TU Wien:Advanced Internet Computing VU (Dustdar)/Prüfung 2025-01-15

Tuwel Test, Open Book, 35 minutes, max 40 points, random selection out of question pool.

Questions

True/False Questions

• Software-as-a-Service and Platform-as-a-Service are worse in terms of vendor lock-in than Infrastructure-as-a-Service.
• Cloud computing is beneficial when considering the use case of batch computational workloads.
• When network neutrality is not enforced, IoT service providers can team up with network operators so that their services' traffic is preferentially treated.
• Xen and Docker are both examples of OS-Level Virtualization technologies.
• NB-IoT enhances device battery life compared to LoRaWAN and SigFox.
• Multi-access Edge Computing (MEC) is typically characterized by low reliability and volatility.
• Dynamic Neural Networks can improve throughput.
• AI on the edge deals with the question of how we can build systems to efficiently run AI models near the source.
• Software-as-a-Service and Platform-as-a-Service are worse in terms of vendor lock-in than Infrastructure-as-a-Service.
• Internet Service Providers are in favor of network neutrality because this facilitates their cooperation with third-party service providers.
• Knowledge Distillation is exclusively a tool for model compression.
• Split inference is faster when partitioning a neural network closer to the output layer.
• In order to reduce the managerial overhead, it is possible for an IoT service provider to connect LoRaWAN devices to community gateways instead of operating private ones.
• LPWAN offers connectivity with lower throughput compared with other WAN technologies. LoRA is an example of an LPWAN technology.
• Knowledge Distillation describes a particular training objective that may, among other purposes, be used for model compression.
• Service Level Objectives are coarse-grained metrics that are presented to clients by cloud providers.
• Dynamic Neural Networks can decrease throughput.

Single-Choice Questions

• An IoT application is in charge of collecting data from thousands of environmental sensors distributed over a country, performing some initial filtering of the data to detect faulty sensor readings, and storing them for future processing. Where would you host the storage?
  • a) On the cloud
  • b) At the edge
• A traffic monitoring application works as follows: Self-driving cars send their location information to an application, which is used by the transport authority to calibrate traffic light timings across the city. Where would you execute the application?
  • a) On the cloud
  • b) On an edge server
• Consider a smart parking service which works as follows. A camera is capturing images from parking spots and transmits them to a remote server at a rate of 10 images/s. The server performs an image processing task where it identifies if the parking spot is occupied and, if so, it detects the license plate of the respective car. Which networking technology would you select to connect the camera device to the remote server?
  • a) LoRaWAN
  • b) NB-IoT
  • c) 4G
• A cancer research institute wants to build a digital library of DNA (genome) samples from tens of thousands of patients for long-term research usage. This requires running a genome sequencing analysis on each of the samples independently. Assume that each individual sample analysis can be handled by a single large-memory Amazon EC2 VM instance. A sample analysis can last for hours, but can be stopped and resumed arbitrarily. The institute’s primary concern is to complete the sample library at the smallest possible cost. Based on the provided information, which EC2 instance type would you use to implement the described use-case? Select the one that is most applicable.
  • a) On-demand
  • b) Dedicated
  • c) Reserved
  • d) Spot Instance
• According to the ETSI MEC specification, one of the tasks of the MEC Orchestrator is to select appropriate MEC hosts for application instantiation. Assume a MEC operator which has implemented a machine learning-based algorithm for this purpose. This algorithm receives monitoring information about the MEC system state (e.g., host CPU/memory resource availability), as well as information about the expected workload that the MEC application will handle, and uses these as input to a machine learning model which predicts the MEC host that would offer the best latency performance, if the application were instantiated there. This is an example of the following paradigm:
  • a) AI on Edge
  • b) AI Operationalization
  • c) Osmotic Computing
  • d) AI for Edge
• Advantages of public clouds when compared to private clouds are:
  • a) There is minimal or no upfront cost for hardware and software.
  • b) Higher compliance with local legal regulations can be achieved.
  • c) Higher reliability can be achieved.

Multiple-Choice Questions

• A provider of computational resources (e.g., the operator of an IT center offering computational resources to customers) would rather provide virtualized resources like Virtual Machines instead of offering Physical Machines for reasons of (select all statements below that apply):
  • Backend parallelization
  • Customer security
  • Consolidation for energy consumption
  • Multitenance

• Which of the following statements are correct? Select all statements that apply.
  • a) When executing a ML model on an IoT device, applying quantization on the model can make inference faster at the expense of memory consumption.
  • b) Unstructured pruning can modify the architecture of the neural network it is applied to.
  • c) Applying unstructured pruning leads to a ML model that is more compressible than the original.
  • d) A disadvantage of artificial bottleneck injection is that it requires to change and retrain an existing ML model.
• Select all methods used for model compression. (A correct answer includes all the correct statements and only them. A wrong answer gets zero points.)
  • a) Structured Pruning
  • b) Transformers and Self-Attention Layers
  • c) Unstructured Pruning
  • d) Activation Functions
  • e) Dark Knowledge-Based Methods
  • f) Dropout Regularization
  • g) Convolutional Neural Networks
  • h) Image Quantization
  • i) Structured Pruning
  • j) Network Quantization

Questions with Calculations

Question 1:

You are the provider of a large social networking service. One of the many functions of this service is the generation of JSON Web Tokens, i.e., cryptographically-signed objects that can be used, e.g., for authenticating user requests to various API endpoints of your service. This function requires 30 MB RAM and is invoked 1 million times per month. Each time it is invoked, it runs for 0.1 s. As a service provider, you have the following two options to host this function:

• Option 1: Lease an on-demand virtual machine from an IaaS provider and host your function there. The cheapest offering which can fit your memory and other requirements costs 0.1 EUR/hour. There is no free tier.
• Option 2: Deploy your function at a FaaS provider. In this case, the following charges apply:
  • Compute charges: 0.00001 EUR per MB-s
  • Request charges: 10 EUR per million requests
  • There is a free tier of 1 000 000 MB-s and 500 000 requests per month.

You need to select the option with the minimum monthly cost. (Assumption: A month has 30 days.)

• a) Which of the two options has the minimum monthly cost?
• b) What is the minimum monthly cost? Please enter a numerical value.

Question 2:

Assume a scenario where 3 mobile devices participate in a federated learning task. Before a given training round, the server selects a subset of these devices, which will then perform training on their local data and submit model updates to the server. It is possible that a selected device fails to submit a report at the end of the round (e.g., it may run out of battery during the process, lose its connectivity, etc.). For each device, the probability that it successfully delivers its report in a round is given below:

• Device 1: p1 = 0.1
• Device 2: p2 = 0.6
• Device 3: p3 = 0.4

At the same time, each device has a fixed cost to participate in a round. If selected, the cost of each device is given below:

• Device 1: c1 = 10
• Device 2: c2 = 2
• Device 3: c3 = 20

The server needs to select the subset of devices to participate in the round which minimizes the total cost, under the constraint that the expected number of reports that will be received by the server is at least E = 1. (The total cost is defined as the sum of the costs of all devices that participate in a round. If a device is not selected, it does not contribute to the overall cost.)

• (a) What is the optimal subset of devices that the server should select? Provide your answer in the following format: (X,Y,...). For example of the server selects devices 1 and 3, your answer should be (1,3), if the server selects device 2 your answer should be (2) etc. If there is no feasible solution, your answer should be (). This format is strict: Do not use whitespace characters in your answer; keep the parentheses.
• (b) What is the cost of the optimal solution? Provide a numerical value. If there is no feasible solution, answer with 0.
• (c) What is the expected number of reports that the server will receive in this round under the optimal solution? Provide a numerical value. If there is no feasible solution, answer with 0.

Question 3:

The figure shows a redundant cloud deployment for a web application. The application can function properly if the load balancer is working, as well as at least one web server and at least one database server.

The figure shows 1 load balancer, 2 web servers and 3 database servers.

Assume that the individual component availabilities are as given below:

• availability(load_balancer) = 0.999
• availability(web_server) = 0.9
• availability(database) = 0.8

What is the overall availability (AT) of the system?

• a) AT <= 0.95 and AT >=0.91
• b) AT > 0.95
• c) AT < 0.91

What is the minimum number of web server replicas we need to add to the above configuration to achieve an availability of "two nines"?

• We need to add at least two more web servers.
• It is not possible to achieve two-nines availability by adding web servers.
• We need to add at least three more web servers.
• We need to add at least one more web server.
• No need to add any web servers.
• The system's availability is already over two nines.

Question 4:

A service provider wishes to deploy an application in the cloud. In order to improve availability, it rents multiple VMs from a cloud provider and each VM hosts a single replica application instance. Client requests are transparently served by any of these instances. Each VM costs the service provider a fixed amount of money, so the total cost of the deployment is linear in the number of VMs leased. Assume that this cost is given by function c(x) = x, where x is the number of VMs leased.

Each application instance fails with probability p = 0.3. The service as a whole is considered available when at least one instance is up and running.

The service provider wishes to maximize service availability (given by function a(x) -- you need to calculate this) while minimizing cost. However, these objectives are conflicting, so the service provider comes up with the following cost function that takes into account both objectives: f(x) = c(x) - w*a(x), where w = 10000. The purpose of the service provider is to create a deployment that minimizes this cost function.

What is the optimal number of VMs that the service provider should lease, i.e., the one that minimizes cost function f? Please enter an integer value (if you have calculated a real value, round it to the closest integer, e.g., 5.1 -> 5, 5.9-> 6).

Hints:

• Instead of brute-forcing the solution, you can calculate the derivative of f(x) to get the nubmer of VMs that result in the minimum cost.

Question 5:

What is the output of the ReLU function if it is given value -10 as input? Please provide a single numerical value.