1. Chapter Twelve Memory Organization

2. Memory Hierarchy Magnetic tapes Main memory I/O processor Magnetic
discs
Auxiliary memory
Cache
CPU

3. Auxiliary vs. Main Memory
Provides backup storage
Low-cost
Large capacity
Holds data file, system programs, etc not currently needed
No direct access by CPU
Holds programs and data currently needed
Smaller capacity
Direct access by CPU
Higher cost
Faster access, but still not fast enough…

4. Cache Memory Very-high speed memory Smaller capacity, higher cost
Access time is close to processor logic clock cycle time
Stores program segments currently being executed and data frequently accessed
Increases performance of computer

5. Memory Hierarchy
• Get the performance of fast expensive memory for the price of slow cheap memory!
• GP Registers (2-5 ns)
• Cache
• Level 1 Cache (2-10 ns)
• Level 2 Cache (5-20 ns)
• Main Memory (40–80 ns)
• Disk (10 ms seek, Mb/s throughput)

6. Foundations • Convention: Registers are lowest, Disk is highest
• Inclusion: Data found at one level of the hierarchy is also found at all higher levels
• A miss at level i implies data is not at any lower level
• Coherence: Copies of data at multiple levels of the hierarchy must (eventually) be consistent
• Write through
• Write back
• Locality
• Temporal: Recently accessed items are likely to be accessed
• Spatial: Items that are close in address are likely to be accessed

7. Basic Operation of Cache
When CPU needs to access memory, first checks cache
If found, then it is a hit -- data is read
If not found, then it is a miss -- access main memory and transfer block to cache

8. Metrics
Hit ratio: hi is the probability that a datum is present at at level i
• Access frequency:
fi = (1-h1)(1-h2) …. (1-hi-1)hi
Effective access time:
Teff = sum fi*ti

9. Goals Performance = level 1 Cost per bit = level n
Overall performance  effective hierarchical access time per memory reference is small
Foundation is based on locality

10. Mapping Procedures Associative mapping – fastest and most flexible
Direct mapping – most rigid
Set-associative mapping – compromise between the two

11. Replacement Policies Choose the victim: • Least Recently Used (LRU)
• Least Frequently Used (LFU)
• First In First Out (FIFO)
• Random
• Optimal

12. Analysis of Write-Back
Simple Write-Back – all replaced pages are written back to disk
Average reference time is
Tm + 2(1-h)Tp
Where
Tm is the memory access time for one word
Tp is the page transfer time
2  one for loading and a for writing back

13. Flagged Write Back Use dirty bit, set when a page is modified
Write back only if it is dirty
Average reference time = Tm + (1-h) (Tp +Wp * Tp)
Where Wp is the probability that a page is modified

14. Write-through Update main memory and disk same time
Main memory always has the same data as disk
Page is loaded on a write-miss and read miss
Ave. time = Tm + (1-h)Tp + Wt (Td-Tm)
Where Wt is the fraction of writes and Td is the disk access time