22. Swaping: Policies

Transcription

1 22. Swaing: Policies Oerating System: Three Easy Pieces 1

2 Beyond Physical Memory: Policies Memory ressure forces the OS to start aging out ages to make room for actively-used ages. Deciding which age to evict is encasulated within the relacement olicy of the OS. 2

3 Cache Management Goal in icking a relacement olicy for this cache is to minimize the number of cache misses. The number of cache hits and misses let us calculate the average memory access time(amat). AMAT = P &'( T * + (P *'-- T. ) Arguement T * T. P &'( P *'-- Meaning The cost of accessing memory The cost of accessing disk The robability of finding the data item in the cache(a hit) The robability of not finding the data in the cache(a miss) 3

4 The Otimal Relacement Policy Leads to the fewest number of misses overall w Relaces the age that will be accessed furthest in the future w Resulting in the fewest-ossible cache misses Serve only as a comarison oint, to know how close we are to erfect 4

5 Tracing the Otimal Policy Reference Row Access Hit/Miss? Evict Resulting Cache State 0 Miss 0 1 Miss 0,1 2 Miss 0,1,2 0 Hit 0,1,2 1 Hit 0,1,2 3 Miss 2 0,1,3 0 Hit 0,1,3 3 Hit 0,1,3 1 Hit 0,1,3 2 Miss 3 0,1,2 1 Hit 0,1,2 Comulsory Caacity Conflict Hit rate is &'(- &'(-0*'--1- = 54. 6% Future is not known. 5

6 A Simle Policy: FIFO Pages were laced in a queue when they enter the system. When a relacement occurs, the age on the tail of the queue(the First-in ages) is evicted. w It is simle to imlement, but can t determine the imortance of blocks. 6

7 Tracing the FIFIO Policy Reference Row Access Hit/Miss? Evict Resulting Cache State 0 Miss 0 1 Miss 0,1 2 Miss 0,1,2 0 Hit 0,1,2 1 Hit 0,1,2 3 Miss 0 1,2,3 0 Miss 1 2,3,0 3 Hit 2,3,0 1 Miss 3,0,1 2 Miss 3 0,1,2 1 Hit 0,1,2 Hit rate is &'(- &'(-0*'--1- = 36. 4% Even though age 0 had been accessed a number of times, FIFO still kicks it out. 7

8 BELADY S ANOMALY We would exect the cache hit rate to increase when the cache gets larger. But in this case, with FIFO, it gets worse. Reference Row Page Fault Count Page Frame Count 8

9 Another Simle Policy: Random Picks a random age to relace under memory ressure. w It doesn t really try to be too intelligent in icking which blocks to evict. w Random does deends entirely uon how lucky Random gets in its choice. Access Hit/Miss? Evict Resulting Cache State 0 Miss 0 1 Miss 0,1 2 Miss 0,1,2 0 Hit 0,1,2 1 Hit 0,1,2 3 Miss 0 1,2,3 0 Miss 1 2,3,0 3 Hit 2,3,0 1 Miss 3 2,0,1 2 Hit 2,0,1 1 Hit 2,0,1 9

10 Random Performance Sometimes, Random is as good as otimal, achieving 6 hits on the examle trace ( trials, i.e. seeds) Frequency Number of Hits Random Performance over 10,000 Trials 10

11 Using History Lean on the ast and use history. w Two tye of historical information. Historical Information recency frequency Meaning The more recently a age has been accessed, the more likely it will be accessed again If a age has been accessed many times, It should not be relaced as it clearly has some value Algorithms LRU LFU Choose a candidate to evict based on: w Less Recently Used (LRU) w Less Frequently used (LFU) Chose a candidate to stay w Most recently used (MRU) / Most frequently used (MFU) w Don t work well in this context (mostly is == random) 11

12 Using History : LRU Relaces the least-recently-used age. Reference Row Access Hit/Miss? Evict Resulting Cache State 0 Miss 0 1 Miss 0,1 2 Miss 0,1,2 0 Hit 1,2,0 1 Hit 2,0,1 3 Miss 2 0,1,3 0 Hit 1,3,0 3 Hit 1,0,3 1 Hit 0,3,1 2 Miss 0 3,1,2 1 Hit 3,2,1 12

13 Workload Examle : The No-Locality Workload Each reference is to a random age within the set of accessed ages. w Workload accesses 100 unique ages over time. w Choosing the next age to refer to at random The No-Locality Workload 100% 80% Hit Rate 60% 40% OPT LRU FIFO RAND When the cache is large enough to fit the entire workload, it also doesn t matter which olicy you use. 20% Cache Size (Blocks) 13

14 Workload Examle : The Workload Exhibits locality: 80% of the references are made to 20% of the ages The remaining 20% of the references are made to the remaining 80% of the ages. The Workload 100% Hit Rate 80% 60% 40% OPT LRU FIFO RAND LRU is more likely to hold onto the hot ages. 20% Cache Size (Blocks) 14

15 Workload Examle : The Looing Sequential Refer to 50 ages in sequence. w Starting at 0, then 1, u to age 49, and then we Loo, reeating those accesses, for total of 10,000 accesses to 50 unique ages. The Looing-Sequential Workload 100% 80% Hit Rate 60% 40% OPT LRU FIFO RAND 20% Cache Size (Blocks) 15

16 Imlementing Historical Algorithms To kee track of which ages have been least-and-recently used, the system has to do some accounting work on every memory reference. w Add a little bit of hardware suort. Full LRU suort could be overwhelming w A 2GB rocess with 4KB-size ages, has 512K ages: LRU is unsustainable Comutational effort of udating accesses and relacing ages Costly TLB accesses, memory overhead, etc 16

17 Aroximating LRU Relax requirements of hardware suort, in the form of a use bit w Whenever a age is referenced, the use bit is set by hardware to 1. w Hardware never clears the bit, though; that is the resonsibility of the OS Clock Algorithm w All ages of the system arranges in a circular list. w A clock hand oints to some articular age to begin with. Any algorithm that clears the use bit eriodically might suffice Bit is udated on TLB misses 17

18 Clock Algorithm The algorithm continues until it finds a use bit that is set to 0. G H F A E B C D Use bit Meaning 0 Evict the age 1 Clear Use bit and advance hand The Clock age relacement algorithm When swa daemon kicks-in, the age the hand is ointing to is insected. The action taken deends on the Use bit 18

19 Workload with Clock Algorithm Clock algorithm doesn t do as well as erfect LRU, it does better then aroach that don t consider history at all. The Workload 100% 80% Hit Rate 60% 40% OPT LRU Clock FIFO RAND 20% Cache Size (Blocks) 19

20 Considering Dirty Pages (in the clock algorithm) Save I/O activity of ossible w Silent evictions are referred over non-silent (i.e. avoid OUT disk oerations) The hardware include a modified bit (a.k.a dirty bit) w Page has been modified and is thus dirty, it must be written back to disk to evict it. w Page has not been modified, the eviction is free. 20

21 Page Selection Policy The OS has to decide when to bring a age into memory. Presents the OS with some different otions. 21

22 Prefetching The OS guess that a age is about to be used, and thus bring it in ahead of time. Page 1 is brought into memory Page 5 Page 4 Page 3 Physical Memory Page n Most OS do On Demand Page 4 Page 3 Page 2 Page 1 Secondary Storage Page 2 likely soon be accessed and thus should be brought into memory too 22

23 Clustering, Grouing Collect a number of ending writes together in memory and write them to disk in one write. w Perform a single large write more efficiently than many small ones. Pending writes Page 1 Page 2 Page 3 Page 4 Page 5 Page n write in one write Physical Memory Page 4 Page 3 Page 2 Page 1 Secondary Storage 23

24 Thrashing Memory is oversubscribed and the memory demands of the set of running rocesses exceeds the available hysical memory. w Decide not to run a subset of rocesses. w Reduced set of rocesses working sets fit in memory. w Linux take : OOM (out-of-memory killer) CPU Utilization Trashing Degree of multirogramming 24

25 Disclaimer: This lecture slide set is used in AOS course at University of Cantabria by V.Puente. Was initially develoed for Oerating System course in Comuter Science Det. at Hanyang University. This lecture slide set is for OSTEP book written by Remzi and Andrea Araci- Dusseau (at University of Wisconsin) 25