Layering and switching

A network task (move a file from an application on one host to an application on another) is split into layers. Each layer offers a service to the layer above, uses the service of the layer below, and talks to its peer layer on the other machine through a protocol. Only the service interface is visible between layers, so one layer can be replaced (say, Wi-Fi for Ethernet) without touching the rest.

Data moves down the stack at the sender, each layer adding its own header (encapsulation), across the medium, and up the stack at the receiver, each layer removing the header meant for it (decapsulation).

Convention: the TCP/IP model is drawn with four layers, or with five if the link layer is split into data link and physical. OSI is a reference model with seven layers; the Internet actually runs TCP/IP. Addressing at each level: the application layer uses names such as domain names or e-mail addresses, the transport layer uses port numbers, the network layer uses IP addresses, the data link layer uses MAC addresses.

At the sender, the application data gets a transport header to form a segment, the segment gets a network header to form a packet, and the packet gets a data link header (and usually a trailer carrying the CRC) to form a frame; the physical layer sends the frame as bits and adds no header. The receiver strips the headers in reverse order. What a layer receives from the layer above is its SDU (its payload); the SDU plus that layer's header is the layer's PDU.

In circuit switching a dedicated path is set up between the two hosts before any data is sent. Along the path each link carries the call in its own share of capacity: a frequency band (FDM) or a repeating time slot (TDM). The resources stay reserved for the whole call, used or not, and are released at teardown.

If a link of rate $R$ is shared by $n$ circuits by TDM or FDM, each circuit ideally (ignoring guard bands and framing overhead) gets $R/n$. Because the path is dedicated, data is not stored and forwarded at each switch: only the propagation delay accumulates along the path, and the message is transmitted once.

Strengths: guaranteed rate and constant delay, in-order delivery, no per-packet header. Weakness: idle capacity is wasted during silences, set-up time is paid even for short transfers, and a call can be blocked if no channel is free.

In packet switching the message is cut into packets that share the links on demand (statistical multiplexing). Every switch stores a whole packet, then forwards it (store-and-forward; assume this unless the question says cut-through); a packet waits in a queue if the outgoing link is busy. No capacity is reserved, so queueing delay and packet loss are possible.

Datagram (connectionless) switching. No set-up. Each packet carries the full destination address, and each router looks it up in its forwarding table independently; routers keep no per-connection state. Packets of one message may take different routes, arrive out of order, or be lost. If a router fails, packets queued at it or sent to it are lost, and later packets can be routed around it once the forwarding tables are updated.

Virtual-circuit switching (connection-oriented). A set-up phase fixes one route and assigns a virtual-circuit identifier (VCI) on every link of it. Each packet carries only a short VCI, not the full address; each switch looks up (incoming port, incoming VCI) in its table, which gives the outgoing port and the outgoing VCI, and swaps in the outgoing VCI. All packets follow the same path and arrive in order. Unlike circuit switching, the path is logical: capacity may or may not be reserved, and packets are still stored and forwarded and can queue. X.25, Frame Relay and ATM are virtual-circuit networks.

Per link, the delay of one packet is $d_{\text{nodal}}=d_{\text{proc}}+d_{\text{queue}}+d_{\text{trans}}+d_{\text{prop}}$: transmission delay $L/R$ (pushing all $L$ bits onto a link of rate $R$), propagation delay $d/s$ (signal travel time over distance $d$ at the speed $s$ given), plus processing and queueing delay, which are zero unless the question gives them. Convention: $1\text{Mbps}=10^6$ bits per second and $1\text{kbps}=10^3$ bits per second for rates; sizes use the unit stated, with 1 byte = 8 bits. A path with $k$ links has $k-1$ intermediate switches or routers.

Why more packets help: with one big packet (message switching) every hop waits for the full message, giving $kM/R$ when headers are ignored. Splitting lets hops work in parallel, giving about $(k+n-1)p/R$. Too many packets, however, mean more headers and more bits to send, so there is an optimum.