Practice · Edge Computing and Computation Offloading

Q1

Which is not a main motivation for edge computing?

A Lower latency for interactive applications B Reducing the bandwidth sent to the cloud C Keeping sensitive data local D Edge servers have more computing power than the cloud

Q2

A task needs 0.2 Gcycles; the phone runs at 2 GHz. How long does local execution take?

A 400 ms B 100 ms C 40 ms D 10 ms

Q3

Using Shannon's formula, what is the uplink rate with a bandwidth of 20 MHz and an SNR of 10 dB?

A About 69 Mb/s B About 200 Mb/s C About 20 Mb/s D About 10 Mb/s

Q4

Uploading 200 KB (1.6 Mb) at 69.2 Mb/s takes about:

A 2.3 ms B 23 ms C 230 ms D 2.9 ms

Q5

Task: 0.2 Gcycles, 1.6 Mb input. Edge: 20 GHz, RTT 5 ms, upload 23.1 ms. What is the total edge execution time?

A 10 ms B 28.1 ms C 38.1 ms D 100 ms

Q6

In the model of this chapter, what energy does the phone spend when it offloads a task?

A The energy of the server's CPU B The transmission energy P_tx × D / R (plus possibly idle waiting power) C Zero D The same as computing locally

Q7

Same task (C = 0.2 Gcycles, D = 1.6 Mb, local 100 ms); edge 20 GHz with RTT 5 ms. Above which uplink rate is offloading to the edge faster than local execution?

A 1.6 Mb/s B 44.4 Mb/s C 18.8 Mb/s D 16 Mb/s

Q8

A task needs 0.05 Gcycles but has a 2 MB input. The phone computes it in 25 ms. The edge is 20 GHz away with RTT 5 ms. Why is the break-even rate so high (≈ 914 Mb/s)?

A Because RTT is large B Because the edge is slow C The task has little computation per bit of input: almost all remote time is spent uploading D Because of DVFS

Q9

Offloading to a cloud 60 ms away can never be faster than local execution when:

A The input is large B The cloud is idle C The local execution time is smaller than RTT + remote computation time D The phone battery is full

Q10

With DVFS, dynamic energy per task is E = κ C f². If a phone halves its CPU frequency, what happens to the task's time and energy?

A Time unchanged, energy ÷2 B Time ×2, energy ÷4 C Time ÷2, energy ÷2 D Time ×2, energy ×2

Q11

Why can the optimal DNN split point change while the user walks across campus?

A Because the phone's CPU gets faster B It cannot change C Because the model changes D Because the uplink rate changes: with a fast network more layers go to the edge; with a slow one more run on the device

Q12

Why is it often good to split a DNN after a pooling layer?

A Pooling greatly reduces the size of the intermediate data, so less must be transmitted. B Pooling increases accuracy. C Pooling layers are the slowest. D Pooling cannot run on a phone.

Q13

Six identical users could each offload to a shared edge server. Individually, offloading is better while the server is shared by at most 5 users. What is the Nash equilibrium?

A 3 offload B Nobody offloads C All 6 offload D 5 offload and 1 computes locally

Q14

In the same example, if all 6 users offloaded, each would wait 105 ms, while computing locally takes 100 ms. What does this illustrate?

A That offloading is always bad B That RTT is zero C That the cloud is better D That uncoordinated use of a shared resource can make everyone worse off — the reason for admission control, pricing or coordination

Q15

An edge server (M/M/1) serves μ = 100 tasks/s. What is the average time in the system when λ = 80 tasks/s?

A 200 ms B 10 ms C 12.5 ms D 50 ms

Q16

An edge server is close to saturation and AR requests keep arriving. Which reaction is not reasonable?

A Ask phones to reduce frame rate or resolution B Admit all requests and let the queue grow without bound C Redirect some requests to the cloud D Reject requests that cannot meet their deadline

Q17

In the placement policy "EDF order + cheapest resource that meets the deadline", why may placing a hopeless task anyway be harmful?

A It occupies a resource and can push later tasks to more expensive resources or make them miss their deadlines. B It reduces the cost. C It makes EDF optimal. D It is never harmful.

Q18

Offloading a workflow DAG across device, edge and cloud is closest to which problem from earlier chapters?

A Scheduling a DAG on heterogeneous processors with communication costs (HEFT/PEFT) B Deadlock detection C Disk scheduling D Page replacement

Q19

A student using CampusAR walks from building A to building B. Their service runs on A's edge server. What are the options, and what is the trade-off?

A Keep serving from A over the campus network (higher latency) or migrate the service to B's edge (migration cost, brief downtime); the choice depends on how long the user stays and the service's state size. B Always move the service to the cloud. C Always migrate immediately. D There is only one option: restart the app.

Q20

Which technique is designed for making good offloading decisions over the long term under random arrivals and channel conditions, while keeping queues stable?

A FIFO page replacement B RAID 5 C Round robin D Lyapunov optimization (drift-plus-penalty)

Q21

Why is deep reinforcement learning used in recent offloading research?

A It guarantees optimal solutions. B It can learn decision policies for dynamic, partially known environments (changing channels, loads, mobility) where exact models or optimization are too slow or unavailable. C It removes the need for networks. D It needs no data.

Q22

What is federated learning, in placement terms?

A A scheduling algorithm for disks B Training on the devices/edge where the data lives and sending only model updates to be aggregated, reducing data movement and protecting privacy C A file system D Sending all raw data to the cloud for training

Q23

Which orchestrators are designed for small edge nodes with unreliable links?

A FAT32 and ext4 B K3s and KubeEdge C RAID 6 and RAID 10 D FIFO and LRU

Q24

Six identical users share an edge server; with k offloaders each gets 45 + 10k ms remotely, while local execution takes 100 ms. Which number of offloaders minimizes the average latency of all six users?

A 3 B 0 C 5 (the Nash equilibrium) D 6

Q25

Converting an SNR of 3 dB to a linear ratio gives approximately:

A 2 B 30 C 3 D 1.5

Q26

AR applications need roughly 20 ms motion-to-photon latency. Why does this practically exclude a cloud region 60 ms away for per-frame processing?

A Because of DVFS. B Clouds are too expensive. C The network round trip alone exceeds the whole latency budget, regardless of how fast the cloud computes. D Clouds cannot run GPUs.

Q27

Which kind of task is the best candidate for offloading?

A Little computation, large input data B Tasks that must read local sensors every microsecond C Heavy computation on small input data (high cycles per bit), without a very tight latency bound D Tasks with no input and no computation

Q28 Short answer

Compute, for a CampusAR frame (C = 0.2 Gcycles, D = 200 KB, phone 2 GHz, P_compute = 0.9 W, P_tx = 1.3 W), the time and phone energy of local, edge (20 GHz, RTT 5 ms) and cloud (50 GHz, RTT 60 ms) execution with (a) good Wi-Fi: 20 MHz, 10 dB; (b) weak Wi-Fi: 5 MHz, 3 dB. Which option would you choose in each case, and what are the break-even rates?

Q29 Short answer

Explain the best-response dynamics for 6 identical users sharing a 20 GHz edge server (C = 0.2 Gcycles each, local time 100 ms, upload 40 ms, RTT 5 ms, capacity shared equally among offloaders). Give each user's remote time as users join, the equilibrium, the average latency, and compare with "all offload" and "none offload". Suggest one mechanism to improve the outcome.

Q30 Short answer

Formulate, in words and symbols, a research problem for EdgeCampus: N phones generate tasks (C_i, D_i, deadline d_i); there are M edge servers (speeds f_j, capacities) and a cloud (speed f_c, price p per Gcycle). Define the decision variables, the objective and the constraints, say why the problem is hard, and propose one heuristic and one evaluation plan.