Memory Management Duties. Overview of Presentation. (User View: code segments) User View: address space

Transcription

1 Overview of Presentation responsibilities of a memory manager user view of memory management memory management strategies, algorithms fixed partitions, variable partitions algorithms, fragmentation, coalescing swapping and the relocation problem material adequately covered by the text dynamic address relocation, swapping memory management 3/5/03-1 Memory Management Duties allocate/assign memory to processes explicit requests: malloc implicit: program loading, stack extension map in and out of process address space when allocated, with each context switch manage migration to/from 2 nd ary storage optimize use of main storage minimize overhead (waste, migrations) memory management 3/5/03-2 User View: address space Code segment Data segment Hole Stack segment 0x xFFFFFFFF DLL DLL DLL ALL of these segments appear to be present whenever process runs memory management 3/5/03-3 (User View: code segments) program code allocated when program loaded initialized with contents of load module Dynamically Loadable Libraries automatically mapped in whenever needed somehow shared by multiple processes all are read-only and fixed size memory management 3/5/03-4

2 (User View: data segments) initialized data allocated when program loaded initialized from load module expansion/contraction of data segment requested via system calls (e.g. sbrk) only added/truncated part is affected process stack allocated and grown automatically memory management 3/5/03-5 Implementing: code segments program loader ask for memory (size and virtual location) copy code from load module into memory run-time loader request DLL be mapped (location and size) edit linkage pointers from program to DLL memory manager allocates memory, maps into process memory management 3/5/03-6 Implementing: data segments program loader ask for memory (location and size) copy data from load module into memory zero the BSS memory manager invoked for allocations and stack extensions allocates and deallocates memory adjusts process address space accordingly memory management 3/5/03-7 Memory Allocation Strategies fixed partitions pre-allocate partitions for N processes each process assigned to one partition variable allocation one massive "heap" of memory divided among processes based on their needs paged allocation... and in that moment he became enlightened memory management 3/5/03-8

3 Fixed Partition Memory Allocation very easy to implement common in old batch processing systems well suited to well-known job mix presumes largest possible process size must reconfigure system for larger processes likely to use memory inefficiently large internal fragmentation losses swapping results in convoys on partitions memory management 3/5/03-9 Internal Fragmentation partition 1 8mb waste 2mb proc 1 6mb partition 2 4mb waste 1mb proc 2 3mb partition 3 4mb waste 3mb proc 3 1mb total waste = = 6/16Mb = 37% memory management 3/5/03-10 (Internal Fragmentation) wasted space in fixed sized blocks caused by a mis-match between the chosen sizes of a fixed-sized blocks the actual sizes that programs request average waste: 50% of each block overall waste reduced by multiple sizes Frequency of allocation requests frequency suppose blocks come in sizes S 1 and S 2 average waste = ((S 1 /2) + (S 2 - S 1 )/2)/2 memory management 3/5/ size of request memory management 3/5/03-12

4 (distribution of allocation requests) fixed partition allocation revisited internal fragmentation in fixed partition allocation resulted from random requests but are memory allocations really random? if we know what the popular sizes are create/reserve special pools of fixed size buffers satisfy matching requests from those pools there are peaks in the frequency v.s. size curve, some sizes are used and reused very often many services use fixed size transient buffers benefit: improved efficiency much simpler than variable partition allocation reduces external fragmentation of heap OS: terminal I/O, network packets, disk blocks,... but... we must know how much to reserve applications: messages, request descriptors,... too little -> contention, too much -> wasted space memory management 3/5/03-13 memory management 3/5/03-14 fixed partition bit-map free list Variable Memory Allocation free partition bit map (one bit per partition) start with one large "heap" of memory when a process requests more memory find a large enough chunk of memory carve off a piece of the requested size partn 1 in use partn 2 free partn 3 free partn 4 in use actual memory partitions partn 5 in use a bit-map is a very simple and compact way to keep track of which partitions are free and in use memory management 3/5/03-15 put the remainder back on the free list when a process frees memory put it back on the free list eliminates internal fragmentation losses memory management 3/5/03-16

5 External Fragmentation P 0 P 0 P 0 P 5 P 1 P 3 P 3 P 3 P 2 P 2 P 2 P 2 P 4 P 4 memory management 3/5/03-17 (External/Global Fragmentation) each allocation creates left-overs over time they become smaller and smaller the small left-over fragments are useless they are too small to satisfy any request a second form of fragmentation waste solutions: try not to create tiny fragments try to recombine fragments into big chunks memory management 3/5/03-18 head variable partition free list free free free Each chunk begins with a header: size of chunk, pointer to next chunk, other optional diagnostic information List might contain only free chunks, or all chunks memory management 3/5/03-19 (Free lists: keeping track of it all) fixed sized blocks are easy to track a bit map indicating which blocks are free variable allocation requires more information a linked list of descriptors, one per chunk each lists size of chunk, whether it is free each has pointer to next chunk on list descriptors often at front of each chunk allocated memory often has descriptors too memory management 3/5/03-20

6 Which chunk: best fit Which chunk: worst fit search for the "best fit" chunk search for the "worst fit" chunk smallest size greater/equal to requested size largest size greater/equal to requested size advantages: advantages: might find a perfect fit tends to create very large fragments disadvantages: disadvantages: have to search entire list every time still have to search entire list every time quickly creates very small fragments memory management 3/5/03-21 memory management 3/5/03-22 Which chunk: first fit take first chunk that is big enough advantages: very short searches creates random sized fragments disadvantages: the first chunks quickly fragment searches become longer ultimately it fragments as badly as best fit memory management 3/5/03-23 head Which chunk: next fit guess free free free after search, leave guess pointing to next free chunk next search starts where previous search left off when guess reaches end of list, wrap back to start memory management 3/5/03-24

7 (next-fit... guess pointers) Coalescing de-fragmentation the best of both worlds all VP algorithms have external fragmentation short searches (maybe shorter than first fit) some do it faster, some spread them out spreads out fragmentation (like worst fit) we need a way to reassemble fragments guess pointers are a general technique check neighbors when ever a chunk is freed think of them as a lazy (non-coherent) cache recombine free neighbors whenever possible if they are right, they save a lot of time free list can be designed to make this easier if they are wrong, the algorithm still works e.g. where are the neighbors of this chunk? they can be used in a wide range of problems this counters forces of external fragmentation memory management 3/5/03-25 memory management 3/5/03-26 Free list must support coalescing Coalescing vs. Fragmentation coalescing happens at free time opposing processes operate in parallel when freeing a region, check its neighbors if either neighbor is free, recombine them which of the two processes will dominate? what fraction of space is typically allocated? it must be easy to find both adjacent regions coalescing works better with more free space e.g. doubly linked list of region descriptors how fast is allocated memory turned over? other coupling between neighbors chunks held for long time cannot be coalesced (e.g. buddy system, where all regions are paired) how variable are requested chunk sizes? high variability increases fragmentation rate memory management 3/5/03-27 memory management 3/5/03-28

8 fixed vs. variable partitions memory allocation within a process fixed partition allocation use OS allocation to get memory for process internal fragmentation is inevitable e.g. sbrk system call to extend data segment multiple sizes may greatly reduce the problem variable partition allocation UNIX malloc (user mode memory allocation) variable partition, first fit-allocation more complex free lists, possible long searches go back to OS to get more if heap is empty external fragmentation evolves over time coalescing can counteract the fragmentation UNIX mfree (return memory when done) return memory to free list w/neighbor coalescing at a cost of more complexity and longer operations memory management 3/5/03-30 memory management 3/5/03-29 Diagnostic Free lists garbage collection common mistakes made w/dynamic memory memory leaks (allocate it and never free it) programmers often forget to free memory garbage collection is an alternative to freeing overruns (use more memory than you allocated) applications allocate objects, but never free them clobbers (keep on using it after you free it) free list can help to catch these problems when we run out, start garbage collection search data space finding every object pointer all chunks in list (whether allocated or free) note address/size of all accessible objects record of who last allocated each chunk compute the compliment (what is inaccessible) guard zones at beginning and end of chunks add all inaccessible memory to the free list memory management 3/5/03-31 memory management 3/5/03-32

9 enabling garbage collection What to do when coalescing loses how to find all accessible data? garbage collection is just another way to free in the general case, this is impossible ongoing activity can always starve coalescing object oriented languages often enable this chunks reallocated before they can be combined all object references are tagged we could stop accepting new allocations all object descriptors include size information convoy on memory manager destroys throughput it is often possible for system resources we need a way to rearrange active memory where all possible references are known (e.g. do any processes still owned by this user) re-pack all processes in one end of memory create one big chunk of free space at other end memory management 3/5/03-33 memory management 3/5/03-34 Memory Compaction Reasons to Relocate a Process to compact fragmented free space P 5 when a process increases its data space size P 3 swap device P 5 P 4 for example, the UNIX sbrk system call migration to and from 2 nd ary storage P 2 P 3 primary memory needed for runnable processes P 4 P 2 swap blocked processes to 2 nd ary storage memory before compaction memory after compaction memory management 3/5/03-35 swap them back-in when they become runnable the original memory may no longer be available? memory management 3/5/03-36

10 The Relocation Problem... how to move a process from memory region where it was initially loaded into a new and different region of memory all addresses in the program will be wrong references in the code segment calls and branches to other parts of the code references to variables in the data segment plus new pointers in data and stack segments memory management 3/5/03-37 The Relocation Problem... it is not feasible to re-relocate a process maybe we could re-relocate references to code if we kept the relocation information around but how can we relocate references to data pointer values may have been changed new pointers may have been created we could never find/fix all address references like the general case of garbage collection can we make processes location independent? memory management 3/5/03-38 Virtual Address Translation physical address space addresses on CPU/memory bus NNNN memory management unit virtual address space addresses seen by processes MMMM memory management 3/5/03-39 Segment Relocation computer has special relocation registers they are called segment base registers used in IBM/360, still used in Intel Pentium point to start of code, data, stack in memory code and data references are relative to these OS initializes base registers when loading program if process must be moved to a new location reset base registers to point to start of new region all code/data references are automatically relocated memory management 3/5/03-40

11 Segment Relocation base registers memory code base code data base stack base data stack physical = virtual + base Paged Memory Management divide physical memory into fixed-size pages typically 1,024, 4,096 or 8,192 bytes long divide virtual memory into same sized pages paging Memory Management Unit maps virtual page #s into physical page #s more flexible than base register relocation consecutive virtual pages need not map to consecutive physical pages memory management 3/5/03-41 memory management 3/5/03-42 Paging Address Translation Paging and Fragmentation virtual pages physical pages A code or data segment is made up of multple pages memory management unit vpage = virtual/pagesize; offset = virtual%pagesize; ppage = map(vpage); physical = (ppage * pagesize) + offset; memory management 3/5/03-43 Internal fragmentation is limited to the unused portion of the last page of the segment. If a page is 4K, the average waste is 2K. If the average segment is 200K, the expected loss due to internal fragmentation is 1%. External fragmentation is non-existent because we only allocate whole pages, we don't carve them up. memory management 3/5/03-44

12 (Paging solves many problems) internal fragmentation average of ½ page per address space segment but there are many pages per segment e.g. a 200K segment and a 4K page yields 1% loss external fragmentation non-existent: pages are never carved up finding enough contiguous memory contiguous physical memory is not required mapping can make any pages appear contiguous memory management 3/5/03-45 For the next lecture read chapter 12 there will be a quiz on this material topics for next lecture paging virtual address translation page faults and demand paging replacement algorithms and working sets segmentation and shared memory user and kernel address spaces memory management 3/5/03-46 key points key points memory manager responsibilities allocation, map-in, map-out, migration fixed partitions, internal fragmentation estimating and mitigating fragmentation variable partition, external fragmentation algorithms, performance, free-lists, coalescing special pools of fixed size buffers garbage collection what it is, why it is done, how it is done the relocation problem what it is, why it is hard how virtual address spaces solve it segment relocation, paged relocation advantages of paged allocation memory management 3/5/03-47 memory management 3/5/03-48