Basic allocator mechanisms The course that gives CMU its Zip! Memory Management II: Dynamic Storage Allocation Mar 6, 2000.

Transcription

1 5-23 The course that gives CM its Zip Memory Maagemet II: Dyamic Storage Allocatio Mar 6, 2000 Topics Segregated lists Buddy system Garbage collectio Mark ad Sweep Copyig eferece coutig Basic allocator mechaisms Sequetial fits (implicit or explicit sigle list) best fit, first fit, or ext fit placemet various splittig ad coalescig optios splittig thresholds immediate or deferred coalescig Segregated lists simple segregated storage -- separate heap for each size class segregated fits -- separate liked list for each size class systems 2 Segregate Storage Each size class has its ow collectio of blocks Ofte have separate collectio for every small size (2,3,4, ) For larger sizes typically have a collectio for each power of 2 3 Simple segregated storage Separate heap ad list for each size class No splittig To allocate a block of size : if list for size is ot empty, allocate first block o list (ote, list ca be implicit or explicit) if list is empty, get a ew page create ew list from all blocks i page allocate first block o list costat time To a block: Add to list If page is empty, retur the page for use by aother size (optioal) Tradeoffs: fast, but ca fragmet badly 4 Page

2 Segregated fits Array of lists, each oe for some size class To allocate a block of size : search appropriate list for block of size m > if a appropriate block is foud: split block ad place fragmet o appropriate list (optioal) if o block is foud, try ext larger class repeat util block is foud To a block: coalesce ad place o appropriate list (optioal) Tradeoffs faster search tha sequetial fits (i.e., log time for power of two size classes) cotrols fragmetatio of simple segregated storage coalescig ca icrease search times deferred coalescig ca help Buddy systems Special case of segregated fits. all blocks are power of two sizes Basic idea: Heap is 2 m words Maitai separate lists of each size 2 k, 0 <= k <= m. equested block sizes are rouded up to earest power of 2. Origially, oe block of size 2 m. 5 6 Buddy systems (cot) To allocate a block of size 2 k : Fid first available block of size 2 j s.t. k <= j <= m. if j == k the doe. otherwise recursively split block util j == k. Each remaiig half is called a ad is placed o the appropriate list 2 m Buddy systems (cot) To a block of size 2 k cotiue coalescig with buddies while the buddies are Block to Not, doe Added to appropriate list 7 8 Page 2

3 Buddy systems (cot) Key fact about systems: give the address ad size of a block, it is easy to compute the address of its e.g., block of size 32 with address xxx...x00000 has xxx...x0000 Tradeoffs: fast search ad coalesce subject to iteral fragmetatio Iteral fragmetatio Iteral fragmetatio is wasted space iside allocated blocks: miimum block size larger tha requested amout e.g., due to miimum block size, list overhead policy decisio ot to split blocks e.g., system Much easier to defie ad measure tha exteral fragmetatio. 9 0 Implicit Memory Maagemet Garbage collector Garbage collectio: automatic reclamatio of heapallocated storage -- applicatio ever has to void foo() { it *p = malloc(28); retur; /* p block is ow garbage */ } Commo i fuctioal laguages, scriptig laguages, ad moder object orieted laguages: Lisp, ML, Java, Perl, Mathematica, Variats (coservative garbage collectors) exist for C ad C++ Caot collect all garbage Garbage Collectio How does the memory maager kow whe memory ca be d? I geeral we caot kow what is goig to be used i the future sice it depeds o coditioals But we ca tell that certai blocks caot be used if there are o poiters to them Need to make certai assumptios about poiters Memory maager ca distiguish poiters from o-poiters All poiters poit to the start of a block Caot hide poiters (e.g. by coercig them to a it, ad the back agai) 2 Page 3

4 Classical GC algorithms Mark ad sweep collectio (McCarthy, 960) Does ot move blocks (uless you also compact ) eferece coutig (Collis, 960) Does ot move blocks Copyig collectio (Misky, 963) Moves blocks Memory as a graph We view memory as a directed graph Each block is a ode i the graph Each poiter is a edge i the graph Locatios ot i the heap that cotai poiters ito the heap are called root odes (e.g. registers, locatios o the stack, global variables) oot odes For more iformatio see Joes ad Li, Garbage Collectio: Algorithms for Automatic Dyamic Memory, Joh Wiley & Sos, 996. Heap odes reachable Not-reachable (garbage) A ode (block) is reachable if there is a path from ay root to that ode. No-reachable odes are garbage (ever eeded by the applicatio) 3 4 Assumptios for this lecture Applicatio ew(): returs poiter to ew block with all locatios cleared read(b,i): read locatio i of block b ito register write(b,i,v): write v ito locatio i of block b Each block will have a header word addressed as b[-], for a block b sed for differet purposes i differet collectors Istructios used by the Garbage Collector is_ptr(p): determies whether p is a poiter legth(b): returs the legth of block b, ot icludig the header get_roots(): returs all the roots Mark ad sweep collectig Ca build o top of malloc/ package Allocate usig malloc util you ru out of space Whe out of space: se extra mark bit i the head of each block Mark: Start at roots ad set mark bit o all reachable memory Sweep: Sca all blocks ad blocks that are ot marked Before mark After mark root Mark Bit Set After sweep 5 6 Page 4

5 # & ' ( -,+ *) ' ( * Mark ad sweep (cot.) Mark usig depth-first traversal of the memory graph Mark ad sweep i C A C Coservative Collector Is_ptr() ca determies if a word is a poiter by checkig if it poits to a allocated block of memory. But, i C poiters ca poit to the middle of a block. ptr head Sweep usig legths to fid ext block So how do we fid the begiig of the block Ca use balaced tree to keep track of all allocated blocks where the key is the locatio Balaced tree poiters ca be stored i head (use two additioal words) head data size left right 7 8 Copyig collectio Keep two equal-sized spaces, from-space ad to-space epeat util applicatio fiishes Applicatio allocates i oe space cotiguously util space is full. Stop applicatio ad copy all reachable blocks to cotiguous locatios i the other space. Flip the roles of the two spaces ad restart applicatio. Before copy (from space) After copy (to space) root Copy does ot ecessarily keep the order of the blocks Has the effect or removig all fragmets Copyig collectio (ew) ptr ew (it ) { if (++ > top) flip(); ewblock = ; += (+); for (i=0; i < +; i++) ewblock[i] = 0; retur ewblock+; } All ew blocks are allocated i, oe after the other A extra word is allocated for the header The Garbage-Collector starts (flips), whe we reach 9 20 Page 5

6 # # # # # ' - - +, + & # # & # # Copyig collectio (copy) Page 6 # # 22 eferece coutig Set referece cout to oe whe creatig a ew block allocate Whe readig a value icremet its referece couter Whe writig decremet the old value ad icremet the ew value & # 24 Copyig collectio (flip) reachable Not reachable & After the first three lies of flip (before the copy). 2 eferece coutig Basic algorithm Keeps cout o each block of how may poiters poit to the block Whe a cout goes to zero, the block ca be d Data structures Ca be built o top of a existig explicit allocator * - ( Add a additioal header word for the referece cout 3 Keepig the cout updated requires that the we modify every read ad write (ca be optimized out i some cases) 23

7 # # # # # # Page 7 eferece coutig example Iitially T 2 S V Now cosider: write(,,nll) This will execute a: decremet(s) 26 eferece coutig example After: decremet(s[0]) T S 0 # V 28 eferece coutig Decremet if couter decremets to zero the the block ca be d whe ig a block, the algorithm must decremet the couters of everythig poited to by the block -- this might i tur recursively more blocks # 25 eferece coutig example After couter o S is decremeted T 2 S 0 # V 27

8 eferece coutig example After: decremet(s[]) eferece coutig cyclic data structures Before After T S 0 0 S 2 S V 0 T 2 T Garbage Collectio Summary Copyig Collectio Pros: prevets fragmetatio, ad allocatio is very cheap Cos: requires twice the space (from ad to), ad stops allocatio to collect Mark ad Sweep Pros: requires little extra memory (assumig low fragmetatio) ad does ot move data Cos: allocatio is somewhat slower, ad all memory eeds to be scaed whe sweepig eferece Coutig Pros: requires little extra memory (assumig low fragmetatio) ad does ot move data Cos: reads ad writes are more expesive ad difficult to deal with cyclic data structures Some collectors use a combiatio (e.g. copyig for small objects ad referece coutig for large objects) 3 Page 8