Lecture 15: Page Replacement - People

Lecture 15: Page Replacement - People

CS162 Operating Systems and Systems Programming Lecture 15 Page Allocation and Replacement October 21, 2009 Prof. John Kubiatowicz http://inst.eecs.berkeley.edu/~cs162 Review: Demand Paging Mechanisms PTE helps us implement demand paging Valid Page in memory, PTE points at physical page Not Valid Page not in memory; use info in PTE to find it on disk when necessary Suppose user references page with invalid PTE? Memory Management Unit (MMU) traps to OS Resulting trap is a Page Fault What does OS do on a Page Fault?: Choose an old page to replace If old page modified (D=1), write contents back to disk Change its PTE and any cached TLB to be invalid Load new page into memory from disk Update page table entry, invalidate TLB for new entry Continue thread from original faulting location TLB for new page will be loaded when thread continued! While pulling pages off disk for one process, OS runs another process from ready queue 10/21/09 UCB

Fallqueue 2009 SuspendedKubiatowicz process CS162 sits on wait Lec 15.2 Goals for Today Precise Exceptions Page Replacement Policies Clock Algorithm Nth chance algorithm Second-Chance-List Algorithm Page Allocation Policies Working Set/Thrashing Note: Some slides and/or pictures in the following are adapted from slides 2005 Silberschatz, Galvin, and 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.3 Gagne Many slides Gagne. generated from my lecture notes Software-Loaded TLB MIPS/Nachos TLB is loaded by software High TLB hit rateok to trap to software to fill the TLB, even if slower Simpler hardware and added flexibility: software can

maintain translation tables in whatever convenient format How can a process run without hardware TLB fill? Fast path (TLB hit with valid=1): Translation to physical page done by hardware Slow path (TLB hit with valid=0 or TLB miss) Hardware receives a TLB Fault What does OS do on a TLB Fault? Traverse page table to find appropriate PTE If valid=1, load page table entry into TLB, continue thread If valid=0, perform Page Fault detailed previously Continue thread Everything is transparent to the user process: It doesnt know about paging to/from disk It doesnt even know about software TLB handling 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.4 OS Load TLB Faulting Inst 2 Faulting Inst 2 TLB Faults

Faulting Inst 1 User Faulting Inst 1 Transparent Exceptions Fetch page/ Load TLB How to transparently restart faulting instructions? Could we just skip it? No: need to perform load or store after reconnecting physical page Hardware must help out by saving: Faulting instruction and partial state Need to know which instruction caused fault Is single PC sufficient to identify faulting position???? Processor State: sufficient to restart user thread Save/restore registers, stack, etc What if an instruction has side-effects? 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.5 Consider weird things that can happen What if an instruction has side effects?

Options: Unwind side-effects (easy to restart) Finish off side-effects (messy!) Example 1: mov (sp)+,10 What if page fault occurs when write to stack pointer? Did sp get incremented before or after the page fault? Example 2: strcpy (r1), (r2) Source and destination overlap: cant unwind in principle! IBM S/370 and VAX solution: execute twice once read-only What about RISC processors? For instance delayed branches? Example: bne somewhere ld r1,(sp) Precise exception state consists of two PCs: PC and nPC Delayed exceptions: 10/21/09 Example: div r1, r2, r3 ld r1, (sp) What if takes many cycles to discover divide by zero, but load has already caused page fault? Kubiatowicz CS162 UCB Fall 2009 Lec 15.6 Precise

Precise state ofExceptions the machine is preserved as if program executed up to the offending instruction All previous instructions completed Offending instruction and all following instructions act as if they have not even started Same system code will work on different implementations Difficult in the presence of pipelining, out-of-order execution, ... MIPS takes this position Imprecise system software has to figure out what is where and put it all back together Performance goals often lead designers to forsake precise interrupts system software developers, user, markets etc. usually wish they had not done this Modern techniques for out-of-order execution 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.7 and branch prediction help implement precise Steps in Handling a Page Fault 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.8

Demand Paging Example Since Demand Paging like caching, can compute average access time! (Effective Access Time) EAT = Hit Rate x Hit Time + Miss Rate x Miss Time Example: Memory access time = 200 nanoseconds Average page-fault service time = 8 milliseconds Suppose p = Probability of miss, 1-p = Probably of hit Then, we can compute EAT as follows: EAT = (1 p) x 200ns + p x 8 ms = (1 p) x 200ns + p x 8,000,000ns = 200ns + p x 7,999,800ns If one access out of 1,000 causes a page fault, then EAT = 8.2 s: This is a slowdown by a factor of 40! What if want slowdown by less than 10%? 200ns x 1.1 < EAT p < 2.5 x 10-6 This is about 1 page fault in 400000! 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.9 What Factors Lead to Misses? Compulsory Misses: Pages that have never been paged into memory before How might we remove these misses?

Prefetching: loading them into memory before needed Need to predict future somehow! More later. Capacity Misses: Not enough memory. Must somehow increase size. Can we do this? One option: Increase amount of DRAM (not quick fix!) Another option: If multiple processes in memory: adjust percentage of memory allocated to each one! Conflict Misses: Technically, conflict misses dont exist in virtual memory, since it is a fully-associative cache Policy Misses: Caused when pages were in memory, but kicked out prematurely because of the replacement policy How to fix? Better replacement policy 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.10 Page Replacement Policies Why do we care about Replacement Policy? Replacement is an issue with any cache Particularly important with pages The cost of being wrong is high: must go to disk Must keep important pages in memory, not toss them out FIFO (First In, First Out)

Throw out oldest page. Be fair let every page live in memory for same amount of time. Bad, because throws out heavily used pages instead of infrequently used pages MIN (Minimum): Replace page that wont be used for the longest time Great, but cant really know future Makes good comparison case, however RANDOM: Pick random page for every replacement Typical solution for TLBs. Simple hardware Pretty unpredictable makes it hard to make realtime guarantees 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.11 Replacement Policies (Cont) LRU (Least Recently Used): Replace page that hasnt been used for the longest time Programs have locality, so if something not used for a while, unlikely to be used in the near future. Seems like LRU should be a good approximation to MIN. How to implement LRU? Use a list! Head Page 6 Page 7 Page 1

Page 2 Tail (LRU) On each use, remove page from list and place at head LRU page is at tail Problems with this scheme for paging? Need to know immediately when each page used so that can change position in list Many instructions for each hardware access In practice, 10/21/09 people approximate LRU Kubiatowicz CS162 UCB Fall 2009 (more later) Lec 15.12 Administrivia Exam not graded yet Hopefully get it back to you in section on Tuesday Will get solutions up early next week Project II Due by Friday at midnight Good Luck! In the news: Sunday (10/18) First use of On Star to recover stolen vehicle Two guys carjacked in 2009 Chevy Tahoe On Star locates vehicle and while police chasing it, give command to SUV to slow down On Star operator flashed lights first so that police could make

sure that they had correct car Hapless crook tries to escape on foot, falls in swimming pool Is this Big Brother, or a Good Thing??? Related: My OS group talked to researchers from Bosch They make lots of things including car electronics Confirmed that high-end cars have over 70 Engine Control Units ECU for short Processors controlling breaks, fuel injection, navigation, climate Interconnected with CAN (Controller Area Network) Thinking about consolidating functions into smaller number of MultiCore chips (Buthow to meet hard realtime Lec 15.13 10/21/09 Kubiatowicz CS162 UCB Fall 2009 requirements??) In the News: Android Android is the new operating system from Google For Mobile devices Phones Ebook Readers (i.e. B&N) Linux version 2.6.x Java virtual machine and runtime system Lots of media extensions WebKit for browsing Media Libraries Cellular Networking Mobile Systems are the hottest new software stack Ubiquitous Computing Worldwide, more than 1 billion new cell phones purchased/year for last few years

Compare: worldwide number PCs purchased/year ~ 250M 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.14 Example: FIFO Suppose we have 3 page frames, 4 virtual pages, and following reference stream: ABCABDADBCB Consider FIFO Page replacement: Ref: Page: A 1 A 2 3 B C A B D A

D B D B C B C A C B FIFO: 7 faults. When referencing D, replacing A is bad choice, since need A again right away 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.15 Example: MIN Suppose we have the same reference stream: ABCABDADBCB Consider MIN Page replacement: Ref: Page: A 1

A 2 3 B C A B D A D B C B C B C D MIN: 5 faults Where will D be brought in? Look for page not referenced farthest in future. What will LRU do? Same decisions as MIN here, but wont always be

10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.16 true! When will LRU perform badly? Consider the following: A B C D A B C D A B C D LRU Performs as follows (same as FIFO here): Ref: Page: A 1 A 2 B C D A B D B 3 C A C

A C D B C D B D B C A D Every reference is a page fault! Ref: A much B C better: D A B MIN Does C D A

B C D Page: 1 2 3 10/21/09 A B B C C D Kubiatowicz CS162 UCB Fall 2009 Lec 15.17 Graph of Page Faults Versus The Number of Frames One desirable property: When you add memory the miss rate goes down Does this always happen? Seems like it should, right? No: BeLadys anomaly Certain replacement algorithms (FIFO) dont have this obvious property!

10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.18 Adding Memory Doesnt Always Help Fault Rate Does adding memory reduce number of page faults? Yes for LRU and MIN Not necessarily for FIFO! (Called Beladys anomaly) Ref: Page: A 1 A 2 B B A 1 A 4 A B B


C B D D E C Ref: Page: 3 D D 3 2 C E B D After adding memory: C With FIFO, contents can be completely different In contrast, with LRU or MIN, contents of memory with 10/21/09 CS162 UCB Fallwith

2009 X+1 PageLec 15.19 X pages are aKubiatowicz subset of contents Implementing LRU Perfect: Timestamp page on each reference Keep list of pages ordered by time of reference Too expensive to implement in reality for many reasons Clock Algorithm: Arrange physical pages in circle with single clock hand Approximate LRU (approx to approx to MIN) Replace an old page, not the oldest page Details: Hardware use bit per physical page: Hardware sets use bit on each reference If use bit isnt set, means not referenced in a long time Nachos hardware sets use bit in the TLB; you have to copy this back to page table when TLB entry gets replaced On page fault: Advance clock hand (not real time) Check use bit: 1used recently; clear and leave alone 0selected candidate for replacement Will always find a page or loop forever? 10/21/09 Even if all use bits set,

will eventually Kubiatowicz CS162 UCB Fall 2009 loop aroundFIFO Lec 15.20 Clock Algorithm: Not Recently Used Set of all pages in Memory Single Clock Hand: Advances only on page fault! Check for pages not used recently Mark pages as not used recently What if hand moving slowly? Good sign or bad sign? Not many page faults and/or find page quickly What if hand is moving quickly? Lots of page faults and/or lots of reference bits set One way to view clock algorithm: Crude partitioning of pages into two groups: young and old Why not partition intoCS162 more than 2 groups? Lec 15.21

10/21/09 Kubiatowicz UCB Fall 2009 Nth Chance version of Clock Algorithm Nth chance algorithm: Give page N chances OS keeps counter per page: # sweeps On page fault, OS checks use bit: 1clear use and also clear counter (used in last sweep) 0increment counter; if count=N, replace page Means that clock hand has to sweep by N times without page being used before page is replaced How do we pick N? Why pick large N? Better approx to LRU If N ~ 1K, really good approximation Why pick small N? More efficient Otherwise might have to look a long way to find free page What about dirty pages? Takes extra overhead to replace a dirty page, so give dirty pages an extra chance before replacing? Common approach: 10/21/09 Clean pages, use N=1 Dirty pages, use N=2 (and write back to disk when N=1) Kubiatowicz CS162 UCB Fall 2009

Lec 15.22 Clock Algorithms: Details Which bits of a PTE entry are useful to us? Use: Set when page is referenced; cleared by clock algorithm Modified: set when page is modified, cleared when page written to disk Valid: ok for program to reference this page Read-only: ok for program to read page, but not modify For example for catching modifications to code pages! Do we really need hardware-supported modified bit? No. Can emulate it (BSD Unix) using read-only bit Initially, mark all pages as read-only, even data pages On write, trap to OS. OS sets software modified bit, and marks page as read-write. Whenever page comes back in from disk, mark read-only 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.23 Clock Algorithms Details (continued) Do we really need a hardware-supported use bit? No. Can emulate it similar to above: Mark all pages as invalid, even if in memory On read to invalid page, trap to OS OS sets use bit, and marks page read-only Get modified bit in same way as previous: On write, trap to OS (either invalid or read-only) Set use and modified bits, mark page read-write When clock hand passes by, reset use and modified

bits and mark page as invalid again Remember, however, that clock is just an approximation of LRU Can we do a better approximation, given that we have to take page faults on some reads and writes to collect use information? Need to identify an old page, not oldest page! Answer: second chance list 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.24 Second-Chance List Algorithm (VAX/VMS) O LRU victim Second Chance List o rfl ve w Directly Mapped Pages Marked: RW List: FIFO Page-in From disk cc A

New Active Pages s s e Marked: Invalid List: LRU New SC Victims Split memory in two: Active list (RW), SC list (Invalid) Access pages in Active list at full speed Otherwise, Page Fault Always move overflow page from end of Active list to front of Second-chance list (SC) and mark invalid Desired Page On SC List: move to front of Active list, mark RW Not on SC list: page in to front of Active list, mark RW; page out LRU victim at end of SC list 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.25 Second-Chance List Algorithm (cont) How many pages for second chance list? If 0 FIFO If all LRU, but page fault on every page reference

Pick intermediate value. Result is: Pro: Few disk accesses (page only goes to disk if unused for a long time) Con: Increased overhead trapping to OS (software / hardware tradeoff) With page translation, we can adapt to any kind of access the program makes Later, we will show how to use page translation / protection to share memory between threads on widely separated machines Question: why didnt VAX include use bit? Strecker (architect) asked OS people, they said they didnt need it, so didnt implement it He later got blamed, but VAX did OK anyway 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.26 Free List Single Clock Hand: Set of all pages in Memory Advances as needed to keep freelist full (background) D D Free Pages For Processes

Keep set of free pages ready for use in demand paging Freelist filled in background by Clock algorithm or other technique (Pageout demon) Dirty pages start copying back to disk when enter list Like VAX second-chance list If page needed before reused, just return to active set Advantage: Faster for page fault Can always use page (or pages) immediately on fault 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.27 Demand Paging (more details) Does software-loaded TLB need use bit? Two Options: Hardware sets use bit in TLB; when TLB entry is replaced, software copies use bit back to page table Software manages TLB entries as FIFO list; everything not in TLB is Second-Chance list, managed as strict LRU Core Map Page tables map virtual page physical page Do we need a reverse mapping (i.e. physical page virtual page)? Yes. Clock algorithm runs through page frames. If sharing, then multiple virtual-pages per physical page Cant push page out to disk without invalidating all PTEs 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.28

Summary Precise Exception specifies a single instruction for which: All previous instructions have completed (committed state) No following instructions nor actual instruction have started Replacement policies FIFO: Place pages on queue, replace page at end MIN: Replace page that will be used farthest in future LRU: Replace page used farthest in past Clock Algorithm: Approximation to LRU Arrange all pages in circular list Sweep through them, marking as not in use If page not in use for one pass, than can replace Nth-chance clock algorithm: Another approx LRU Give pages multiple passes of clock hand before replacing Second-Chance List algorithm: Yet another approx LRU Divide pages into two groups, one of which is truly LRU and managed on page faults. 10/21/09 Kubiatowicz CS162 UCB Fall 2009 Lec 15.29

Recently Viewed Presentations

  • Invasive Species

    Invasive Species

    In New Zealand for example, the introduction of non-native species has caused an irreversible decline in many of the islands' bird species, such as the kakapo, the world's only flightless parrot. Water discharged from ships can potentially transport a huge...
  • Joint Relevance and Freshness Learning From Clickthroughs for ...

    Joint Relevance and Freshness Learning From Clickthroughs for ...

    Random bucket clicks. To validate the effectiveness of the proposed JRFL model in real news search tasks, we quantitatively compare it with all our baseline methods on: random bucket clicks, normal clicks, and editorial judgments. All the click-based learning algorithms...
  • Words &amp; Meaning: - ualberta.ca

    Words & Meaning: - ualberta.ca

    You might even believe it was true that 'The morning star is not the planet Venus' even though you know the reference of each term individually- and therefore of both terms Similarly many examples of a word's referent does not...
  • Medi-Cal Childrens Health Advisory Panel May 22, 2015

    Medi-Cal Childrens Health Advisory Panel May 22, 2015

    The SMHS medical necessity determination is based on an assessment of the beneficiary by the MHP. If medical necessity criteria for Medi-Cal SMHS are not met, the MHP will refer the beneficiary back to the MCP for mental health services...
  • Evidence Based Primary Care

    Evidence Based Primary Care

    Pyramida důkazů Vyhledávání důkazů Kde vyhledávat (multizdrojový přístup) Jak vyhledávat Rešeršní strategie (terminologie) Logické vazby mezi termíny (Booleovské operátory - AND, OR, NOT) Hodnotit kvalitu vyhledaných důkazů Hierarchie důkazů (typologie publikací) Kritická analýza Bibliometrie (citační ohlasy ...
  • The Path to Retirement - North American Securities ...

    The Path to Retirement - North American Securities ...

    Even guaranteed, fixed interest accounts, such as Certificates of Deposit, are subject to 'buying power risk' - the risk that the interest you earn will not be greater than the rate of inflation and, therefore the dollars returned to you...
  • Theorizing the spread of American English metaphorical idioms

    Theorizing the spread of American English metaphorical idioms

    The spread of metaphorical idioms in Englishes and beyond.A suggestion for a theoretical framework. Zoltán Kövecses. ELTE, Budapest
  • Diapositiva 1 - UCM

    Diapositiva 1 - UCM

    Es la plataforma de acceso público y gratuito a las bases de datos de la NLM (National Library of Medicine, USA).La principal de ellas es Medline.. Se crea en 1996/97, pero se puede localizar información desde 1946 aproximadamente. Es el...