Packet switching and circuit switching: 22-Elec-B4 Information Technology Networks, May 2015, Q1

Question 1: Packet switching and circuit switching (25 marks)

Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.

Part (a) — the two switching disciplines. The difference is when and how transmission capacity is allocated. A circuit-switched network allocates capacity before any user data flows: a signalling phase establishes an end-to-end path and reserves a physical resource on every link of that path — a frequency band, a repeating time slot, a spreading code or a wavelength — and that resource belongs to the connection until it is torn down, whether or not anything is being sent. The consequences follow directly. The path is fixed, so data arrive in order; the rate is constant, so delay is deterministic and jitter is essentially zero; no addressing information need travel with the data, because the switches already know where the bits from this slot must go; and admission is binary, so a call is either accepted with a guarantee or blocked outright.

A packet-switched network allocates capacity as data arrive. The message is divided into packets, each carrying a header with the destination address and enough control information to be handled independently, and each switch stores an arriving packet, examines the header, queues it for the appropriate outgoing link and forwards it. Links are statistically multiplexed: many flows share the same link, and a flow consumes capacity only in the instants it actually transmits. That gives far higher utilisation for traffic that is not continuously busy, but it also means that when several packets contend for one link they queue, so delay varies with load and buffers can overflow and drop packets. Nothing is reserved, so the network can accept an unlimited number of flows — it degrades gracefully instead of blocking, and any quality guarantee must be manufactured on top by scheduling and admission control rather than being inherent in the switching fabric.

Part (b) — why LTE abandoned circuit switching. GSM, and UMTS after it, ran two core networks in parallel: a circuit-switched domain built around the mobile switching centre, which carried voice as 64 kbit/s (later compressed) circuits inherited from the fixed telephone network, and a packet-switched domain (GPRS, then the UMTS packet core) bolted on later to carry data. LTE's Evolved Packet Core deletes the circuit-switched half entirely: the radio bearer, the backhaul and the core carry IP packets and nothing else, and voice is provided as an application over IP (VoLTE, using an IMS core) rather than as a network service.

Four forces drove that decision. First, the traffic mix inverted — by the time LTE was specified, data had overtaken voice by orders of magnitude, so dedicating a whole parallel core to the smaller and shrinking of the two services was no longer defensible. Second, statistical multiplexing is where the capacity is: mobile data is extremely bursty, and an OFDMA scheduler that reassigns resource blocks every 1 ms extracts multiplexing gain that a reserved circuit throws away. Third, operating one network is far cheaper than operating two, in capital cost, in spare inventory, in staff skills and in the number of interworking functions that must be tested. Fourth, the technical objection to packetised voice had disappeared: LTE's short transmission-time interval and its QoS class identifiers let a voice bearer be given a guaranteed bit rate and a priority that hold delay and jitter inside the range the vocoder needs, so a packet bearer can now do the one job the circuit was retained for. The residual problem — what happens where LTE coverage ends — was handled by fallback mechanisms (CSFB) to the legacy circuit domain rather than by keeping a circuit domain in LTE itself.

Part (c) — classifying the four methods. Two of these are multiple-access techniques at the physical layer and two are network technologies, and the honest answer distinguishes what the method is from how it is normally deployed.

Item	Most useful for	One-sentence reason
i. CDMA	Circuit switching	A spreading code is assigned to a connection and held for its whole duration, which is a reservation in the code domain exactly as a time slot is one in the time domain — although packet-mode variants exist (EV-DO schedules the shared code channel per packet), the classic IS-95 use is a dedicated code per call.
ii. TCP/IP	Packet switching	IP is a connectionless packet-forwarding protocol in which every datagram carries its own destination address and is routed independently, and TCP exists precisely to repair the loss and reordering that packet switching introduces — on a circuit there would be nothing for either to do.
iii. TDMA	Circuit switching	The same numbered slot in every frame is reserved for one connection, giving it a fixed rate for the call's duration; dynamic slot assignment can make TDMA serve packets, but the technique itself is a time-domain reservation.
iv. Ethernet	Packet switching	Ethernet transports self-addressed frames that contend for the medium or for a switch's output port with no reservation of any kind, so its frames queue and can be discarded exactly like the packets of any store-and-forward network.

Part (d) — stream features that favour packet switching. (i) Burstiness — a high peak-to-mean ratio. A source whose peak rate greatly exceeds its average rate (a web session, a database query, a machine sending telemetry once a minute) would hold a circuit sized for its peak while transmitting nothing most of the time, whereas statistical multiplexing lets the idle capacity of one bursty source carry another's burst, and the multiplexing gain grows with the number of sources sharing the link. (ii) Elasticity — no fixed rate requirement. A stream that is content to finish sooner or later rather than at one particular rate (file transfer, e-mail, software update, backup) can absorb whatever capacity happens to be spare and back off when the network is congested, which is a behaviour a reserved circuit cannot express at all. A third feature worth a mark if the first two are exhausted is short duration: for a transaction of a few packets, circuit set-up and tear-down signalling would cost more time than the data transfer itself.

Part (e) — stream features that favour circuit switching. (i) A constant bit rate sustained for a long time. A stream that genuinely occupies its rate continuously — a toll-quality voice call, a studio video feed, a leased inter-exchange trunk — keeps its reservation fully utilised, so circuit switching wastes nothing while avoiding per-packet header overhead and per-packet forwarding work entirely. (ii) A hard delay and jitter constraint, or a required guarantee. A stream that must arrive with bounded latency and negligible variation (real-time process control, protective relaying between substations, legacy telephony) gets that deterministically from a reserved circuit, because there is no queue anywhere to introduce variable delay and no possibility of congestive loss, whereas a packet network can only approximate the guarantee with admission control and priority scheduling and can never quite eliminate the queueing variance.

Topic. Resource allocation as the defining difference between switching disciplines.

Key relations. Circuit-switched capacity for a link of rate $C$ carrying calls of rate $r$: $N_{\text{max}} = \lfloor C/r \rfloor$ regardless of activity. Packet-switched sharing of the same link by $N$ sources of peak $r$ and activity factor $p$: the link is adequate whenever $\Pr\{\text{simultaneously active} > C/r\}$ is acceptably small, i.e. $\sum_{k > C/r} \binom{N}{k} p^{k}(1-p)^{N-k} \le \varepsilon$. Multiplexing gain $= Np/(C/r)$.

Why this works. A circuit is a promise made in advance and therefore paid for in advance; a packet is a request made at the moment of need and therefore paid for only when it is granted. Every downstream property follows from that one difference. Because a circuit's resources are removed from the pool at set-up, its delay cannot depend on anyone else's behaviour, which is exactly what a real-time stream wants and exactly what a bursty stream wastes. Because a packet network commits nothing, it can serve a population whose aggregate peak far exceeds the link rate, at the price of queueing delay that grows without bound as offered load approaches capacity. The industry's convergence on packet switching — of which LTE is one instance — happened when scheduling became fast enough to synthesise the circuit's delay guarantee on demand, at which point there was no longer a service the circuit alone could provide.

Common pitfall. Treating "packet switching" as synonymous with "connectionless" and "circuit switching" with "connection-oriented". They are independent axes: virtual-circuit packet networks (X.25, ATM, MPLS) are connection-oriented but still switch packets and still queue. A second pitfall is asserting that packet switching is simply more efficient; for a continuously-busy constant-rate stream it is slightly less efficient, because the headers and the per-packet processing buy nothing.

Source. Leon-Garcia and Widjaja, Communication Networks, 2nd ed., ch. 7 (packet-switching networks) and ch. 4 (circuit switching and multiplexing); Tanenbaum and Wetherall, Computer Networks, 5th ed., §2.5.4; 3GPP TS 23.401 §4 for the all-IP Evolved Packet Core.

Question 1 of 5: Packet switching and circuit switching

Question 1: Packet switching and circuit switching (25 marks)