Practice · Virtualization, Containers and Cloud Resource Management

Q1

Which is a type 1 hypervisor?

A KVM / VMware ESXi / Xen B VMware Workstation C Docker D VirtualBox on Windows

Q2

In trap-and-emulate virtualization, what happens when the guest kernel executes a privileged instruction?

A It executes directly on the real hardware. B The instruction is ignored. C The CPU traps to the hypervisor, which emulates the instruction's effect on the VM's virtual state. D The guest crashes.

Q3

What does paravirtualization change?

A Nothing B The hardware C The applications D The guest OS (or its drivers) is modified to call the hypervisor directly via hypercalls, avoiding costly traps and emulation

Q4

With nested paging and 4-level page tables in both the guest and the host, how many memory references can a TLB miss require in the worst case?

A 8 B 4 C 16 D 24

Q5

A host has 64 physical cores and runs VMs with 256 vCPUs in total. What is the CPU overcommit ratio, and why does it usually work?

A 4:1; the hypervisor creates extra cores B 1:4; VMs are slow anyway C 4:1; most VMs do not use all their vCPUs at the same time D 256:1; vCPUs are free

Q6

What is memory ballooning?

A A memory leak B Compressing the VM image C A driver inside the guest allocates ("inflates") memory and hands those pages back to the hypervisor, letting the guest OS choose what to evict D Growing the host's physical memory

Q7

Which kernel features create a Linux container?

A A hypervisor and a guest kernel B Virtual memory only C Namespaces (isolated views: PIDs, network, mounts…) and cgroups (resource limits) D Only chroot

Q8

Why can a vulnerability in the Linux kernel be more dangerous for containers than for VMs?

A Containers do not use the kernel. B All containers share the host kernel; exploiting it can break out of every container, whereas each VM has its own kernel behind a hypervisor. C It is not more dangerous. D VMs have no kernel.

Q9

Ten containers run from the same 500 MB image. Approximately how much disk space do the image layers use?

A About 500 MB, plus a small writable layer per container B 0 MB C 50 MB D 5 GB

Q10

In Kubernetes, which value does the scheduler use to decide whether a pod fits on a node?

A The pod's requests B The image size C The pod's actual usage D The pod's limits

Q11

A pod has requests equal to limits for both CPU and memory. What is its QoS class?

A Critical B Burstable C BestEffort D Guaranteed

Q12

A container has cpu.max = 20000 100000 (0.2 CPU). A single-threaded request needs 50 ms of CPU and arrives at the start of a period. When does it finish?

A 210 ms B 120 ms C 50 ms D 250 ms

Q13

A 4-thread service has a CPU limit of 1 CPU (100 ms per 100 ms period). A request uses all 4 threads for 50 ms each. When does it finish?

A 50 ms B 125 ms C 200 ms D 100 ms

Q14

Hosts have 10 CPUs. VMs need 2, 5, 4, 7, 1, 3 and 8 CPUs (in that order). How many hosts does first fit use?

A 4 B 5 C 7 D 3

Q15

Why does sorting VMs in decreasing size before first fit usually help?

A Sorting reduces the number of VMs. B It makes the algorithm optimal in all cases. C Large items are placed while there is still room for them; small items then fill the gaps. D Small VMs are more important.

Q16

Power model: 100 W idle, 250 W at full load, linear in CPU utilization. What does a host at 60 % utilization draw?

A 210 W B 150 W C 250 W D 190 W

Q17

What is the main risk of aggressive consolidation (packing VMs tightly and switching hosts off)?

A Higher energy use B SLA violations when demand grows: hosts become overloaded before VMs can be migrated or new hosts powered on C VMs become larger D Nothing

Q18

HPA with a target of 60 % CPU. 3 replicas run at 100 %. How many replicas does the HPA request?

A 6 B 5 C 3 D 4

Q19

HPA target 60 %; 8 replicas run at 25 %. What does the HPA request (minimum 1)?

A 1 B 4 C 2 D 8

Q20

Why do autoscaled services usually target 50–70 % CPU instead of 95 %?

A Because Kubernetes forbids 95 % B To reduce the number of pods C New pods and nodes take seconds to minutes to start; headroom absorbs load increases in the meantime (and queueing delay grows quickly near 100 %) D To waste money

Q21

What is a cold start in serverless computing?

A A function that runs at night B A server booting after a power failure C A request that finds no warm instance, so the platform must create one (start the sandbox, load the runtime and code) before executing it D A request with a small payload

Q22

Increasing a serverless platform's keep-alive time has which effect?

A Fewer cold starts but more idle memory kept warm (cost for the provider) B Longer execution time C More cold starts and less idle memory D No effect

Q23

Two tenants' VMs on the same host both have CPU and memory limits, yet one tenant's latency doubles when the other runs a memory-intensive job. What is this called?

A Thrashing B A noisy-neighbour effect: contention for shared caches and memory bandwidth that CPU/memory limits do not isolate C A deadlock D A cold start

Q24

What did Intel VT-x and AMD-V add to x86 processors?

A A hardware guest mode in which privileged operations of a guest trap cleanly to the hypervisor, removing the need for binary translation B More cores C Larger caches D GPU support

Q25

Why do cloud GPU instances usually use device pass-through (or SR-IOV) instead of emulated devices?

A GPUs cannot be virtualized at all. B Emulated GPUs are more secure. C Pass-through is cheaper for the provider. D The VM accesses the physical device (or a hardware slice of it) directly, reaching near-native performance, which emulation cannot provide.

Q26

Under memory pressure on a Kubernetes node, which pods are evicted first?

A The oldest pods B BestEffort (no requests or limits) C Burstable using less than their requests D Guaranteed

Q27

How do micro-VMs such as AWS Firecracker combine the advantages of VMs and containers?

A They share the host kernel like containers. B They are containers renamed. C They run without a kernel. D A minimal hypervisor with very few emulated devices boots a small guest in ~100 ms, giving hardware-level isolation with near-container start time and overhead.

Q28 Short answer

Place VMs A (2 CPU, 4 GB), B (5, 8), C (4, 4), D (7, 8), E (1, 2), F (3, 4), G (8, 4) on identical hosts with 10 CPUs and 32 GB using first fit (in the given order) and first fit decreasing (sorted by dominant normalized demand, largest first). For each, list the hosts' contents, the number of hosts and the total power (100 W idle + 150 W × CPU utilization per powered host). What is the lower bound on the number of hosts?

Q29 Short answer

Simulate the HPA (target 60 %, min 1, max 10, each pod serves 100 requests/s at 100 % CPU, starting with 2 replicas) for the load sequence 80, 150, 300, 450, 450, 200, 100, 60 requests/s (the new replica count applies at the next step). Give utilization and desired replicas at each step. What weaknesses of reactive autoscaling does this show?

Q30 Short answer

EdgeCampus must choose between (a) one VM per student research project on the edge GPU servers, (b) containers in Kubernetes on bare-metal edge servers, and (c) containers inside per-department VMs. Compare them for startup time, density, isolation between students, GPU sharing and operations effort, and recommend one.