Jordan University
Computer Engineering Department
Computer Performance Evaluation Project
Supervisor: Dr. Ghieth Abandah
Student: Aieshah F. Almaslam
1
Outline
 Introduction
 Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
2
3
Why cache?
High cache performance leads to high total system performance
Cache performance parameters
 Cache size
 Cache block size
 Cache levels
 Cache mapping
 Replacement policy
 Unified cache or splitted
Very big range of combinations, so we need to
parameter performance evaluation.
4
Outline
 Introduction
 Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
5
6
SMPCache simulator
Trace driven simulator
Windows compatible
User friendly interface
Wide range of configuration
Uniprocessor build in memory traces
Multiprocessor downloaded memory traces
Creating your own memory traces
Text and graph results
Simulator
SMPCache User-friendly interface
7
8
Organization of memory SMP or DSM
Number of processors 1,2,4,8,16,32,64or 128
Snoopy Protocol MSI, MESI or DRAGON
Bus arbitration random, LFU or LRU
Directory protocol SGI or off
Word Width (bits( 8,16,32or 64
Words in a block 1,2,4,8,16,32,64,128,up to 1024
Memory Blocks 1,2,4,8,16,32,64,up to 4194304.
Cache Levels 1,2,3or 4
Unified or splitted unified or data and instructions
Cache Blocks 1,2,4,8,16,32,64,128,256,512,1024or 2048
Mapping Direct, set-associative or fully associative
Cache sets in case of set-associative mapping
Replacement policy Random, LRU, LFU or FIFO
Writing strategy Writeback
Architectural characteristics
supported by SMPCache
Outline
 Introduction
 SMPCache Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
9
10
Memory traces from SPEC’92 Benchmarks
Uni-processor traces
Real applications
Build in simulator software
Different types of applications
Integer and floating point
Examples: Hydro, Nasa7, Cexp, Mdljd, Ear,
Comp,Wave, Swm and UComp
Workload
Outline
 Introduction
 SMPCache Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
11
12
Miss rate
Less miss rate less main memory access
 indication on execution delay
 less ink in the final graph
Metric performance
Outline
 Introduction
 SMPCache Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
13
14
Design alternatives
for Cache size factor
Main memory size 64Gbytes
Block size 16Kbytes
Cache mapping fully associative
cache replacement Policy LRU
Cache levels 1
Cache levels size 16/32/64/128/256/1000Kbytes
Memory traces Comp/Nasa7/hydro
15
Design alternatives
for Cache multi-level factor
Main memory size 64Gbytes
Block size 16Kbytes
cache mapping fully associative
cache replacement Policy LRU
Cache levels 1/2/3/4
Cache levels size 16/32/64/128Kbytes in order
Memory traces Comp/Nasa7/hydro
16 / 32 / 64 / 128 Kbytes
16
Design alternatives
for Cache maping factor
Main memory size 64Gbytes
Block size 16Kbytes
cache mapping direct , 2,4,8,16,32 set associatiev
and fully associative
cache replacement Policy LRU
Cache levels 1
Cache levels size 16Kbytes
Memory traces Comp/Nasa7/hydro
Outline
 Introduction
 SMPCache Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
17
Implementation
In each simulation experiment
 Determine Main memory cache configuration
 Select the desired option from list menu
 Run the simulation
 Record the result
18
Outline
 Introduction
 SMPCache Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
19
An Example
of one experiment result
20
Cache Performance when
cache size is changed
21
Cache Performance when
cache levels is changed
22
Cache Performance when
cache mapping is changed
23
Outline
 Introduction
 SMPCache Simulator
 Workload
 Metrics
 Design alternatives
 Implementation
 Results Analysis
 Conclusion & Future work
24
25
Higher cache performance
 Higher cache size
 Higher cache levels
 Higher associativity
- But there is a ”Tradeoff”
Conclusion
26
Evaluate cache performance by studying
More factors
Factor interaction
Future Work
Any Question?
27
Eng. Aiesha F. Al-maslam

Cache Performance Evaluation under Multi-parameters Using SMPCache simulator

  • 1.
    Jordan University Computer EngineeringDepartment Computer Performance Evaluation Project Supervisor: Dr. Ghieth Abandah Student: Aieshah F. Almaslam 1
  • 2.
    Outline  Introduction  Simulator Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 2
  • 3.
    3 Why cache? High cacheperformance leads to high total system performance
  • 4.
    Cache performance parameters Cache size  Cache block size  Cache levels  Cache mapping  Replacement policy  Unified cache or splitted Very big range of combinations, so we need to parameter performance evaluation. 4
  • 5.
    Outline  Introduction  Simulator Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 5
  • 6.
    6 SMPCache simulator Trace drivensimulator Windows compatible User friendly interface Wide range of configuration Uniprocessor build in memory traces Multiprocessor downloaded memory traces Creating your own memory traces Text and graph results Simulator
  • 7.
  • 8.
    8 Organization of memorySMP or DSM Number of processors 1,2,4,8,16,32,64or 128 Snoopy Protocol MSI, MESI or DRAGON Bus arbitration random, LFU or LRU Directory protocol SGI or off Word Width (bits( 8,16,32or 64 Words in a block 1,2,4,8,16,32,64,128,up to 1024 Memory Blocks 1,2,4,8,16,32,64,up to 4194304. Cache Levels 1,2,3or 4 Unified or splitted unified or data and instructions Cache Blocks 1,2,4,8,16,32,64,128,256,512,1024or 2048 Mapping Direct, set-associative or fully associative Cache sets in case of set-associative mapping Replacement policy Random, LRU, LFU or FIFO Writing strategy Writeback Architectural characteristics supported by SMPCache
  • 9.
    Outline  Introduction  SMPCacheSimulator  Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 9
  • 10.
    10 Memory traces fromSPEC’92 Benchmarks Uni-processor traces Real applications Build in simulator software Different types of applications Integer and floating point Examples: Hydro, Nasa7, Cexp, Mdljd, Ear, Comp,Wave, Swm and UComp Workload
  • 11.
    Outline  Introduction  SMPCacheSimulator  Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 11
  • 12.
    12 Miss rate Less missrate less main memory access  indication on execution delay  less ink in the final graph Metric performance
  • 13.
    Outline  Introduction  SMPCacheSimulator  Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 13
  • 14.
    14 Design alternatives for Cachesize factor Main memory size 64Gbytes Block size 16Kbytes Cache mapping fully associative cache replacement Policy LRU Cache levels 1 Cache levels size 16/32/64/128/256/1000Kbytes Memory traces Comp/Nasa7/hydro
  • 15.
    15 Design alternatives for Cachemulti-level factor Main memory size 64Gbytes Block size 16Kbytes cache mapping fully associative cache replacement Policy LRU Cache levels 1/2/3/4 Cache levels size 16/32/64/128Kbytes in order Memory traces Comp/Nasa7/hydro 16 / 32 / 64 / 128 Kbytes
  • 16.
    16 Design alternatives for Cachemaping factor Main memory size 64Gbytes Block size 16Kbytes cache mapping direct , 2,4,8,16,32 set associatiev and fully associative cache replacement Policy LRU Cache levels 1 Cache levels size 16Kbytes Memory traces Comp/Nasa7/hydro
  • 17.
    Outline  Introduction  SMPCacheSimulator  Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 17
  • 18.
    Implementation In each simulationexperiment  Determine Main memory cache configuration  Select the desired option from list menu  Run the simulation  Record the result 18
  • 19.
    Outline  Introduction  SMPCacheSimulator  Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 19
  • 20.
    An Example of oneexperiment result 20
  • 21.
    Cache Performance when cachesize is changed 21
  • 22.
    Cache Performance when cachelevels is changed 22
  • 23.
    Cache Performance when cachemapping is changed 23
  • 24.
    Outline  Introduction  SMPCacheSimulator  Workload  Metrics  Design alternatives  Implementation  Results Analysis  Conclusion & Future work 24
  • 25.
    25 Higher cache performance Higher cache size  Higher cache levels  Higher associativity - But there is a ”Tradeoff” Conclusion
  • 26.
    26 Evaluate cache performanceby studying More factors Factor interaction Future Work
  • 27.