Memory Management 1 Overview Basic memory management Address Spaces Virtual memory Page replacement algorithms Design issues for paging systems Implementation issues Segmentation 2 1
Memory Management Ideally programmers want memory that is large fast non volatile Memory hierarchy small amount of fast, expensive memory cache some medium-speed, medium price main memory gigabytes of slow, cheap disk storage Memory manager handles the memory hierarchy Protects processes from each other. 3 Approaches Single Process, Contiguous Memory Multiple Processes, Contiguous Memory Multiple Processes, Discontiguous Memory Multiple Processes, Only partially in memory 4 2
Basic Memory Management Single Process without Swapping or Paging Three simple ways of organizing memory - an operating system with one user process 5 Binding If a program has a line: int x; When is the address of x determined? What are the choices? At compile time At load time At run time 6 3
Base / Limit Registers Binding done at run time. Addresses are added to base value to map to physical address Addresses larger than limit value are an error 7 Swapping Memory allocation changes as processes come into memory leave memory Shaded regions are unused memory 8 4
Managing Free Memory Assume a process being loaded can ask for any size chunk of memory needed. We need to be able to find a chunk the right size. How can we keep track of free chunks efficiently? 9 Memory Management with Bit Maps Part of memory with 5 processes, 3 holes tick marks show allocation units shaded regions are free Corresponding bit map Same information as a list 10 5
Memory Management with Linked Lists Four neighbor combinations for the terminating process X 11 Order of Search for Free Memory We can search for a large enough free block of free memory starting from the beginning. That s called first fit. If we already skipped over the first N holes because they were too small, maybe it s a waste of time to look there again. Try next fit. Should we be more selective in our choice? After all we re just grabbing the first thing that works How does first (or next) fit effect fragmentation? Could we do better? Can we find a hole that fits faster? What are the downsides? 12 6
The Contiguous Constraint So far a process s memory has been contiguous. What if it didn t have to be? What problems would that help solve? How would the hardware need to change? What additional work would the OS have to do? 13 It s All Gotta Be in Memory (or does it?) We have assumed that the entire process has to be in memory whenever it is running. What if it didn t have to be? What problems would that help solve? How would the hardware need to change? What additional work would the OS have to do? 14 7
Virtual Memory The position and function of the MMU 15 Paging The relation between virtual addresses and physical memory addresses given by page table 16 8
Page Tables Internal operation of MMU with 16 4 KB pages 17 Page Tables 2-level Second-level page tables Top-level page table 32 bit address with 2 page table fields Two-level page tables 18 9
Page Table Entry Present/absent is also called Valid Modified is also called Dirty Referenced is also called Accessed Why would caching be disable? 19 Pentium PTE 20 10
TLBs Translation Lookaside Buffers A TLB to speed up paging 21 Inverted Page Tables Comparison of a traditional page table with an inverted page table 22 11
Page Replacement Algorithms Page fault forces choice If there are no free page frames, we have to make room for incoming page Which page should be removed? Modified page frame must first be saved before being evicted An unmodified page frame can just overwritten Better not to choose an often used page Likely to be brought back in again soon 23 Optimal Page Replacement Algorithm Replace page needed at the farthest point in future Optimal but unrealizable Estimate by logging page use on previous runs of process although this is impractical 24 12
Not Recently Used Page Replacement Algorithm Each page has Reference bit, Modified bit bits are set when page is referenced, modified Pages are classified 1. not referenced, not modified 2. not referenced, modified 3. referenced, not modified 4. referenced, modified NRU removes page at random from lowest numbered non empty class 25 FIFO Page Replacement Algorithm Maintain a linked list of all pages in order they came into memory Page at beginning of list replaced Disadvantage page in memory the longest may be often used 26 13
Second Chance Page Replacement Algorithm Operation of a second chance pages sorted in FIFO order Page list if fault occurs at time 20, A has R bit set (numbers above pages are loading times) 27 The Clock Page Replacement Algorithm 28 14
Least Recently Used (LRU) Principle: assume a page used recently will be used again soon throw out page that has been unused for longest time Implementation Keep a linked list of pages most recently used at front, least at rear update this list every memory reference!! Or keep time stamp with each PTE choose page with oldest time stamp Again, this must be updated with every memory reference. 29 LRU (Another Hardware Solution) LRU using a matrix pages referenced in order 0,1,2,3,2,1,0,3,2,3 30 15
LRU Approximation? How could we approximate LRU? We can t track every time a page is referenced, but we can sample the data. How often? Once per clock tick, perhaps. Update a counter for each page that has been referenced in the last clock tick. Take the page with the lowest count i.e. Not Frequently Used (NFU) How well does this work? 31 Simulating LRU in Software Aging The aging algorithm simulates LRU in software Note 6 pages for 5 clock ticks, (a) (e) 32 16
Thrashing A program causing page faults every few instructions is said to be thrashing. What causes thrashing? If a process keeps accessing random new pages, then it is hard to anticipate what it will use next. Most programs exhibit temporal and spatial locality. Temporal locality: if the process accessed a particular address, it is likely to do so again soon. Spatial locality: if the process accessed a particular address, it is likely to access nearby addresses soon. Processes that follow this principle tend not to thrash unless they have to fight for memory. 33 Causing Thrashing within a single process const int ROWS = 10000; const int COLS = 1024; int arr[rows][cols]; int main() { for (int row = 0; row < ROWS; ++row) for (int col = 0; col < COLS; ++col) arr[row][col] = row * col; } for (int col = 0; col < COLS; ++col) for (int row = 0; row < ROWS; ++row) arr[row][col] = row * col; The first nested loop demonstrates spatial locality The second thrashes. 34 17
The Working Set Page Replacement Algorithm (1) The working set is the set of pages used by the k most recent memory references w(k,t) is the size of the working set at time, t 35 The Working Set Page Replacement Algorithm (2) The working set algorithm 36 18
The WSClock Page Replacement Algorithm Operation of the WSClock algorithm 37 Review of Page Replacement Algorithms 38 19
Modeling Page Replacement Algorithms Belady's Anomaly FIFO with 3 page frames FIFO with 4 page frames P's show which page references show page faults 39 Stack Algorithms State of memory array, M, after each item in reference string is processed 40 20
The Distance String Probability density functions for two hypothetical distance strings 41 The Distance String Computation of page fault rate from distance string the C vector the F vector 42 21
Design Issues for Paging Systems Local versus Global Allocation Policies Original configuration Local page replacement Global page replacement 43 Page Fault Rate Page fault rate as a function of the number of page frames assigned Use to determine if a process should be granted additional pages. 44 22
Load Control Despite good designs, system may still thrash When PFF algorithm indicates some processes need more memory but no processes need less Solution : Reduce number of processes competing for memory swap one or more to disk, divide up frames they held reconsider degree of multiprogramming 45 Advantages Page Size (1) Small page size less internal fragmentation better fit for various data structures, code sections less unused program in memory Disadvantages programs need many pages, larger page tables 46 23
Page Size (2) Overhead due to page table and internal fragmentation page table space Where s = average process size in bytes p = page size in bytes e = page entry internal fragmentation Optimized when 47 Separate Instruction and Data Spaces One address space Separate I and D spaces 48 24
Shared Pages Two processes sharing same program sharing its page table 49 Cleaning Policy Need for a background process, paging daemon periodically inspects state of memory When too few frames are free selects pages to evict using a replacement algorithm. Evicted pages are kept in a pool in case they are wanted again. Dirty pages are scheduled to be written out. 50 25
Implementation Issues Operating System vs. Hardware for Paging 1. Process creation Create and initialize page table. Pre-fetch pages. 2. Context Switch Point MMU to page table for new process TLB flushed 3. Memory Reference Map virtual address to physical address Determine if page fault If fault, determine whose fault and resolve 4. Page Replacement Record page access / modifies Determine page to replace. 5. Process termination time release page table, pages 51 Instruction Backup An instruction causing a page fault 52 26
Global Flag Some pages are used by every process in the system. Which? What would we like to have happen special for those pages during a context switch. How can the hardware know? 53 Locking Pages in Memory Virtual memory and I/O occasionally interact Proc issues call for read from device into buffer while waiting for I/O, another processes starts up has a page fault buffer for the first proc may be chosen to be paged out Need to specify some pages locked exempted from being target pages 54 27
Backing Store (a) Paging to static swap area (b) Backing up pages dynamically 55 Separation of Policy and Mechanism Page fault handling with an external pager 56 28
Memory Management NT & Unix Exact details not as easy as in process scheduling. Attempt to keep a set of free pages using a page daemon. (e.g. Linux: kswapd, NT: Working Set Manger) Essentially demand paging. NT brings in clusters, typically 8 pages for code, 4 for data. Unix s swapper used to bring in all referenced pages, when swapping a process back in. Now uses demand paging. OS present in every process s virtual address space. Reduces changes needed in TLB and cache. No changes needed for system call. Load control: Swap out entire processes as needed. This is currently something that one hopes to avoid, but the feature is still present. Some structure to describe where pages are currently, e.g. location in paging file. Unix: Core Map. NT: Page Frame Database. A portion (e.g., 64KB) of the top/bottom of address space unmapped. 57 MM in Unix Early versions totally swapping Current based on paging. Page Daemon ¼ sec. Refill free list when less then some value. BSD: lotsfree = ¼ memory. SysV: has some min/max. If less than min, fill until max. Goal is to reduce frequency of page daemon, thereby reducing thrashing. Originally used Global Clock. As memory sizes increased BSD went to 2-handed clock. Front clears referenced bit Back hand picks victim, Result: a referenced page is really recent. SysV instead keeps frequency count of pages not referenced. Only boots page out if count greater than some value Seems they had the opposite problem of the BSD folk. 58 29
MM in Unix Load Control. If page daemon has to work too often then swap someone out. Every few seconds look for someone to bring back. Decision based on swapped out time, size, niceness, if it was sleeping Only bring back the user structure portion of PCB and the page tables. (I m surprised page tables are removed. Must record swapped out page for process somewhere ) Core Map describes frames, free list, location to swap to. 59 MM in NT Demand paging of Clusters. (code: 8, data: 4). XP does some prepaging after application s first run. Maintains a working set per process, not thread (remember scheduling is per thread) WS is based on size, not time window. Keeps a min/max per process If we have > max and a page fault then steal process s page. Otherwise add. ~ 1sec. Balance Set Manager, checks for sufficient free pages. Calls Working Set Manager to free pages. Sort processes by desirability. Large idle, first. Foreground last. If < min and enough page faults since last check, then leave alone. Determine #pages to remove from process. Depends on need, size of WS relative to min/max. Examine all pages in a process. Uses an unreferenced count, like AT&T Unix. Pages with high unreferenced removed. Continue examining additional processes till sufficient pages free. More aggressive passes as needed. 60 30
MM in NT Load Control. Swapper runs every 4 sec. If thread asleep for 7 sec. Mark thread s kernel stack in transition. When all threads in process marked, then swap out. 61 MM in NT 62 31
MM in FreeBSD 1GB for kernel, unless you want lots of small processes using lots of kernel services, then configure for 2GB First few virtual pages reserved to catch bad pointers. Shared libraries placed by default just below default stack limit. Memory Lists: Wired Active Inactive. Dirty (min: 0%, max 4.5%) Pageout daemon moves pages from Inactive to cache attempts to balance i/o load by not starting too many concurrent writes. Runs as a kernel process. This way it has access to kernel data structures, etc., but can be scheduled to run as a process so it can sleep. Cache. Clean. (min: 3%, max 6%) Free. (min: 0.7%, max 3%) One end is zeroed, the other is not. Idle process tries to keep 75% of frames on Free list zeroed. 63 MM in FreeBSD Page coloring. Attempt to avoid cache conflicts. Cache holds pages. A 1MB cache with 4KB pages has 256 cache pages. Each physical page on a 1MB boundary maps to the same cache page. Thus each cache page represents as many physical pages as there are MB s of physical memory. When allocating frames, color coding attempts to avoid such potential conflicts by making contiguous virtual pages, get frames that map to contiguous cache pages. Pure demand paging when swapping back in. Back in BSD4.3, count of resident pages was used. DIoF says might return. Page Replacement Least actively used algorithm. Page has count of three when first brought in Incremented each time reference bit found set to a max of 64 Decremented each time referenced bit found not set. At zero page moved from Active to Inactive list. 64 32
Segmentation (1) One-dimensional address space with growing tables One table may bump into another 65 Segmentation (2) Allows each table to grow or shrink, independently 66 33
Segmentation (3) Comparison of paging and segmentation 67 Implementation of Pure Segmentation 68 34
Segmentation with Paging: MULTICS (1) Descriptor segment points to page tables Segment descriptor numbers are field lengths 69 Segmentation with Paging: MULTICS (2) A 34-bit MULTICS virtual address 70 35
Segmentation with Paging: MULTICS (3) Conversion of a 2-part MULTICS address into a main memory address 71 Segmentation with Paging: MULTICS (4) Simplified version of the MULTICS TLB Existence of 2 page sizes makes actual TLB more complicated 72 36
Segmentation with Paging: Pentium (1) A Pentium selector 73 Segmentation with Paging: Pentium (2) Pentium code segment descriptor Data segments differ slightly 74 37
Segmentation with Paging: Pentium (3) Conversion of a (selector, offset) pair to a linear address 75 Segmentation with Paging: Pentium (4) Mapping of a linear address onto a physical address 76 38
Segmentation with Paging: Pentium (5) Level Protection on the Pentium 77 39