Memory Management -2

Paging
 Review of last class
●
CPU/Processors have Memory Management Unit
(MMU)
– Provides basic hardware mechanism for translating address
issued by execution unit, into another address
– Logical Address
– Physical Address
– The translation is done in hardware, as a part of execution of
every machine instruction
 Review of last class
●
Different types of MMUs
– No translation (no MMU)!
– Base + limit (also called relocation + offset) scheme
●
Leads to what type of OS ?
– Multiple base + limit scheme --> segmentation
●
Offers which benefits?
●
Problem: fragmentation
– Paging – logical division of the physical memory into equal sized pages, in
memory page table + PTBR in hardware, OS’s job and compiler’s job in this
case
 Dynamic Linking
●
Linker is normally invoked as a part of compilation process
– Links
●
function code to function calls
●
references to global variables with “extern” declarations
●
Dyanmic Linker
– Does not combine function code with the object code file
– Instead introduces a “stub” code that is indirect reference to actual code
– At the time of “loading” the program in memory, the “link-loader” (part of OS!)
will pick up the relevant code from the library machine code file (e.g. libc.so.6)
 Dynamic Loading
●
Loader
– Loads the program in memory
– Part of exec() code
– Needs to understand the format of the executable file (e.g. the ELF format)
●
Dynamic Loading
– Load a part from the ELF file only if needed during execution
– Delayed loading
– Needs a more sophisticated memory management by operating system – to be
seen during this series of lectures
 Dynamic Linking, Loading
●
Dynamic linking necessarily demands an advanced
type of loader that understands dynamic linking
– Hence called ‘link-loader’
– Static or dynamic loading is still a choice
●
Question: which of the MMU options will alllow for
which type of linking, loading ?
Continous memory management
 What is Continous memory management?
●
Entire process is hosted as one continous chunk
in RAM
●
Memory is typically divided into two partitions
– One for OS and other for processes
– OS most typically located in “high memory” addresses,
because interrupt vectors map to that location (Linux,
Windows) !
●
Hardware support needed: base + limit (or
relocation + limit)
 Problems faced by OS
●
Find a continuous chunk for the process being forked
●
Different processes are of different sizes
– Allocate a size parameter in the PCB
●
After a process is over – free the memory occupied by it
●
Maintain a list of free areas, and occupied areas
– Can be done using an array, or linked list
Variable partition scheme
 Problem: how to find a “hole” to fit in new
process
●
Suppose there are 3 free memory regions of
sizes 30k, 40k, 20k
●
The newly created process (during fork() +
exec()) needs 15k
●
Which region to allocate to it ?
 Problem: how to find a “hole” to fit in new
process
●
Best fit
●
Worst fit
●
First fit
●
 Problem : External fragmentation
●
Suppose there are 3 free memory regions of
sizes 30k, 40k, 20k
●
The newly created process (during fork() +
exec()) needs 50k
●
Total free memory: 30+40+20 = 90k
– But can’t allocate 50k !
 Solution to external fragmentation
●
Compaction !
●
OS moves the process chunks in memory to make available
continous memory region
– Then it must update the memory management information in PCB
(e.g. base of the process) of each process
●
Time consuming
●
Possible only if the relocation+limit scheme of MMU is
available
 Another solution to external
fragmentation
●
Fixed partition scheme
●
Memory is divided by OS into chunks of equal size: e.g., say, 50k
– If total 1M memory, then 20 such chunks
●
Allocate one or more chunks to a process, such that the total size is >= the size of
the process
– E.g.if request is 50k, allocate 1 chunk
– If request is 40k, still allocate 1 chunk
– If request is 60k, then allocate 2 chunks
●
Leads to internal fragmentation
– space wasted in the case of 40k or 60k requests above
Internal fragmentation
 Fixed partition scheme
●
OS needs to keep track of
– Which partition is free and which is used by which
process
– Free partitions can simply be tracked using a
bitmap or a list of numbers
– Each process’s PCB will contain list of partitions
allocated to it
 Solution to internal fragmentation
●
Reduce the size of the fixed sized partition
●
How small then ?
– Smaller partitions mean more overhead for the
operating system in allocating deallocating
Paging
 An extended version of fixed size
partitions
●
Partition = page
– Process = logically continuous sequence of bytes, divided in ‘page’
sizes
– Memory divided into equally sized page ‘frames’
●
Important distinction
– Process need not be continous in RAM
– Different page sized chunks of process can go in any page frame
– Page table to map pages into frames
Logical address seen as
Paging hardware
 MMU’s job
To translate a logical address generated by the CPU to a physical address:
1. Extract the page number p and use it as an index into the page table.
(Page table location is stored in a hardware register
Also stored in PCB of the process, so that it can be used to load the
hardware register on a context switch)
2. Extract the corresponding frame number f from the page table.
3. Replace the page number p in the logical address with the frame number f .
 Job of OS
●
Allocate a page table for the process, at time of fork()/exec()
– Allocate frames to process
– Fill in page table entries
●
In PCB of each process, maintain
– Page table location (address)
– List of pages frames allocated to this process
●
During context switch of the process, load the PTBR using the
PCB
 Job of OS
●
Maintain a list of all page frames
– Allocated frames
– Free Frames (called frame table)
– Can be done using simple linked list
– Innovative data structures can also be used to
maintain free and allocated frames list (e.g. xv6 code)
–
 Disadvantage of Paging
●
Each memory access results in two memory
accesses!
– One for page table, and one for the actual memory
location !
– Done as part of execution of instruction in hardware
(not by OS!)
– Slow down by 50%
 Speeding up paging
●
Translation
Lookaside
Buffer (TLB)
●
Part of CPU
hardware
●
A cache of
Page table
entries
●
Searched in
parallel for a
page number
 Speedup due to TLB
●
Hit ratio
●
Effective memory access time
– Hit ratio * 1 memory access time + miss ratio * 2
memory access time
●
 Memory
protection
with paging
X86 PDE and PTE
 Shared
pages (e.g.
library)
with paging
 Paging: problems
●
64 bit address
●
Suppose 20 bit offset
– That means 2^20 = 1 MB pages
– 44 bit page number: 2^44 that is trillion sized page
table!
– Can’t have that big continuous page table!
 Paging: problems
●
32 bit address
●
Suppose 12 bit offset
– That means 2^12 = 4 KB pages
– 20 bit page number: 2^20 that is a million entries
– Can’t always have that big continuous page table
as well, for each process!
Hierarchical
paging
More hierarchy
 Problems with hierarchical paging
●
More number of memory accesses with each
level !
– Too slow !
●
OS data structures also needed in that
proportion
Hashed page table
Inverted page table
Normal page table – one per
process --> Too much memory
consumed

Inverted page table : global

Examples of systems using

virtual address
consists of a triple:
<process-id, page-number,
offset>
Case Study: Oracle SPARC Solaris
●
64 bit SPARC processor , 64 bit Solaris OS
●
Uses Hashed page tables
– one for the kernel and one for all user processes.
– Each hash-table entry represents a contiguous area of mapped
virtual memory (set of continguous pages)
– Each entry has a base address
– and a span indicating the number of pages the entry represents
Case Study: Oracle SPARC Solaris
●
CPU implements a TLB that holds translation table entries
( TTE s) for fast hardware lookups.
●
A cache of these TTE s resides in a translation storage
buffer ( TSB ), which includes an entry per recently accessed
page
●
When a virtual address reference occurs, the hardware
searches the TLB for a translation.
●
If none is found, the hardware walks through the in memory
TSB looking for the TTE that corresponds to the virtual
address that caused the lookup
Case Study: Oracle SPARC Solaris
●
If a match is found in the TSB , the CPU copies the TSB
entry into the TLB , and the memory translation completes.
●
If no match is found in the TSB , the kernel is interrupted to
search the hash table.
●
The kernel then creates a TTE from the appropriate hash
table and stores it in the TSB for automatic loading into the
TLB by the CPU memory-management unit.
●
Finally, the interrupt handler returns control to the MMU ,
which completes the address translation and retrieves the
requested byte or word from main memory.
Swapping
 Swapping
●
Standard swapping
– Entire process swapped in or swapped out
– With continous memory management
●
Swapping with paging
– Some pages are “paged out” and some “paged in”
– Term “paging” refers to paging with swapping now
 Words of caution about ‘paging’
●
Not as simple as it sounds when it comes to
implementation
– Writing OS code for this is challenging