are the routers drawn with only three layers?
2.Which pair correctly separates forwarding from routing?
3.predictStep through Figure 4.2 and then Figure 4.3. What is different below the dashed line?
Look at the table, the lookup and the output interface in both.
4.The book imagines a network where humans configure every forwarding table by hand. What is the point of that thought experiment?
5.What exactly does the Internet's best-effort service guarantee?
6.predictATM offered guaranteed delivery, bounded delay and guaranteed bandwidth. Why did the Internet's weaker model win?

What to remember

The network layer runs in every host and every router. The application and transport layers run only at the two ends. This is the structural difference that makes chapter 4 and chapter 5 necessary.
Forwarding is local and fast; routing is global and slow. Nanoseconds against seconds, hardware against software. The data plane is forwarding and the control plane is routing, which is why chapter 4 is one and chapter 5 the other.
The Internet’s network layer offers best-effort service: no guarantee of delivery, order, delay or bandwidth. A network delivering nothing would meet the definition. It was enough because capacity was provisioned generously and applications learned to adapt.