CENG 3420

Computer Organization & Design

Lecture 17: Cache-3 Discussions
Bei Yu
CSE Department, CUHK
[email protected]

(Textbook: Chapters 5.3–5.4)

Spring 2022
 Overview

1 Example 1

2 Example 2

3 Example 3

4 Performance Issues

2/28
Example 1
 Cache Example

short A[10][4];
int sum = 0; • Assume separate instruction and data caches
int j, i;
double mean; • So we consider only the data
// forward loop • Cache has space for 8 blocks
for (j = 0; j <= 9; j++)
sum += A[j][0];
• A block contains one word (byte)
mean = sum / 10.0;
• A[10][4] is an array of words located at
// backward loop
7A00-7A27 in row-major order
for (i = 9; i >= 0; i--)
A[i][0] = A[i][0]/mean;

4/28
Cache Example

Memory word
address in hex Memory word address in binary Array Contents (40 elements)

(7A00) 0 1 1 1 1 0 1 0 0 0 0 0 0 0 0 0 A[0][0]
(7A01) 0 1 1 1 1 0 1 0 0 0 0 0 0 0 0 1 A[0][1]
(7A02) 0 1 1 1 1 0 1 0 0 0 0 0 0 0 1 0 A[0][2]
(7A03) 0 1 1 1 1 0 1 0 0 0 0 0 0 0 1 1 A[0][3]
(7A04) 0 1 1 1 1 0 1 0 0 0 0 0 0 1 0 0 A[1][0]

(7A24) 0 1 1 1 1 0 1 0 0 0 1 0 0 1 0 0 A[9][0]
(7A25) 0 1 1 1 1 0 1 0 0 0 1 0 0 1 0 1 A[9][1]
(7A26) 0 1 1 1 1 0 1 0 0 0 1 0 0 1 1 0 A[9][2]
(7A27) 0 1 1 1 1 0 1 0 0 0 1 0 0 1 1 1 A[9][3]

Tag for Direct Mapped

8 blocks in cache, 3 bits encodes cache block number
Tag for Set-Associative
4 blocks/ set, 2 cache sets, 1 bit encodes cache set number
Tag for Associative

To simplify discussion: 16-bit word (byte) address; i.e. 1 word = 1 byte.

5/28
Direct Mapping
• Least significant 3-bits of address determine location
• No replacement algorithm is needed in Direct Mapping
• When i == 9 and i == 8, get a cache hit (2 hits in total)
• Only 2 out of the 8 cache positions used
• Very inefficient cache utilization

Content of data cache after loop pass: (time line)

j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8 j=9 i=9 i=8 i=7 i=6 i=5 i=4 i=3 i=2 i=1 i=0

0 A[0][0] A[0][0] A[2][0] A[2][0] A[4][0] A[4][0] A[6][0] A[6][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[6][0] A[6][0] A[4][0] A[4][0] A[2][0] A[2][0] A[0][0]
1
2
Cache 3
Block
number 4 A[1][0] A[1][0] A[3][0] A[3][0] A[5][0] A[5][0] A[7][0] A[7][0] A[9][0] A[9][0] A[9][0] A[7][0] A[7][0] A[5][0] A[5][0] A[3][0] A[3][0] A[1][0] A[1][0]

5
6
7

Tags not shown but are needed.

6/28
Associative Mapping

• LRU replacement policy: get cache hits for i = 9, 8, . . . , 2

• If i loop was a forward one, we would get no hits!

Content of data cache after loop pass: (time line)

j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8 j=9 i=9 i=8 i=7 i=6 i=5 i=4 i=3 i=2 i=1 i=0

0 A[0][0] A[0][0] A[0][0] A[0][0] A[0][0] A[0][0] A[0][0] A[0][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[0][0]

1 A[1][0] A[1][0] A[1][0] A[1][0] A[1][0] A[1][0] A[1][0] A[1][0] A[9][0] A[9][0] A[9][0] A[9][0] A[9][0] A[9][0] A[9][0] A[9][0] A[9][0] A[1][0] A[1][0]

2 A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0] A[2][0]

Cache 3 A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0] A[3][0]
Block
number 4 A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0] A[4][0]

5 A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0] A[5][0]

6 A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0]

7 A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0]

Tags not shown but are needed; LRU Counters not shown but are needed.
7/28
Set Associative Mapping
• Since all accessed blocks have even addresses (7A00, 7A04, 7A08, ...), only
half of the cache is used, i.e. they all map to set 0
• LRU replacement policy: get hits for i = 9, 8, 7 and 6
• Random replacement would have better average performance
• If i loop was a forward one, we would get no hits!

Content of data cache after loop pass: (time line)

j=0 j=1 j=2 j=3 j=4 j=5 j=6 j=7 j=8 j=9 i=9 i=8 i=7 i=6 i=5 i=4 i=3 i=2 i=1 i=0

0 A[0][0] A[0][0] A[0][0] A[0][0] A[4][0] A[4][0] A[4][0] A[4][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[8][0] A[4][0] A[4][0] A[4][0] A[4][0] A[0][0]
1 A[1][0] A[1][0] A[1][0] A[1][0] A[5][0] A[5][0] A[5][0] A[5][0] A[9][0] A[9][0] A[9][0] A[9][0] A[9][0] A[5][0] A[5][0] A[5][0] A[5][0] A[1][0] A[1][0]
Set 0
2 A[2][0] A[2][0] A[2][0] A[2][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[6][0] A[2][0] A[2][0] A[2][0]
Cache 3 A[3][0] A[3][0] A[3][0] A[3][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[7][0] A[3][0] A[3][0] A[3][0] A[3][0]
Block
number 4
5
Set 1
6
7

Tags not shown but are needed; LRU Counters not shown but are needed.
8/28
Comments on the Example

• In this example, Associative is best, then Set-Associative, lastly Direct Mapping.

• What are the advantages and disadvantages of each scheme?

• In practice,
• Low hit rates like in the example is very rare.
• Usually Set-Associative with LRU replacement scheme is used.

• Larger blocks and more blocks greatly improve cache hit rate, i.e. more cache
memory

9/28
Example 2
Question:
How many total bits are required for a direct-mapped cache with 16 KiB of data and
4-word blocks, assuming a 32-bit address?

11/28
Question:
How many total bits are required for a direct-mapped cache with 16 KiB of data and
4-word blocks, assuming a 32-bit address?

Answer:
• In a 32-bit address CPU, 16 KiB is 4096 words.
• With a block size of 4 words, there are 1024 blocks.
• Each block has 4 × 32 or 128 bits of data plus a tag, which is (32 − 10 − 2 − 2) = 18
bits, plus a valid bit.
• Thus, the total cache size is 210 × (4 × 32 + 18 + 1) = 210 × 147 bits.
147
• the total number of bits in the cache is about 1.15 = times as many as needed
32 × 4
just for the storage of the data.

11/28
Question:
How many total bits are required for an associated-mapped cache with 16 KiB of data and
4-word blocks, assuming a 32-bit address?

12/28
Question:
How many total bits are required for an associated-mapped cache with 16 KiB of data and
4-word blocks, assuming a 32-bit address?

Answer:
• In a 32-bit address CPU, 16 KiB is 4096 words.
• With a block size of 4 words, there are 1024 blocks.
• Each block has 4 × 32 or 128 bits of data plus a tag, which is (32 − 2 − 2) = 28 bits,
plus a valid bit.
• Thus, the total cache size is 210 × (4 × 32 + 28 + 1) = 210 × 157 bits.
157
• the total number of bits in the cache is about 1.27 = times as many as needed
32 × 4
just for the storage of the data.

12/28
Example 3
Question
We have designed a 64-bit address direct-mapped cache, and the bits of address used to
access the cache are as shown below:

Table: Bits of the address to use in accessing the cache

Tag Index Offset

63-10 9-5 4-0

1 What is the block size of the cache in words?

2 Find the ratio between total bits required for such a cache design implementation
over the data storage bits.
3 Beginning from power on, the following byte-addressed cache references are
recorded as shown below.

Table: Recored byte-addressed cache references

Hex 00 04 10 84 E8 A0 400 1E 8C C1C B4 884

Dec 0 4 16 132 232 160 1024 30 140 3100 180 2180

Find the hit ratio.

14/28
1 Each cache block consists of four 8-byte words. The total offset is 5 bits. Three of
those 5 bits is the word offset (the offset into an 8-byte word). The remaining two bits
are the block offset. Two bits allows us to enumerate 22 = 4 words.
2 The ratio is 1.21. The cache stores a total of
32lines × 4words/block × 8bytes/word = 1024bytes = 8192bits. In addition to the
data, each line contains 54 tag bits and 1 valid bit. Thus, the total bits required is
8192 + 54 × 32 + 1 × 32 = 9952 bits.
4
3 The hit ratio is 12 = 33%

15/28
Performance Issues
Q1: Where A Block Be Placed in Upper Level?

Scheme name # of sets Blocks per set

Direct mapped # of blocks 1
# of blocks
Set associative Associativity Associativity
Fully associative 1 # of blocks

17/28
Q1: Where A Block Be Placed in Upper Level?

Scheme name # of sets Blocks per set

Direct mapped # of blocks 1
# of blocks
Set associative Associativity Associativity
Fully associative 1 # of blocks

Q2: How Is Entry Be Found?

Scheme name Location method # of comparisons
Direct mapped Index 1
Set associative Index the set; compare set’s tags Degree of associativity
Fully associative Compare all tags # of blocks

17/28
Q3: Which Entry Should Be Replaced on a Miss?
• Direct mapped: only one choice
• Set associative or fully associative:
• Random
• LRU (Least Recently Used)

Note that:
• For a 2-way set associative, random replacement has a miss rate 1.1× than LRU
• For high level associativity (4-way), LRU is too costly

18/28
Q4: What Happen On A Write?
• Write-Through:
• The information is written in both the block in cache & the block in lower level
of memory
• Combined with write buffer, so write waits can be eliminated
•
L
L:
• :

• Write-Back:
• The information is written only to the block in cache
• The modification is written to lower level, only when the block is replaced
• Need dirty bit: tracks whether the block is clean or not
• Virtual memory always use write-back
•
L
:
•
L
:

19/28
Q4: What Happen On A Write?
• Write-Through:
• The information is written in both the block in cache & the block in lower level
of memory
• Combined with write buffer, so write waits can be eliminated
•
L
L: read misses don’t result in writes
• : easier to implement

• Write-Back:
• The information is written only to the block in cache
• The modification is written to lower level, only when the block is replaced
• Need dirty bit: tracks whether the block is clean or not
• Virtual memory always use write-back
•
L
:
•
L
:

19/28
Q4: What Happen On A Write?
• Write-Through:
• The information is written in both the block in cache & the block in lower level
of memory
• Combined with write buffer, so write waits can be eliminated
•
L
L: read misses don’t result in writes
• : easier to implement

• Write-Back:
• The information is written only to the block in cache
• The modification is written to lower level, only when the block is replaced
• Need dirty bit: tracks whether the block is clean or not
• Virtual memory always use write-back
•
L
: write with speed of cache
•
L
: repeated writes require only one write to lower level

19/28
 Performance Consideration

Performance
How fast machine instructions can be brought into the processor and how fast they can be
executed.

• Two key factors are performance and cost, i.e., price/performance ratio.
• For a hierarchical memory system with cache, the processor is able to access
instructions and data more quickly when the data wanted are in the cache.
• Therefore, the impact of a cache on performance is dependent on the hit and miss
rates.

20/28
 Cache Hit Rate and Miss Penalty

• High hit rates over 0.9 are essential for high-performance computers.

• A penalty is incurred because extra time is needed to bring a block of data from a
slower unit to a faster one in the hierarchy.

• During that time, the processor is stalled.

• The waiting time depends on the details of the cache operation.

Miss Penalty
Total access time seen by the processor when a miss occurs.

21/28
 Miss Penalty

Example: Consider a computer with the following parameters:

Access times to the cache and the main memory are t and 10t respectively. When a cache
miss occurs, a block of 8 words will be transferred from the MM to the cache. It takes 10t
to transfer the first word of the block and the remaining 7 words are transferred at a rate
of one word per t seconds.

• Miss penalty = t + 10t + 7 × t + t

• First t: Initial cache access that results in a miss.
• Last t: Move data from the cache to the processor.

22/28
Average Memory Access Time

h × C + (1 − h) × M

• h: hit rate
• M: miss penalty
• C: cache access time

• High cache hit rates (> 90%) are essential

• Miss penalty must also be reduced

23/28
Question: Memory Access Time Example
• Assume 8 cycles to read a single memory word;
• 15 cycles to load a 8-word block from main memory (previous example);
• cache access time = 1 cycle
• For every 100 instructions, statistically 30 instructions are data read/ write
• Instruction fetch: 100 memory access: assume hit rate = 0.95
• Data read/ write: 30 memory access: assume hit rate = 0.90

Calculate: (1) Execution cycles without cache; (2) Execution cycles with cache.

24/28
Caches on Processor Chips

• In high-performance processors, two levels of caches are normally used, L1 and L2.
• L1 must be very fast as they determine the memory access time seen by the processor.
• L2 cache can be slower, but it should be much larger than the L1 cache to ensure a
high hit rate. Its speed is less critical because it only affects the miss penalty of the L1
cache.
• Average access time on such a system:

h1 · C1 + (1 − h1 ) · [h2 · C2 + (1 − h2 ) · M]

• h1 (h2 ): the L1 (L2) hit rate

• C1 the access time of L1 cache,
• C2 the miss penalty to transfer data from L2 cache to L1
• M: the miss penalty to transfer data from MM to L2 and then to L1.

25/28
Larger Block Size

• Take advantage of spatial locality.

• , If all items in a larger block are needed in a computation, it is better to load these
items into the cache in a single miss.

• / Larger blocks are effective only up to a certain size, beyond which too many items
will remain unused before the block is replaced.

• / Larger blocks take longer time to transfer and thus increase the miss penalty.

• Block sizes of 16 to 128 bytes are most popular.

26/28
 Miss Rate v.s. Block Size v.s. Cache Size

Miss rate goes up if the block size becomes a significant fraction of the cache size
because the number of blocks that can be held in the same size cache is smaller
(increasing capacity misses)

27/28
 Enhancement

Write buffer:
• Read request is served first.
• Write request stored in write buffer first and sent to memory whenever there is no
read request.
• The addresses of a read request should be compared with the addresses of the write
buffer.
Prefetch:
• Prefetch data into the cache before they are needed, while the processor is busy
executing instructions that do not result in a read miss.
• Prefetch instructions can be inserted by the programmer or the compiler.
Load-through Approach
• Instead of waiting the whole block to be transferred, the processor resumes execution
as soon as the required word is loaded in the cache.

28/28