#4 Inverted page table. The need for more bookkeeping. Inverted page table architecture. Today. Our Small Quiz

Size: px

Start display at page:

Download "#4 Inverted page table. The need for more bookkeeping. Inverted page table architecture. Today. Our Small Quiz"

Priscilla Peters
6 years ago
Views:

ADRIAN PERRIG & TORSTEN HOEFLER Networks and Operatng Systems (-6-) Chapter 6: Demand Pagng http://redmne.replcant.

nformaton on machne Hashng used to locate an entry effcently Examples: PowerPC, a6, UltraSPARC Inverted table archtecture The need for more bookkeepng CPU logcal pd p d d physcal Physcal Most OSes

Portablty Trackng objects Software vrtual physcal translaton Physcal vrtual translaton search pd p table Our Small Quz True or false (rase hand).

1 ADRIAN PERRIG & TORSTEN HOEFLER Networks and Operatng Systems (-6-) Chapter 6: Demand Pagng () # Inverted table One system-wde table now maps PFN -> VPN One entry for each real of Contans VPN, and whch process owns the Bounds total sze of all nformaton on machne Hashng used to locate an entry effcently Examples: PowerPC, a6, UltraSPARC Inverted table archtecture The need for more bookkeepng CPU logcal pd p d d physcal Physcal Most OSes keep ther own translaton nfo Per-process herarchcal table (Lnux) System wde nverted table (Mach, MacOS) Why? Portablty Trackng objects Software vrtual physcal translaton Physcal vrtual translaton search pd p table Our Small Quz True or false (rase hand). Base (relocaton) and lmt regsters provde a full vrtual. Base and lmt regsters provde protecton. Segmentaton provdes a base and lmt for each segment. Segmentaton suffers from external fragmentaton. Segmentaton allows lbrares to share ther code 6. Segmentaton provdes lnear ng 7. Segment tables are set up for each process n the CPU 8. Segmentng prevents nternal fragmentaton 9. Pagng prevents nternal fragmentaton.protecton nformaton s stored at the physcal frame.pages can be shared between processes.the same may be wrteable n proc. A and wrte protected n proc. B.The same physcal can be referenced through dfferent es from (a) two dfferent processes (b) the same process?.inverted tables are faster to search than herarchcal (asymptotcally) Today TLB shootdown Uses for vrtual Copy-on-wrte Demand pagng Page fault handlng Page replacement algorthms Frame allocaton polces Thrashng and workng set Book: OSPP Sectons 9., 9.7 (all of 9 as refresh) As always, the book does not cover %!

2 TLB management TLB shootdown Recall: the TLB s a cache. Machnes have many MMUs on many cores many TLBs Problem: TLBs should be coherent. Why? Securty problem f mappngs change E.g., when s reused TLB management TLB management Core Process ID VPN PPN acce ss x x r/w xf8 x r/w Core Process ID VPN PPN acce ss x x r/w xf8 x r/w Change to read only Core x x r/w x x read Core x x r/w x x read Core xf8 x r/w x x read Core xf8 x r/w x x read TLB management TLB management Core Core Process ID VPN PPN acce ss x x r/w xf8 x r/w x x r/w x x read Change to read only Core Core Process ID VPN PPN acce ss x x r/w xf8 x r/w x x r/w x x read Change to read only Core xf8 x r/w x x read Core xf8 x r/w x x read Process on core can only contnue once shootdown s complete!

3 Keepng TLBs coherent. Hardware TLB coherence Integrate TLB mgmt wth cache coherence Invaldate TLB entry when PTE changes Rarely mplemented. Vrtual caches Requred cache flush / nvaldate wll take care of the TLB Hgh context swtch cost! Most processors use physcal (last level) caches. Software TLB shootdown Most common OS on one core notfes all other cores - typcally an IPI Each core provdes local nvaldaton. Hardware shootdown nstructons Broadcast specal access on the bus Interpreted as TLB shootdown rather than cache coherence message E.g., PowerPC archtecture Summary/recap: vrtual User logcal physcal. Only part of the program must be n RAM for executon Logcal can be larger than physcal Address s can be shared by several processes More effcent process creaton Vrtualze usng software+hardware The many uses of translaton Copy-on-wrte (COW) Process solaton Memory mapped fles IPC Vrtual Shared code segments Checkpont and restart Program ntalzaton Persstent data structures Effcent dynamc allocaton Process mgraton Cache management Informaton flow control Program debuggng Dstrbuted shared Effcent I/O and many more Photo by Josh Hammerlng Recall fork() Can be expensve to create a complete copy of the process Especally just to do exec()! vfork(): shares, doesn t copy Fast Dangerous two wrters to same heap Better: only copy when you know somethng s gong to get wrtten Requres MMU/ vrtualzaton Copy-on-wrte COW allows both parent and chld processes to ntally share the same s n If ether process modfes a shared, only then s the coped COW allows more effcent process creaton as only modfed s are coped Free s are allocated from a pool of zeroed-out s

Example: processes sharng an area of Example: processes sharng an area of Not necessarly the same vrtual es (but would be after fork()) A A B B C C Process physcal Process

After process wrtes to C Intally mark all s as read-only Ether process wrtes fault Fault handler allocates new frame Makes copy of n new frame Maps each copy nto resp.

4 Example: processes sharng an area of Example: processes sharng an area of Not necessarly the same vrtual es (but would be after fork()) A A B B C C Process physcal Process Process physcal Process How does t work? After process wrtes to C Intally mark all s as read-only Ether process wrtes fault Fault handler allocates new frame Makes copy of n new frame Maps each copy nto resp. processes wrteable A Only modfed s are coped Less usage, more sharng Cost s fault for each mutated B C copy of C Process physcal Process After process wrtes to C After process wrtes to C A B Stll readonly A B Stll readonly C C copy of C Now wrteable copy of C Process physcal Process Process physcal Process

had happened General: allows emulaton of as well as multplexng.

5 General prncple Pagng concepts Mark a VPN as nvald or read-only trap ndcates attempt to read or wrte On a fault, change mappngs somehow Restart nstructon, as f nothng had happened General: allows emulaton of as well as multplexng. E.g. on-demand zero-fllng of s And v vrtual table physcal dsk Pagng concepts Pagng concepts -resdent v vrtual table physcal dsk v vrtual Keep track of where s are on dsk table physcal dsk Pagng concepts Pagng concepts Wrte drty s out to dsk Read n s from dsk on demand v vrtual Keep track of where s are on dsk table physcal dsk v vrtual table physcal dsk

6 Demand pagng Brng a nto only when t s needed Less I/O needed Less needed Faster response More users Turns RAM nto a cache for processes on dsk! Demand pagng Page needed reference (load or store) to t nvald reference abort not-n- brng to Lazy swapper never swaps a nto unless wll be needed Swapper that deals wth s s a r Can do ths wth segments, but more complex or whole processes Strct demand pagng: only n when referenced Page fault If there s a reference to a, frst reference to that wll trap to operatng system: fault. Operatng system looks at another table to decde: Invald reference abort Just not n. Get empty frame. Swap nto frame. Reset tables. Set vald bt v 6. Restart the nstructon that caused the fault Recall: handlng a fault CPU Chp CPU VA MMU PTEA ) Processor sends vrtual to MMU Cache/ Memory -) MMU fetches PTE from table n ) Vald bt s zero, so MMU trggers fault excepton ) Handler dentfes vctm (and, f drty, s t out to dsk) 6) Handler s n new and updates PTE n 7) Handler returns to orgnal process, restartng faultng nstructon PTE Dsk Recall: handlng a fault Recall: handlng a fault Excepton Page fault handler Excepton Page fault handler CPU Chp PTEA CPU Chp PTEA CPU VA MMU PTE Cache/ Memory Dsk CPU VA MMU PTE Cache/ Memory Dsk ) Processor sends vrtual to MMU -) MMU fetches PTE from table n ) Vald bt s zero, so MMU trggers fault excepton ) Handler dentfes vctm (and, f drty, s t out to dsk) 6) Handler s n new and updates PTE n 7) Handler returns to orgnal process, restartng faultng nstructon ) Processor sends vrtual to MMU -) MMU fetches PTE from table n ) Vald bt s zero, so MMU trggers fault excepton ) Handler fnds a frame to use for mssng 6) Handler s n new and updates PTE n 7) Handler returns to orgnal process, restartng faultng nstructon

Recall: handlng a fault Recall: handlng a fault Excepton Excepton Page fault handler PTEA CPU Chp CPU VA MMU Page fault handler PTE Cache/ Memory PTEA CPU Chp New 6 Dsk CPU VA MMU 7 PTE New Cache/

7 Recall: handlng a fault Recall: handlng a fault Excepton Excepton Page fault handler PTEA CPU Chp CPU VA MMU Page fault handler PTE Cache/ Memory PTEA CPU Chp New 6 Dsk CPU VA MMU 7 PTE New Cache/ Memory 6 ) Processor sends vrtual to MMU ) Processor sends vrtual to MMU -) MMU fetches PTE from table n -) MMU fetches PTE from table n ) Vald bt s zero, so MMU trggers fault excepton ) Vald bt s zero, so MMU trggers fault excepton ) Handler fnds a frame to use for mssng ) Handler fnds a frame to use for mssng 6) Handler s n new and updates PTE n 6) Handler s n new and updates PTE n 7) Handler returns to orgnal process, restartng faultng nstructon 7) Handler returns to orgnal process, restartng faultng nstructon Performance of demand pagng Demand pagng example Page fault rate p. Memory access tme = nanoseconds f p = : no faults f p = : every reference s a fault Dsk Average -fault servce tme = mllseconds Effectve Access Tme (EAT) EAT = ( p) x access + p ( fault overhead + swap out + swap n + restart overhead ) EAT = ( p) x + p ( mllseconds) = ( p) x + p x,, = + p x,999,9 If one access out of, causes a fault, then EAT = mcroseconds. Ths s a slowdown by a factor of 8!! Page Replacement What happens f there s no free frame? Page replacement fnd lttle used resdent to dscard or wrte to dsk vctm needs selecton algorthm performance want an algorthm whch wll result n mnmum number of faults Same may be brought nto several tmes Photo: Urs Flueeler, source:

Page replacement Page replacement Try to pck a vctm whch won t be referenced n the future Varous heurstcs

drty ones (save a dsk wrte) frame vald f v f vctm Page table Physcal Page replacement Page replacement

Change vctm PTE to nvald Page table Page table Physcal Physcal Page replacement Page replacement frame

8 Page replacement Page replacement Try to pck a vctm whch won t be referenced n the future Varous heurstcs but ultmately t s a guess Use modfy bt on PTE Don t wrte clean (unmodfed) to dsk Try to pck clean s over drty ones (save a dsk wrte) frame vald f v f vctm Page table Physcal Page replacement Page replacement frame vald. Swap vctm to dsk frame vald. Swap vctm to dsk f v f vctm f vctm. Change vctm PTE to nvald Page table Page table Physcal Physcal Page replacement Page replacement frame vald frame vald. Change fault PTE to vald f vctm f v f vctm Page table Page table. Load desred n from dsk. Load desred n from dsk Physcal Physcal

9 Number of faults Page replacement algorthms Page faults vs. number of frames Want lowest -fault rate Evaluate algorthm by runnng t on a partcular strng of references (reference strng) and computng the number of faults on that strng Very lttle : thrashng (see later) What we mght expect E.g. 7,,,,,,,,,,,,,,,,, 7,, Plenty of : more doesn t help much Number of frames FIFO (Frst-In-Frst-Out) replacement FIFO (Frst-In-Frst-Out) replacement FIFO (Frst-In-Frst-Out) replacement FIFO (Frst-In-Frst-Out) replacement

10 FIFO (Frst-In-Frst-Out) replacement FIFO (Frst-In-Frst-Out) replacement FIFO (Frst-In-Frst-Out) replacement FIFO (Frst-In-Frst-Out) replacement FIFO (Frst-In-Frst-Out) replacement More s better? Reference strng:,,,,,,,,,,, rames: Here, faults. Belady s Anomaly: more frames more faults

11 More s better? More s better? Reference strng:,,,,,,,,,,, frames ( s can be n ): Reference strng:,,,,,,,,,,, frames ( s can be n ): Belady s Anomaly: more frames more faults Belady s Anomaly: more frames more faults More s better? More s better? Reference strng:,,,,,,,,,,, frames ( s can be n ): Reference strng:,,,,,,,,,,, frames ( s can be n ): Belady s Anomaly: more frames more faults Belady s Anomaly: more frames more faults More s better? More s better? Reference strng:,,,,,,,,,,, frames ( s can be n ): Reference strng:,,,,,,,,,,, frames ( s can be n ): 9 faults 9 faults Belady s Anomaly: more frames more faults Belady s Anomaly: more frames more faults

12 Number of faults More s better? More s better? Reference strng:,,,,,,,,,,, frames ( s can be n ): Reference strng:,,,,,,,,,,, frames ( s can be n ): 9 faults 9 faults faults! Belady s Anomaly: more frames more faults Belady s Anomaly: more frames more faults More s better? FIFO showng Belady s Anomaly Reference strng:,,,,,,,,,,, frames ( s can be n ): 9 faults faults! Belady s Anomaly: more frames more faults Number of frames Optmal algorthm Optmal replacement Replace that wll not be used for longest perod of tme frames example: faults Here, 9 faults. How do you know ths? you can t! Used for measurng how well your algorthm performs

13 Least Recently Used (LRU) algorthm LRU replacement Reference strng: Counter mplementaton Every entry has a counter; every tme s referenced through ths entry, copy the clock nto the counter When a needs to be changed, look at the counters to determne whch are to change Here, faults. LRU stack algorthm Stack mplementaton keep a stack of numbers n a double lnk form: Page referenced: move t to the top requres 6 ponters to be changed No search for replacement General term: stack algorthms Have property that addng frames always reduces faults (no Belady s Anomaly) Use a stack to record most recent references Reference strng 7 7 LRU approxmaton algorthms Second-chance (clock) replacement algorthm Reference bt Wth each assocate a bt, ntally = When s referenced bt set to Replace a whch s (f one exsts) We do not know the order, however Second chance Need reference bt Clock replacement If to be replaced (n clock order) has reference bt = then: set reference bt leave n replace next (n clock order), subject to same rules Next vctm ( clock hand ) Crcular queue of s Reference bts

14 Second-chance (clock) replacement algorthm Second-chance (clock) replacement algorthm Next vctm ( clock hand ) Crcular queue of s Next vctm ( clock hand ) Crcular queue of s Reference bts Reference bts Second-chance (clock) replacement algorthm Frame allocaton polces (mult-process) Next vctm ( clock hand ) Crcular queue of s Reference bts Allocaton of frames Each process needs mnmum number of s Example: IBM 7 6 s to handle SS MOVE nstructon: nstructon s 6 bytes, mght span s s to handle from s to handle to Two major allocaton schemes fxed allocaton prorty allocaton Fxed allocaton Equal allocaton all processes get equal share Proportonal allocaton allocate accordng to the sze of process s sze of process p S s m total number of frames s a allocaton for p m S m 6 s s 7 a a 6 9 7

Useful CPU utlzaton Prorty allocaton Proportonal allocaton scheme Usng prortes rather than sze If process P generates a

local allocaton Global replacement process selects a replacement frame from the set of all frames; one process can take a

process does not have enough s, the fault rate s very hgh.

15 Useful CPU utlzaton Prorty allocaton Proportonal allocaton scheme Usng prortes rather than sze If process P generates a fault, select:. one of ts frames, or. frame from a process wth lower prorty Global vs. local allocaton Global replacement process selects a replacement frame from the set of all frames; one process can take a frame from another Local replacement each process selects from only ts own set of allocated frames Thrashng Thrashng If a process does not have enough s, the fault rate s very hgh. Ths leads to: low CPU utlzaton operatng system thnks that t needs to ncrease the degree of multprogrammng another process added to the system Thrashng begns! Thrashng a process s busy swappng s n and out Demand for vrtual (e.g., more procs) Source: wkpeda Demand pagng and thrashng Localty n a reference pattern Why does demand pagng work? Localty model Process mgrates from one localty to another Localtes may overlap Why does thrashng occur? sze of localtes > total sze

16 Rate of faults Workng-set model workng-set wndow a fxed maxmum number of references Example:, nstructon WSS (workng set of process P ) = total number of s referenced n the most recent (vares n tme) too small wll not encompass entre localty too large wll encompass several localtes = wll encompass entre program Allocate demand frames D = WSS total demand frames Intuton: how much s really needed D > m Thrashng Polcy: f D > m, suspend some processes Workng-set model Page WS(t ) = {,,,6,7} WS(t ) = {,} t t Keepng track of the workng set Approxmate wth nterval tmer + a reference bt Example: =, Tmer nterrupts after every, tme unts Keep n bts for each Whenever a tmer nterrupts shft+copy and sets the values of all reference bts to If one of the bts n = n workng set Why s ths not completely accurate? Hnt: Nyqust-Shannon! Keepng track of the workng set Page-fault frequency scheme Approxmate wth nterval tmer + a reference bt Example: =, Tmer nterrupts after every tme unts Keep n bts for each Whenever a tmer nterrupts shft+copy and sets the values of all reference bts to If one of the bts n = n workng set Why s ths not completely accurate? Improvement = bts and nterrupt every, tme unts Establsh acceptable -fault rate If actual rate too low, process loses frame If actual rate too hgh, process gans frame Increase number of frames Upper bound Lower bound Number of frames Decrease number of frames

If you miss a key. Chapter 6: Demand Paging Source:

If you miss a key. Chapter 6: Demand Paging Source: ADRIAN PERRIG & TORSTEN HOEFLER ( -6- ) Networks and Operatng Systems Chapter 6: Demand Pagng Source: http://redmne.replcant.us/projects/replcant/wk/samsunggalaxybackdoor If you mss a key after yesterday