Memory Management

1
Memory Management

• Operating Systems - Spring 2003
• Gregory (Grisha) Chokler
• Ittai Abraham
• Zinovi Rabinovich

2
Background

• A compiler creates an executable code
• An executable code must be brought into main
  memory to run
  
• Operating system maps the executable code onto
  the main memory
  
• The mapping depends on the hardware
• Hardware accesses the memory locations in the
  course of the execution

3
Memory management

• Principal operation Bringing programs into the
  memory for execution by the processor
  
• Address Binding
• Compile time, e.g. MS-DOS .com programs
• Load time relocatable code
• Execution time moved around in memory

4
Memory management (II)

• Requirements and issues
• Relocation
• Protection
• Sharing
• Logical and physical organization

5
Relocation

• Compiler generated addresses are relative
• A process can be (dynamically) loaded at
  different memory location
  
• Actual physical mapping is not known at compile
  time
  
• Simple if processes are allocated contiguous
  memory ranges
  
• Complicated otherwise,
• e.g. segmentation or paging is used

6
Protection

• Processes must be unable to reference addresses
  of other processes or those of the operating
  system
  
• Cannot be enforced at compile time (especially if
  relocation is present)
  
• Each memory access must be checked for validity
• Enforced by hardware

7
Sharing

• Protection must be flexible enough to allow
  controlled sharing of the same portion of the
  main memory
  
• E.g., shared libraries load once into main
  memory and allow several programs access
  simultaneously

8
Logical organization

• Memory is organized as a linear one-dimensional
  address space
  
• Programs are typically organized into modules
• can be written/compiled independently, have
  different degrees of protection, shared
  
• A mechanism is desirable for supporting a logical
  structure of the user programs

9
Physical organization

• Two level hierarchical storage
• Main memory fast, expensive, scarce, volatile
• Secondary storage slow, cheap, abundant,
  persistent
  
• Memory management involves 2-directional
  information flow between the levels
  
• Swapping

10
Memory management (III)

• Management How physical and virtual address
  spaces of multiple processes correlate?
  
• Allocation of physical memory for virtual space
  intervals
  
• Placement algorithms
• Replacement algorithms

11
Allocator Issues

• Patterns of memory reuse
• Reuse preference for recently-freed intervals
• Preference for intervals in certain regions
• Splitting and Coalescing
• Are large free intervals split or allocated as a
  whole?
  
• Are adjacent free intervals unified into one?
• Fits
• If several possibilities exist for free intervals
  allocation, which is used?
  
• Splitting thresholds
• If some portion of allocated interval is not used
  by virtual counterpart, will the remainder become
  separate free interval ?

12
More Problems

• Fragmentation inability to reuse memory that is
  free
  
• External fragmentation there are free unused
  fragments in physical memory outside allocated
  intervals
  
• Internal fragmentation there is free unused
  fragment space inside allocated interval

13
Fragmentation Sources

• Isolated deaths
• Should allocator place in memory objects that
  die (expire) together, fragmentation would be
  low
  
• Time-varying behavior
• For example free small blocks, request larger
  ones
  
• Ramps/Peaks/Plateaus
• Should these changes follow a known pattern,
  allocations could be optimized

14
Memory management Allocation

• Partitioning now obsolete, but useful for
  demonstration of basic principles
  
• Fixed
• Dynamic
• By list of intervals
• By hierarchy of intervals
• Paging and segmentation
• without and with virtual memory

15
Allocation by Fixed partitioning

• Main memory is divided into a number of static
  partitions
  
• all partitions of the same size
• a collection of different sizes
• A process can be loaded into a partition of equal
  or greater size
  
• A process cannot be scattered among many
  partitions

16
Fixed partitioning - Problems

• Internal fragmentation
• wasted space is internal to an allocated region
• Executing big programs requires explicit memory
  management by the programmer
  
• overlays

17
Fixed Partitioning Placement algorithms
Operating System
Operating System
New Processes
New Processes
18
Fixed Partitioning Replacement algorithm

• If all partitions are occupied some process
  should be swapped out
  
• Select a process that
• occupies the smallest partitions that will hold
  the incoming process
  
• blocked process

19
Allocation by Dynamic partitioning

• Partitions created dynamically
• Each process is loaded into a partition of
  exactly the same size as that process
  
• What data structure and How is it used to support
  information about free physical memory intervals?

20
Dynamic partitioning example
Operating System
128 K
224 K
896 K
352 K
21
Dynamic partitioning example
Operating System
Operating System
320 K
320 K
Process 1
Process 1
224 K
224 K
Process 2
Process 3
288 K
Process 3
288 K
64 K
64 K
22
Dynamic partitioning example
Operating System
Operating System
320 K
Process 2
224 k
96 K
128 K
128 K
Process 4
Process 4
96 K
96 K
Process 3
288 K
Process 3
288 K
64 K
64 K
23
Problem External fragmentation

• Dynamic partitioning is subject to external
  fragmentation
  
• situation when the memory which is external to
  the allocated partitions becomes fragmented
  
• Can be reduced using compaction
• wastes the processor time

24
Placement algorithms

• Free regions are organized into a linked list
  ordered by the memory addresses
  
• simplifies coalescing of free regions
• increases insertion time
• First-fit allocate the first region large enough
  to hold the process
  
• optimization use a roving pointer next-fit
• Best-fit allocate the smallest region which is
  large enough to hold the process

25
An example
8K
8K
12K
12K
First Fit
22K
6K
Best Fit
Last allocated block (14K)
18K
2K
8K
8K
6K
6K
Allocated block
14K
Free block
14K
Next Fit
36K
20K
Before
After
26
Placement algorithms discussion

• First-fit
• efficient and simple
• tends to create more fragmentation near the
  beginning of the list
  
• solved by next-fit
• Best-fit
• less efficient (might search the whole list)
• tends to create small unusable fragments

27
Buddy systems

• Hierarchy of free intervals
• Approximates the best-fit principle
• Efficient search
• first (best) fit has a linear complexity
• Simplified coalescing

28
Buddy Systems Variations

• Binary buddies
• system allocates memory in blocks which are
  powers of 2
  
• Fibonacci buddies
• Block sizes are Fibonacci numbers, recall every
  Fibonacci number is sum of previous two
  
• Weighted buddies
• Block sizes are either power of two, or three
  times the power of two 2,3,4,6,8,12
  
• Double buddies
• Two different buddy system with staggered sizes,
  e.g. one with 2,4,8,16 and second with 3,6,12,24

29
Binary buddy system

• Memory blocks are available in size of 2K where L
  
• 2L smallest size of block allocated
• 2U largest size of block allocated(generally,
  the entire memory available)
  
• Initially, all the available space is treated as
  a single block of size 2U

30
Creating buddies

• A request to allocate a block of size s
• If 2U-1
• Otherwise, split the block into two equal buddies
  of size 2U-1
  
• If 2U-2 one of the buddies, otherwise split further
  
• Proceed until the smallest block is allocated

31
Maintaining buddies

• Existing non-allocated buddies (holes) are kept
  on several lists
  
• Holes of size 2i are kept on the ith list
• Hole is removed from the (i1)th list by
  splitting it into two bodies and putting them
  onto the ith list
  
• Whenever two adjacent buddies on the ith list are
  freed they are coalesced and moved to the (i1)th
  list

32
Finding a free buddy

• get_hole(i)
• if (iU1) then failure
• if (list i is empty)
• get_hole(i1)
• split hole into buddies
• put buddies on list i
• take first hole on list i

33
An example
34
Remarks

• Buddy system is a reasonable compromise between
  the fixed and dynamic partitioning
  
• Internal fragmentation is a problem
• Expected case is about 28 which is high
• Buddy System is not used for (system) memory
  management nowadays
  
• Used for memory management by user level
  libraries (malloc)

35
Relocation with partitioning
Relative address
Base Register
Text
Adder
Absolute address
Bounds Register
Comparator
Data
Interrupt to operating system
Stack
Process image in main memory
36
Paging

• The most efficient and flexible method of memory
  allocation
  
• Process memory is divided into fixed size chunks
  of the same size, called pages
  
• Pages are mapped onto frames in the main memory
• Internal fragmentation is possible with paging
  (negligible)

37
Paging support

• Process pages can be scattered all over the main
  memory
  
• Process page table maintains mapping of process
  pages onto frames
  
• Relocation becomes complicated
• Hardware support is needed to support translation
  of relative addresses within a program into the
  memory addresses

38
Paging example
0
0
0
---
1
1
1
---
2
2
2
---
3
3
Process B
Process A
0
7
4
13
1
8
5
14
2
9
6
Free Frame List
3
10
11
12
Process C
Process D
39
Address translation

• Page (frame) size is a power of 2
• with page size 2r, a logical address of lr
  bits is interpreted as a tuple (l,r)
  
• l page number, r offset within the page
• Page number is used as an index into the page
  table

40
Hardware support
Logical address
Page
Offset
Frame
Offset
Register
Page Table Ptr
Page Table
Offset
Page Frame
P

Frame
Program
Paging
Main Memory
41
Segmentation

• A program can be divided into segments
• Segments are of variable size
• Segments reflect the logical (modular) structure
  of a program
  
• Text segment, data segment, stack segment
• Similar to dynamic partitioning except segments
  of the same process can be scattered
  
• subject to external fragmentation

42
Address translation

• Maximum segment size is always a power of 2
• Process segment table maps segment numbers into
  their base addresses in the memory
  
• With the maximum segment size of 2r, a logical
  address of lr bits is interpreted as a pair
  (l,r)
  
• l segment number, r offset within the
  segment
  
• l is used as an index into the segment table

43
Hardware support
Virtual Address
Segment Table

Base d
Seg
Offset d
Register
Seg Table Ptr
Segment Table
d
Segment
S

Length Base
Program
Segmentation
Main Memory
44
Remarks

• The price of paging/segmentation is a
  sophisticated hardware
  
• But the advantages exceed by far
• Paging decouples address translation and memory
  allocation
  
• Not all the logical addresses are necessary
  mapped into physical memory at every given moment
  - virtual memory
  
• Paging and segmentation are often combined to
  benefit from both worlds