Virtual Memory

CSE 410 Spring 2009CSE 410, Spring 2009Computer Systems

http://www.cs.washington.edu/410

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 1

Reading and Referencesg

• Reading • Computer Organization and Design, Patterson and Hennessy

» Section 5.4 Virtual Memory» Section 5.5 A Common Framework for Memory Hierarchies

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 2

Memory Management Goalsy g

We want to share main memory such that:y• Each process thinks it has a private memory of

2-4 GB (or more), even if it doesn’t use it all2 4 GB (or more), even if it doesn t use it all• Real memory is allocated efficiently to parts of

process memory actually being used (locality)process memory actually being used (locality)• No process can interfere with or even see

b l i t thmemory belonging to another» Unless we want that to happen

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 3

Layout of program memoryy p g y

7FFF EFFF stack (grows down)

reserved (4KB)7FFF FFFF

stack (grows down)

~1792 MB

1000 FFFF

Not toScale!heap (grows up)1001 0000

0FFF FFFF1000 00001000 FFFF

program (252 MB)

global data (64 KB)

reserved (4 MB)0000 0000

003F FFFF

0040 0000

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 4

The Big Ideag

• Separate program notion of memory addresses p p g yfrom actual physical memory locations» Program memory = virtual addressesg y» Physical memory = real addresses» Use hardware to map between the two» Use hardware to map between the two

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 5

Memory Mappingy pp g

programdd

physicaldd

memoryi

heap

stackaddresses addressesmapping

programphysical

memorystack

heapprogram

stack

heapprogram

stack

disk

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 6

Virtual Physical Paging01

01

Virtual Page #

Physical Page #

0x0000 0x0000

• Divide a process's virtual address space into fixed-size chunks (called pages)

1234

1234size chunks (called pages)

• Divide physical memory into pages of the same sizeA i t l b

45

4567

0x6000

0 0000• Any virtual page can be located at any physical page

012

78910

0x0000

• Translation box converts from virtual pages to physical pages

310111213Translation

0x4000

0 00013 0xE000

Multiple Processes Vi t l Ph i l Multiple ProcessesShare Memory

01

01

Virtual Page #

Physical Page #

0x0000 0x0000

• Each process thinks it t t t dd

1234

1234starts at address

0x0000 and has all of memory

45

4567

0x6000

memory• A process doesn't

know anything about

012

78910

0x0000

know anything about physical addresses and doesn't care

310111213Translation

0x4000

and doesn t care 13 0xE000

Protection

• A process can only use its 0 0

Virtual Page #

Physical Page #

0x0000 0x0000p yown virtual addresses

• A process can't corrupt

123

123

another process's memory» It has no address to refer to it

45

456

0x6000

• How can Blue write to Green's page 2?

d dd f

012

789

0x0000

» needs an address to refer to physical page 7, but it doesn't have one

23

101112

Translation

0x4000

13Translation0xE000

Sparse AddressSpaces

0

Virtual Page #

Physical Page #

0x0000 0x00000• Memory addresses that

aren't being used at all don't have to be

0123

0x0000 0x00000123don't have to be

assigned real addresses» Code can start at a very

3456

30x4000

Unusedylow logical address

» Stack can start at a very high logical address

6789 0xA000

997998

0x0FFC000

high logical address» No physical pages

allocated for unused addresses in bet een

9

Translation0x1000000

0xA0009991000

addresses in between

Sh i M Virtual Physical Sharing Memory• Two processes can share 0

1

Virtual Page #

Physical Page #

0x0000 0x000001

memory by mapping two virtual pages to the same physical page

1234

1234physical page

• The code for Word can be shared for two Word

45

Word0x6000

00x0000

4567

processes» code pages are read only

h h i

0123

0x0000 78910• Each process has its own

data pages» possible to share data pages Translation

0 000

345

W d0x6000

10111213» possible to share data pages

too, but less common0xE000Word 13

Store Memory on DiskStore Memory on Disk

• Memory that isn't being 0

Virtual Page #

Physical Page #

0x0000 0x00000used can be saved on disk» swapped back in when it is

referenced via page fault

0123

0123referenced via page fault

• Programs can address more memory than is

456

456

physically available• This is one important

f i t l

789 0xA000

789reason for virtual memory

» too hard for programs to do this on their own

101112

Translation0xE000 Disk13

Memory Hierarchy Revisitedy y

• Once the translation hardware is there we have a caching problem again» Want size ≈ disk, performance ≈ memory

K i di k l i 100 000 i• Key issue: disk latency is 100,000 times memory, so design motivation is to avoid accessing disk

• Minimizing miss rate (“page faults”):• Minimizing miss rate (“page faults”):» VM “pages” are much larger than cache blocks = size

of disk blocks, usually 4K or 8K or more, y» Use fully associative lookup with approximate LRU» Question: should it be write-back or write-through?

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 13

Finding the Right Page (frame)g g g ( )• If fully associative, how do we find the right page

ith t i ll f ?without scanning all of memory?• Answer: index is called the page table

» Each process has a separate page table» Each process has a separate page table• Processor “page table register” points to active one – part of

process state» Page table indexed with virtual page number (VPN)» Page table indexed with virtual page number (VPN)

• The bits that aren’t part of the page offset» Each entry contains a valid bit and a physical page

b (PPN)number (PPN)• PPN is concatenated with page offset to get physical address

» No index tag needed – full VPN is index

5/26/2009 cse410-14-virtual-memory © 2006-09, Perkins, DW Johnson and University of Washington 14

Page Table pictureg pVirtual address

Page table register

31 30 29 28 27 15 14 13 12 11 10 9 8 3 2 1 0

Page offsetVirtual page number

Physical page numberValid

20 12

Page tableg

18

Page offsetPhysical page number

If 0 then page is not�present in memory

29 28 27 15 14 13 12 11 10 9 8 3 2 1 0

15

Physical address

How big is the page table?g p g

• From the previous slide:p» Virtual page number is 20 bits.» Physical page number is 18 bits + valid bit -> y p g

round up to 32 bits.• Or 20 bits + valid bit if 32-bit physical addressing

16

Dealing with large page tablesg g p g• Multi-level page tables

» “Any problem in CS can be solved by adding a level of indirection”or two…

Page Table Base Pointer

1st

2nd

3rd

A 3-level page table

VPN1 VPN2 VPN3 ff t

PPN

PPN offset

• Since most processes don’t use the whole address space, you don’t allocate the tables that aren’t needed

VPN1 VPN2 VPN3 offset

17

» Also, the 2nd and 3rd level page tables can be “paged” to disk.

Waitaminute!• We’ve just replaced every memory access MEM[addr] with:

MEM[MEM[MEM[MEM[PTBR VPN1 2] VPN2 2] VPN3 2] ff ]MEM[MEM[MEM[MEM[PTBR + VPN1<<2] + VPN2<<2] + VPN3<<2] + offset]

» i.e., 4 memory accesses

A d h ’t t lk d b t th b d t (i f lt )• And we haven’t talked about the bad case yet (i.e., page faults)…

“A bl i CS b l d b ddi l l f i di i ”“Any problem in CS can be solved by adding a level of indirection”» except too many levels of indirection…

H d d l ith t l l f i di ti ?• How do we deal with too many levels of indirection?

18

Caching Translationsg• Virtual to Physical translations are cached in a Translation Lookaside

Buffer (TLB).Page offsetVirtual page number

Virtual address

1220

31 30 29 15 14 13 12 11 10 9 8 3 2 1 0

Physical page numberValid Dirty Tag

TLB hit

TLB

Page offset

20

Cache index Byte�

Physical page number

Physical address tag

�

�

Physical address

Valid Tag Data

16 14

Cache index

2offsetPhysical address tag

32

Cache

19

DataCache hit

What about a TLB miss?

• If we miss in the TLB, we need to “walk the page t bl ”table”» In MIPS, an exception is raised and software fills the TLB» In x86, a “hardware page table walker” fills the TLB» In x86, a hardware page table walker fills the TLB

• What if the page is not in memory?» This situation is called a page fault.» The operating system will have to read the page from disk.» It will need to select a page to replace» It will need to select a page to replace.

• The O/S tries to approximate LRU (coming next)» The replaced page will need to be written back if dirty.

20

Summaryy• Virtual memory is great:

» It means that we don’t have to manage our own memory» It means that we don t have to manage our own memory.» It allows different programs to use the same physical memory.» It provides protect between different processes.

It ll t ll d h i b t ( lb it h t» It allows controlled sharing between processes (albeit somewhat inflexibly).

• The key technique is indirection:Y t th l i CS t i k ’ i thi l» Yet another classic CS trick you’ve seen in this class.

» Many problems can be solved with indirection.• Caching made a few appearances, too:

» Virtual memory enables using physical memory as a cache for disk.

» We used caching (in the form of the Translation LookasideBuffer) to make Virtual Memory’s indirection fast

21

Buffer) to make Virtual Memory s indirection fast.