Value Compression to Reduce Power in Data Caches

1
Value Compression to Reduce Power in Data Caches
Europar03, Klagenfurt - August 26-29, 2003

• Carles Aliagas, Carlos Molina and Montse García
• Universitat Rovira i Virgili Tarragona,
  Spaincaliagas, cmolina, mgarciaf_at_etse.urv.es
• Antonio González and Jordi Tubella
• Universitat Politècnica de Catalunya Barcelona,
  Spain antonio,jordit_at_ac.upc.es

2
Motivation

• Caches spend close to 50 of total die area

• Caches may be responsible for 10 to 20 of total
  power dissipated by a processor

3
Cache Storage
Cache access
TAG
DATA
address
Cache Line
TAG Comparison
HIT
Value
4
Structures Accessed
Cache access
TAG
DATA
address
Cache Line
TAG Comparison
HIT
Value
5
Useful Information
Cache access
TAG
DATA
address
Cache Line
TAG Comparison
HIT
Value
6
Compressed Data
30 of total capacity
TAG

Compressed Data
address
Cache Line
TAG Comparison
Value decompression
HIT
Value
7
Objective

• Reduction of
• Die area.
• Energy consumption.
• Maintaining
• Miss ratio.
• Latency.

8
Compression Types

• Integer compression
• The value is an integer.
• The compression applied is
• Sign extension.
• 0 bit elimination.
• Address compression
• The value is a pointer.
• From pointer to pointer only few bits are
  different.
• Address are
• Data pointers, code pointers and stack pointers
• Try to recognize a common pattern.
• The patterns are stored on a common structure.
• Only the different bits are stored in cache.

9
Integer Compression

• Six different compression forms, depending on
  significant bits position.
• Two 32-bit words or one 64-bit word

T1
0
0
T2
0 0
0 0
T3
0 0
T4
T5
0 0
0 0
T6
10
Address Compression

• One compression type
• Pointer pattern variable bits.
• Several patterns stored in a separate table
  Address Table (AT)
• Only 8 patterns represents 83.

0 0
TA
Pattern
0
Pattern-1
Pattern-2
Pattern-3
Pattern-k
11
Data Value Compression

• Close to 70 of all data stored in a 16KB data
  cache can be compressed from 64 to 22 bits.

Percentage of each type of compression applied
12
Simulation Enviroment

• Simulators
• Cacti tool version 3.0 (Static Analysis)
• Alpha version of SimpleScalar 3.0 (Dynamic
  Analysis)
• Benchmarks
• Spec2000
• Maximum Optimization Level
• DEC C F77 compilers with -non_shared -O5
• Statistics Collected for 1 billion instructions
• Skipping initializations

13
The Pattern Cache
Address

• - Integer type
• compression
• Address pattern
• index
• - Non compressed

- Compressed value or - Pointer to
uncompressed value
LT
4 bytes
Tag
Direct
VT
Pointer
access
Location
Address Pattern
AT
Word 8 bytes
13 bits
4x 22bits
AB
Free bit map
Tag
Line 32 bytes
14
Cache Behavior
Data access
93
7
HIT
MISS
Integer compression
Address compression
Non compressed
Integer comp.
Address comp.
Non comp.

75
25
Next Level cache
70
23
Static analysis
Dynamic analysis
15
Static analysis

• Base cache
• Direct mapped 16KB, 32Bytes/line.
• Pattern Cache
• Same associative and number of lines. LT
• Three configurations reducing the area of VT to
  40, 30 and 20 of the original data area.
• AT 8 entries.
• Assist Buffer
• Full associative cache of 16 entries
• Working as a Victim cache in the Base cache.

16
Die Area
17
Access Time
18
Energy Consumption
19
Dynamic Analysis

• Energy consumption of 1st level cache plus 2on
  Level cache
• EC (Hit DL1_EC) (Miss (2 DL1_EC
  DL2_EC))
• 1st level cache 64KB to 4KB
• 2on level cache 512KB

20
Energy Consumption (2)
21
Results

• Caches ranging from 4KB to 64KB

22
Conclusions

• High degree of value compression can be applied
  in data caches.
• The compression allow to store the same
  information with less die area.
• This involves an energy consumption reduction.
• While maintaining the same access time.
• With minor miss ratio increment.

23
Value Compression to Reduce Power in Data Caches
Europar03, Klagenfurt - August 26-29, 2003

• Carles Aliagas, Carlos Molina and Montse García
• Universitat Rovira i Virgili Tarragona,
  Spaincaliagas, cmolina, mgarciaf_at_etse.urv.es
• Antonio González and Jordi Tubella
• Universitat Politècnica de Catalunya Barcelona,
  Spain antonio,jordit_at_ac.upc.es