Discovering and Exploiting Program Phases

1
Discovering and Exploiting Program Phases

• Timothy Sherwood, Erez Perelman,
• Greg Hamerly, Suleyman Sair, Brad Calder
• CSE 231 Presentation by Justin Ma

2
400 Million Instructions
Non-Existent Processor
New Processor
New Compiler
Spec2000
Benchmark
Simulator
3
400 Million Instructions

• Suppose you have a time budget
• Less than half second of execution time
• What would you simulate?
• Beginning?
• Middle?
• End?

4
400 Million Instructions
Programs exhibit diverse modes of behavior
gzip
gcc
5
400 Million Instructions

• Suppose you have a time budget
• Less than half second of execution time
• What would you simulate?
• Beginning?
• Middle?
• End?
• Samples of different modes of behavior

6
Program Phases

• Observation programs exhibit various modes of
  periodic behavior
• These modes are program phases
• Challenge Extract these automatically

7
Phase Basics

• Intervals slices in times
• Phases intervals with similar behavior

IPC
Time (Instruction Count)
8
Phase Basics

• Intervals slices in times
• Phases intervals with similar behavior

IPC
Time (Instruction Count)
9
Defining Similar Behavior

• Metric for comparing intervals?
• Cache misses?
• IPC?
• Branch misprediction rates?
• Problem Performance alone is too architecture
  dependent

10
Defining Similar Behavior

• Code path traversal
• Directly affects time-varying behavior
• Execute same code, same performance
• Architecture independent
• Metrics for code path traversal
• Frequency of branches
• Frequency of function calls
• Frequency of basic block calls

11
Basic Block Vector
B1
0
0
0
0
Time t
B1
B2
B3
B4
B2
B3
B4
12
Basic Block Vector
B1
1
1
0
1
Time t
B1
B2
B3
B4
B2
B3
B4
13
Basic Block Vector
B1
2
1
1
2
Time t
B1
B2
B3
B4
B2
B3
B4
14
Basic Block Vector
B1
2
1
1
2
Time t
B1
B2
B3
B4
B2
B3
0
0
0
0
Time t 1
B1
B2
B3
B4
B4
15
Basic Block Vector
B1
2
1
1
2
Time t
B1
B2
B3
B4
B2
B3
1
1
0
1
Time t 1
B1
B2
B3
B4
B4
16
Basic Block Vector
B1
2
1
1
2
Time t
B1
B2
B3
B4
B2
B3
2
2
0
2
Time t 1
B1
B2
B3
B4
B4
Manhattan Distance 1 2 1 0
2 Euclidian Distance sqrt((1 2)2 (1 0)2)
sqrt(2)
17
Basic Block Similarity Matrix

• gzip

18
Basic Block Similarity Matrix

• gcc

BBV similarity between intervals reflects
performance similarity
19
Automatic Phase Classification

• Classify intervals into phases
• We do not know which BBVs correspond to
  particular phases a priori
• k-means clustering
• Iterative clustering algorithm
• Dimension Reduction
• Random Linear Projection
• Try different k values
• Use BIC to choose best

20
Automatic Phase Classification
21
Automatic Phase Classification
Clustering accurately distinguishes phases
automatically
22
SimPoint

• Simulate large programs on a budget
• Perform detailed simulation on representative
  code snippets
• Choose centroid interval from each phase (10
  million instructions)
• Extrapolate large program performance
• Weighted by frequency of phase

23
SimPoint

• Simulate 400 million instructions total

Accurate estimate despite instruction budget
24
Why SimPoint Succeeds

• Program behavior varies over time
• SimPoint intelligently chooses which intervals to
  simulate
• Regularity within program phases allows accurate
  extrapolation

25
Online Classification

• Detect phases as program is running
• Applications
• Thread scheduling
• Power management
• Predicting future phases
• Challenges
• One pass of input
• Limited storage

26
Online Classification
27
Online Classification
High variance in metrics across full trace
Low variance shows online classification succeeds
in finding phases
28
Conclusions

• Phases are a vital abstraction
• Performance varies greatly w/in program
• Attributable to different modes of behavior
• Can discover phases automatically
• Offline k-means clustering
• Online
• Code path characterization
• Strong correlation with actual performance
• SimPoint exploits this with great success

29
(No Transcript)
30
Outline

• Introduction (motivate)
• Basics (definitions, BBV, BBMatrix)
• Offline Phase Classification
• SimPoints
• Online Phase Classification
• Conclusions

31
Limitations of Clustering
32
Bayesian Information Criterion

• Fit to Gaussians

33
Self-Modifying Code
Self-modifying code
85o
Program Phases
34
Learning Phases
35
Learning Phases