1. Memory Management
© 2004, D. J. Foreman

2. Building a Module -1 Compiler Linker or Link Editor
generates references for function addresses
may be inside current file
may be external - in a separate compile
addresses are 'marked' in compiler output - the "object" file
Linker or Link Editor
Resolves addresses (references)
Makes them relative to start of module
Saves relocatable module
© 2004, D. J. Foreman

3. Building a Module -2 Loader
Builds a table of addresses (ref table)
References
Definitions
replaces every relative address with sum of actual load address + relative address
gives control to the module
original relocatable file is still on disk
This is called "static binding" of the logical address to the physical address
© 2004, D. J. Foreman

4. Building a Module -3 Dynamic memory - malloc, etc
Linker examines all calls to malloc
Adds up requests (as much as possible)
Adds space for stack data
Loader assigns space for both
Heap space allocated from one end of block
Stack space allocated from other end
Grow toward middle
Ask O/S for more, when space is gone
© 2004, D. J. Foreman

5. Strategies Fixed-size partition Variable-size partition
Divide all RAM into N blocks
Program size must be < block size
Internal fragments exist - cannot be used
Variable-size partition
Blocks defined as needed
Must keep track of blocks
External fragmentation occurs and gets worse
Smaller holes left, as processes fill old holes
Favors smaller and smaller processes
Eventually must move processes
© 2004, D. J. Foreman

6. Managing fragments Move programs - compaction Allocation strategies
Best fit - search for block with closest size
Worst fit - find biggest block - leaves largest hole when program terminates
First fit - use first hole in list, from front
Next fit - circular list, just use next entry in list
Part of block points to next
Helps manage 'garbage collection' - the process of building big blocks from little ones
© 2004, D. J. Foreman

7. Current Strategies All use variable-sized partitions
Most use fixed sized blocks within the partition - will be called pages
Simplifies list mgmt
Other considerations:
Program segments don't (usually) change size
If pgm seg changes size, pgm must be moved
Loader must re-bind - increases overhead
© 2004, D. J. Foreman

8. Dynamic Binding All addresses are relative to "0"
Adjust all addresses as in static binding
Repeat when moving program
Can be done with h/w relocation register
Simple relocation done easily
h/w adjusts every memory reference
Relocation value added as needed
© 2004, D. J. Foreman

9. Improvements Multiple relocation registers
Code segment register - during fetch cycle
Stack segment register - for stack instructions
Data segment register - during execute cycle
Automatic relocation-register management
Compiler generates code to update register when references cross segment boundary
No auto-help for
Assembler programs
Programs with data segments > register capability
Programs that change segment registers
© 2004, D. J. Foreman

10. Making it work Trusted s/w for register manipulation
Privileged instructions
Compiler generates a segment trap
NOT a violation trap
O/S changes the relocation registers
O/S changes the Program Counter
Also privileged
Program proceeds
© 2004, D. J. Foreman

11. Isolation Need to ensure access protection "Limit register" added
Contains length of segment
Compute: segment register + relative address
If relative address <= limit then proceed
Else segment fault (protection error)
© 2004, D. J. Foreman

12. Swapping Swapping removes the entire address space Cost analysis:
If a process has S memory space units
And a disk block holds D memory units
K=Ceil (S/D) disk writes required to save
And K reads to restore it later = 2*K overhead
If Pi is blocked for T time, cost = T*S
T must be > X (assume 1 I/O = 1 Time unit)
© 2004, D. J. Foreman

13. Another strategy Swapping removes an entire process
What if we only remove specific blocks?
Must recognize spatial locality
Sometimes called locality of reference
We can remove/restore selected pages
© 2004, D. J. Foreman

14. Virtual Memory Mgmt Responsibilities: Problems: Determine locality
Keep referenced pages in memory
Problems:
Partitioning
Scatter-loading
Re-binding
© 2004, D. J. Foreman

15. Shared Memory Multiprocessors need to share
Increases bus load, reduces speed
Speedup via cache, but new problems:
Stale data (coherency)
Expense
Interconnection networks:
More speed
More expense
Must detect changes everywhere
Must force cache update synch
© 2004, D. J. Foreman