Exploring Processor Design Using Genetic Programming

1
Exploring Processor Design Using Genetic
Programming

• Borys Bradel
• Kirk Stewart

ECE1718 28 April 2004
2
Outline

• Introduction
• Genetic Programming
• Processor Parameters
• Methodology
• Results
• Conclusion

3
Introduction

• Processors have many parameters
• Good parameters are not obvious
• Good parameters may be application specific
• Simulation takes a long time
• Impractical to search entire space
• Efficient algorithms need to be used

4
Genetic Programming

• Mimics real life
• Entities are defined by chromosomes
• Entities evolve over many generations
• Survival of the fittest
• Propagation, crossover, mutation

5
Propagation

• Random selection
• Weighted by fitness function

Example 1
Example 2
0.0
0.0
0.0
2 4 2 1 5
3 2 1 5 5
1 f0.5 2 f0.7 3 f0.9 4 f1.1 5 f1.8
x
x
choose four times randomly
x
x
x
x
x
propagate best
x
5.0
5.0
5.0
6
Crossover
1 4 5 6 4 5 3 6 2 3
1 4 5 3 2 4 0 6 2 3
2 4 0 3 2 4 0 5 2 1
2 4 0 6 4 5 3 5 2 1
start
end
2 2 3 6 5 4 0 2 2 5
0 2 3 6 5 4 0 2 5 1
0 1 5 3 6 4 3 6 5 1
2 1 5 3 2 4 3 6 2 5
start
end
mutation
7
Transition Between Generations
5 4 3 3 4 3 6 0 3 4 1 4 5 0 0 2 0 0 2 2 6 3 2 2 5
0 1 5 3 4 5 6 3 5 6 5 3 4 4 3 6 2 5 4 1 4 4 1 4
2 6 0 4 1 1 3 5 5 0 6 3 4 0 4 5 0 2 2 1 4
5 6 3 5 6 5 3 4 4 3 6 3 2 2 5 0 1 5 3 4 1 4 5 0 0
2 0 0 2 2 6 2 5 4 1 4 4 1 4 2 3 4 0 4 5 0 2 2 1
4 6 0 4 1 1 3 5 5 0 6 5 4 3 3 4 3 6 0 3 4
5 6 3 5 6 5 3 1 4 3 4 3 1 2 5 0 1 4 3 4 6 2 5 4 1
2 0 0 4 2 1 4 0 0 0 4 4 1 2 2 6 4 0 4 3 0 2 5 0
6 3 0 4 1 1 3 5 5 1 4 5 4 3 3 4 3 6 0 3 4
crossover and mutate
shuffle
evaluate fitness and propagate based on it
8
Chromosomes

• Instruction fetch queue length
• Branch prediction algorithm
• Decode width
• Issue width
• Commit width
• Load/store queue length
• L1 D-cache size
• L1 I-cache size

• L2 size
• L2 associativity
• L1 replacement algorithm
• L2 replacement algorithm
• of integer ALUs
• of integer multiplier/dividers
• of floating point ALUs
• of floating point multiplier/dividers

47 bits ? 247 140 thousand billion
9
Processor Fitness

• Fast execution is good
• Large area is bad
• Large power consumption is bad
• Fitness IPC/power
• Hard upper limit on area
• Alternates IPC/power2 or IPC2/power

10
Methodology

• Simulations performed in Wattch
• An extension of SimpleScalar v3.0 sim-outorder
• Calculates static and dynamic power consumption
• Used area model from Exploring the Design Space
  of Future CMPs, Huh, Keckler and Burger
• Perl subroutine written by authors
• Very rough estimate

11
Methodology

• C program to calculate fitness, create new
  generations
• Perl scripts coordinate simulation, population
  generation
• Ran on cluster of 30 dual AMD Athlon MP 2600

genetic.cpp run_sim.pl
sim-outorder
sim-outorder
sim-outorder
12
Results Fitness Over Time
13
Results IPC Over Time
14
Results Power Over Time
15
Results Area Over Time
16
Results Effect of Mutation Rate
crafty
17
Results Number of New Members
0.5
2.5
18
Conclusion

• Created a fast algorithm to search a large design
  space
• Explored effects of varying algorithm parameters
• Evaluated algorithm performance through extensive
  simulation
• Algorithm is effective at finding good processor
  configurations

19
Future Work

• Improve area estimation
• Use area as part of fitness
• Evaluate best configurations performance
• Simulate benchmarks not used in genetic algorithm
• Use results from average runs on individual
  benchmarks
• Compare to existing architectures