Question 2 of 5: Layered architecture (25 marks)

Question 2 — Layered architecture (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.

Given. The ISO/IEC 7498-1 OSI reference model, with seven layers, and eight named mechanisms to be placed within it. The question explicitly rewards brevity in part (a), so each layer gets one sentence and no more. Find. The seven layers named and described; the layer or layers hosting each of the eight items; and a reasoned account of when strict layering costs more than it saves.

Part (a) — The seven layers

Part (a) — one sentence per layer, from the top down, which is the order in which data is handed downward at the sender:

The mnemonic worth carrying into the exam room is All People Seem To Need Data Processing (Application, Presentation, Session, Transport, Network, Data Link, Physical). It is also worth remembering that layers 1–3 are the chained layers — they operate hop by hop and are present in every intermediate node — while layers 4–7 are end-to-end and exist only in the two end systems.

Part (b) — Placing eight mechanisms in the model

Part (b) — the assignments, with the one-line reason that earns the mark:

Two of the eight deserve a comment. Item iv is the one that takes two layers: “Ethernet” as a standard is IEEE 802.3, which spans the MAC sublayer of layer 2 and the physical layer, so naming only layer 2 answers half the question — the question's own phrasing (“or layers, if more than one”) is the hint. Item vi is a reminder that the OSI model and real practice diverge: encryption belongs to the presentation layer in OSI, but in the TCP/IP stack there is no presentation layer, and TLS sits between the application and TCP while IPsec sits at layer 3. The OSI answer is layer 6; saying so and adding the practical note is the complete answer.

Part (c) — When layering is a disadvantage

Layering buys modularity: each layer can be reimplemented without disturbing its neighbours, provided the service interface is unchanged. The price is paid in four distinct currencies.

Duplicated function. The same job is often done more than once because no layer can rely on the one below. Error detection appears in the Ethernet CRC at layer 2, in the IP header checksum at layer 3, in the TCP checksum at layer 4 and again in an application-level digest at layer 7; flow control appears at both layer 2 and layer 4. Each repetition costs processing on every packet.

Header overhead. Each layer prepends its own header, and on small payloads the headers dominate. A G.729 voice packet carries 20 ms of speech at 8 kbit/s, that is 20 bytes of payload, under an RTP header of 12 bytes, a UDP header of 8 and an IPv4 header of 20:

$$\text{overhead} = \frac{12 + 8 + 20}{12 + 8 + 20 + 20} = \frac{40}{60} = 66.7\%$$

so the 8 kbit/s codec occupies $60 \times 8 / 0.020 = 24$ kbit/s on the wire — three times the codec rate, before the Ethernet header is added. This is precisely why header-compression schemes such as RoHC exist, and every one of them is a deliberate layering violation.

Information hiding at the wrong moment. A layer cannot act on information it has been designed not to see, and sometimes that information is decisive. The standard example is TCP over a wireless link: TCP treats every lost segment as evidence of congestion and halves its window, but on a radio link most loss is bit corruption, so the correct response would have been to retransmit at the same rate. TCP cannot tell the two apart because layer 2 does not tell it. The same blindness prevents an application from expressing a delay requirement that the network layer could act on, which is why quality-of-service mechanisms (DiffServ, MPLS traffic engineering) all reach across layer boundaries.

Ossification. A layer boundary drawn in the wrong place is expensive to move. The OSI session and presentation layers turned out to be nearly empty and are simply absent from the TCP/IP stack; conversely NAT, firewalls, load balancers and MPLS are all widely deployed and all violate strict layering, because the architecture had no clean way to express what operators needed. Add the per-boundary cost in latency and data copying, and the honest summary is that layering is the right default but a poor absolute: it should be relaxed deliberately, by a documented cross-layer interface, rather than accidentally.