Presented by: Sameer Kulkarni

1
Improving Both the Performance Benefits and Speed
of Optimization Phase Sequence Searches-
Kulkarni, Jantz and Whalley

• Presented by Sameer Kulkarni
• Dept of Computer Information Sciences
• University of Delaware

2
Terms used

• Phase Ordering
• Genetic Algorithms
• Performance measurements
• Benchmarks
• Search granularity

3
Introduction

• Function vs. program level Granularity
• Embedded Systems
• Emulation
• Cost benefits
• Hybrid Search

4
Ideal Solution?

• Oracle ? Perfect sequence at the very start
• Wise Man Solution ? Given the present code
  predict the best optimization solution

5
Wise Man

• Understand Compilers
• Optimizations
• Source Code

?
6
Possible Solutions

• Pruning the search space
• Genetic Algorithms
• Estimating running times
• Precompiled choices

7
Genetic Algorithms
Fast Searches for Effective Optimization Phase
Sequences, Kulkarni et al. PLDI 04
8
Exhaustive vs Heuristic 2
9
Related Work

• Genetic Algorithms
• Other Evolutionary Techniques
• HMMs (CGO 06)
• Other Statistical methods
• Optimization Space Exploration

10
Present work

• Granularity
• Function Level
• File Level
• Program Level
• Hybrid

11
Experimental Setup

• VPO (Very Portable Optimizer)
• Base Genetic Algorithm
• Redundancy elimination

12
VPO

• Single IR
• Simplified phase ordering
• Configurable/modifiable

13
Redundancy Elimination

• Identical Sequence
• Identical Active Sequence
• Identical Function instance
• Equivalent Function Instance

14
Improvement?
120 days gt 12.5 days
15
Granularities Studied

• Function Level
• File Level
• Program Level
• Hybrid
• all compared to batch compilation

16

• Graphs, graphs and more graphs ?

17
Search Requirements
http//www.ittc.ku.edu/kulkarni/research/papers/l
ctes59f-preprint.pdf
18
Are they any good?
http//www.ittc.ku.edu/kulkarni/research/papers/l
ctes59f-preprint.pdf
19
Performance ??
http//www.ittc.ku.edu/kulkarni/research/papers/l
ctes59f-preprint.pdf
20
Future Work

• Other machine learning algorithms
• Reduce granularity
• Use a cluster to reduce search / learning time

21
Conclusion

• Reduced search overhead

22
Questions