An Offline Approach for Whole-Program Paths Analysis using Suffix Arrays

1
An Offline Approach for Whole-Program Paths
Analysis using Suffix Arrays

• G. Pokam, F. Bodin

2
Introduction

• Many optimizations/program analysis rely on
  control flow information
• The more accurate the better
• Trace analysis to identify hot sub-paths
• Similar problem as finding patterns in strings
• Or search for DNA sequences
• Adaptive cache reconfiguration
• Identification of the configuration change
• Reducing energy consumption

3
Whole Program Path Analysis

• Embedded systems
• No mechanism for run-time monitoring
• Offline analysis
• Interprocedural analysis
• Phase detection
• Consider a signature for each basic block
• Cache misses, ILP, static cycles (VLIW),
• What happens between hot sub-paths
• How are the hot sub-paths interleaved

4
Example
5
Main steps

• Reduce the potential size of the trace
• Keep only representative basic block
• Instrument the code to get the BB signature
• Run the program
• Compute the hot sub-paths
• Find the repeating patterns
• Exploit the sub-paths information
• Insertion of cache configuration instruction

6
Reducing the size of the trace

• Keep only a subset of the basic blocks
• Use strong regions Ball93
• Dont keep iterations of simple loops
• Keep only control condition basic blocks (lossy)

7
Suffix Arrays

• Karp Miller and Rosenberg algorithm
• Complexity is low Log(N) iteration, N the length
  of the trace
• Memory space used is linear to the size of the
  trace
• Can be used to
• Find the longest repeated sub-path
• Find the n-length repeated sub-path of BBWS
• Determine the frequency of each sub-path
• Identify the position of each instance of a
  sub-path

8
Suffix Arrays
9
KMR Algorithm
10
Characterizing the Hot Sub-Paths

• Three metrics
• Local Coverage how long does a sub-path last
• Global Coverage how representative a sub-path is
• Reuse Distance dispersion in the trace

11
Experiments

• Offline analysis ranges from a few minutes (40MB)
  to hours (GB trace)

12
Experiments (cont.)Trace Compression
13
Experiments (cont.)Coverage Adaptive Cache
Reconfiguration
Basic blocks signature is a set of data misses
14
Conclusion

• Suffix arrays are an efficient tool to deal with
  traces
• Accurate description of the sub-paths sequences
• But the CFG has to be simplified
• Has been used to dynamically adapt the cache
  configuration for reducing energy consumption

15
Future Works

• Convert hot sub path in speculative threads
• System on chip
• Identification of computation to migrate on
  co-processors
• More trace compression technique
• Abstraction of the control flow