Practice · Virtual Memory: Demand Paging, Replacement and Thrashing

Q1

What happens when a process accesses a page whose valid bit is 0 but whose address is legal?

A The CPU reads from disk directly. B A page fault: the OS loads the page into a frame, updates the page table and restarts the instruction. C The process is terminated with a segmentation fault. D The access returns 0.

Q2

Memory access 100 ns, page-fault service time 8 ms, fault rate p = 0.001. What is the effective access time, approximately?

A 108 ns B 8 ms C 900 ns D 8.1 µs

Q3

With a 100 ns memory access and 8 ms page-fault time, what page-fault rate keeps the slowdown under 10 %?

A p < 0.1 B p < 0.001 C p < 1.25 × 10⁻⁶ (about one fault per 800,000 accesses) D p < 0.01

Q4

What is the difference between a minor and a major page fault?

A There is no difference. B Minor faults kill the process; major faults do not. C Minor faults happen only in the kernel. D A minor fault is resolved without disk I/O (e.g., page already in the page cache or a new zero page); a major fault must read from disk.

Q5

After fork() with copy-on-write, when is a page actually copied?

A Never B When either reads it C Immediately for all pages D When either the parent or the child writes to that shared page

Q6

A 4-page process forks. The child writes page 1, the parent writes page 1, the parent writes page 2, the child writes page 3. How many pages are copied?

A 3 B 1 C 4 D 8

Q7

Reference string 7 0 1 2 0 3 0 4 2 3 0 3 2 1 2 0 1 7 0 1 with 3 frames. How many faults does FIFO produce?

A 15 B 14 C 12 D 9

Q8

Same reference string, 3 frames. How many faults does OPT produce?

A 12 B 9 C 15 D 10

Q9

Why is the optimal (OPT) algorithm not used in real operating systems?

A It suffers from Belady's anomaly. B It requires dirty bits. C It produces too many faults. D It needs to know future references.

Q10

Frames contain pages 7, 0, 1 (loaded in that order). The next reference is 2, and the recent references were … 7, 0, 1, 0, 7. Which page does LRU evict?

A 0 B 1 C 7 D 2

Q11

Why is exact LRU rarely implemented in hardware?

A It needs knowledge of the future. B It requires updating a timestamp or reordering a list on every memory access, which is too costly; reference bits give a cheap approximation. C It is patented. D It suffers from Belady's anomaly.

Q12

In the clock (second-chance) algorithm, the hand points to a page with reference bit 1. What happens?

A The page is written to disk. B The page is evicted. C The algorithm stops. D The bit is cleared and the hand moves to the next page (the page gets a second chance).

Q13

In the enhanced clock algorithm, which class of page is the best victim?

A (reference 1, dirty 0) B (reference 1, dirty 1) C (reference 0, dirty 0) D (reference 0, dirty 1)

Q14

With FIFO, the string 1 2 3 4 1 2 5 1 2 3 4 5 causes 9 faults with 3 frames. How many with 4 frames?

A 10 B 9 C 6 D 8

Q15

Why can LRU never exhibit Belady's anomaly?

A Because it is a stack algorithm: the pages in memory with n frames are always a subset of those with n + 1 frames B Because it knows the future C It can; LRU also has the anomaly D Because it uses reference bits

Q16

62 free frames; process P1 has 10 pages and P2 has 127 pages. With proportional allocation, approximately how many frames does P1 get?

A 10 B 4 C 31 D 57

Q17

What is thrashing?

A A disk failure B Frequent context switches between threads C A state where processes spend more time paging than executing because they lack frames for their working sets D Very high CPU utilization from computation

Q18

During thrashing, a naive long-term scheduler sees low CPU utilization. What does it do, and why is that harmful?

A It admits more processes, which further reduces the frames per process and makes thrashing worse. B Nothing. C It suspends processes, which is harmful. D It increases the clock speed.

Q19

Δ = 4. The last four references of a process are 4, 5, 6, 7 and the last four of another are 13, 14, 15, 16. Only 7 frames are available. What does the working-set policy do?

A Nothing; the processes will adapt. B D = 8 > 7: suspend one process to free its frames and prevent thrashing. C Admit a third process. D Increase Δ to 8.

Q20

What does page-fault frequency (PFF) control do when a process's fault rate is above the upper threshold?

A Give it more frames (or suspend a process if no frames are free) B Lower its CPU priority C Take frames away from it D Kill it

Q21

Why do many AI inference engines mmap model files instead of reading them into memory?

A mmap is required by GPUs. B Reading files is impossible in Python. C Pages are loaded on demand and cached by the OS: startup is fast, only touched pages use RAM, and several processes can share the same physical pages. D mmap compresses the file.

Q22

What is the advantage of zram/zswap (compressed memory) over swapping to disk?

A It avoids page tables. B An evicted page is compressed and kept in RAM, so bringing it back costs microseconds of decompression instead of a disk access. C It removes the need for replacement algorithms. D It increases physical RAM.

Q23

During pre-copy live migration of a VM, which page-table information tells the hypervisor which pages must be re-sent in the next round?

A The valid bit B The execute bit C The page size D The dirty bit (pages modified since the last round)

Q24

vLLM's PagedAttention improves LLM serving mainly by:

A Storing each request's KV cache in fixed-size blocks mapped by a block table (like paging), which avoids fragmentation and lets many more requests share GPU memory B Using the CPU instead of the GPU C Removing attention D Making the model smaller

Q25

Page-fault service time is 100 µs (NVMe SSD) and memory access is 100 ns. With p = 0.001, what is the effective access time, approximately?

A 100 µs B 1 µs C 200 ns D 100 ns

Q26

Linux uses global page replacement. What is a drawback for a latency-sensitive service sharing the machine with a batch job, and how do containers mitigate it?

A None; global replacement isolates processes. B The batch job can steal the service's frames, causing page faults and latency spikes; cgroup memory limits/reservations give each container a bounded share. C The service cannot allocate memory at all. D The kernel disables paging for services.

Q27

Why do serverless platforms use memory snapshots restored with demand paging to reduce cold starts?

A Snapshots make functions run faster after start. B They disable virtual memory. C Initialization (runtime, libraries, model loading) is done once; new instances map the snapshot and fault in only the pages they touch, with copy-on-write for private changes. D They avoid using memory.

Q28 Short answer

Simulate LRU and clock (second chance; new pages get reference bit 1, the hand advances past a newly loaded page) on the reference string 7 0 1 2 0 3 0 4 2 3 0 3 2 1 2 0 1 7 0 1 with 3 frames. Give the number of faults for each, compare with FIFO (15) and OPT (9), and explain the ranking.

Q29 Short answer

EdgeCampus migrates a user's AR service (a VM with 4 GB of memory) between two edge servers as the user walks across campus. The link provides 1 GB/s, and the service dirties memory at 200 MB/s. Using pre-copy, compute the duration of each round until less than 50 MB remains, the total migration time and the downtime. What happens if the dirty rate is 1.2 GB/s?

Q30 Short answer

A server's CPU utilization drops from 70 % to 8 % after the number of worker processes is increased from 8 to 24, while disk utilization rises to 100 % and response times increase 50×. Diagnose the problem, name two Linux metrics or commands that would confirm it, and propose two fixes.