IoT Exam Question Bank

Use this for closed-book recall. Answer aloud or on paper before opening the model points. The full source sheets remain in master-cs/raw/iot/.

1. Systems and architecture

What are the connected-things and connected-data views of IoT?

Connected things are networked smart objects cooperating through Internet technologies (M2M). Connected data is the continuously generated, modeled, stored, queried, and analyzed representation of the physical world.
Compare cloud-, edge-, and mesh-based IoT systems.

Cloud: most functions in cloud; edge: gateways take over cloud functions; mesh: devices perform more functions and communicate directly. State latency, autonomy, resource, and complexity trade-offs.
Are gateways always necessary?

No. Direct IP devices can use an IP router. Gateways are useful for protocol translation, buffering, security isolation, and edge processing.

Differentiate PHY and MAC.

PHY transmits bits and provides radio/channel measurements; MAC manages frames, addressing, association, acknowledgements, and medium access.
Compare unslotted and slotted CSMA/CA.

Unslotted has no beacons/slots/GTS and is simple for sparse traffic. Slotted is beacon-synchronized and supports superframes/GTS, useful when coordination, scheduled access, or sleep timing matters.
Why can doubling the backoff window fail to increase one trial’s delay?

The new random draw can be smaller than the old draw. The larger window primarily lowers collision probability; it does not force every individual wait to be longer.
CSMA/CA or frequency hopping: what does each solve?

CSMA/CA chooses when to transmit by sensing/backoff. Frequency hopping chooses the carrier channel over time to avoid interference/improve reliability.

Name the three NB-IoT deployment scenarios.

Standalone in a separate 200 kHz carrier, guard-band in an unused LTE guard band, and in-band in an LTE PRB.
What is the difference between signaling and data channels, and where is paging sent?

Signaling channels carry control/access/radio-management information; data channels carry user traffic. Paging is signaling/control traffic.
How does a UE join and track a cell?

It scans broadcast information, checks joinability and signal quality, selects an acceptable best cell, and attaches. Tracking areas group cells; crossing a TAI boundary causes TAU, and paging searches the last tracking area.
Contrast eDRX and PSM.

eDRX sleeps between scheduled paging windows, so downlink remains possible at opportunities. PSM sleeps indefinitely and cannot be paged until it wakes; both trade energy for reachability/latency.

Why is 6LoWPAN needed?

802.15.4 frames are tiny, IPv6 headers are large, IPv6 requires a 1280-byte link MTU, and constrained address configuration needs adaptation. 6LoWPAN supplies fragmentation, compression, autoconfiguration, and optional mesh headers.
What is the fifth-fragment offset if payload capacity is 72 bytes?

Fragment 5 starts at 4 × 72 = 288 bytes. Offset is in 8-byte units: 288 / 8 = 36.
Does 6LoWPAN fragmentation replace IP fragmentation?

No. It is link-specific fragmentation below IPv6 that enables an IPv6 datagram to cross 802.15.4; it does not make all IP fragmentation concepts disappear.

What do + and # match, and can publishers use them?

+ matches one level; # matches zero or more trailing levels. They are subscription wildcards, not valid publish-topic wildcards.
Why store topic-alias state at the broker?

Later PUBLISH packets carry only the alias. The broker needs the per-connection alias-to-topic mapping to resolve the concrete topic and route the message.
C1 publishes QoS 2; C2 subscribes QoS 0. What QoS is used on each link?

C1-to-broker: QoS 2. Broker-to-C2: QoS 0, because delivery is limited by the subscription: min(2, 0).
What is the difference between retain and Last Will?

Retain stores the latest value for new subscribers; Last Will is published when a client disconnects unexpectedly.
What should a client publish for a shared subscription $share/exam/iot/temp?

Publish to the ordinary topic iot/temp; $share/... is only the subscription filter/group syntax.

Update protocol vs stream compression?

An update protocol sends fewer measurements/updates; stream compression encodes the transmitted data in fewer bytes. They reduce different costs and can be combined.
When is dead reckoning good or bad?

Good for motion that follows a known model, such as constant velocity. Bad for turns, wheel slip, obstacles, or wind because prediction error causes corrections and wasted computation/transmissions.
Give one relation a key-value model does not represent naturally.

Sensor S locatedAt Room R or Container C carriedBy Vessel V; linked data makes entities and relations first-class and queryable.
Compare stream processing and CEP.

Stream processing queries continuous timestamped data with windows, joins, and aggregates. CEP detects patterns of atomic events and derives/acts on complex events using ECA rules.

Pruning vs quantization?

Pruning removes/zeros connections and creates sparsity; quantization lowers numeric precision while normally keeping the dense connections.
Why can magnitude pruning hurt accuracy?

A small-magnitude weight may be critical in context; magnitude alone ignores activations and network paths. Pruning can remove a critical route or disconnect a neuron.
Why binarize a network?

With +1/-1 weights and activations, multiplication can be approximated using XNOR/popcount logic, which is hardware-efficient, at possible accuracy cost.

Not attempted yet. Record date, score, and missed concept links after each closed-book run.