Highlevel Software Energy Macromodeling

1
High-level Software Energy Macro-modeling

• T. K. Tan and N. K. Jha
• Princeton University
• DAC01

2
Introduction

• Paradigm shift towards software power estimation
  technique
• Previous work instruction-level modeling or
  structural modeling of the underlying hardware
  architecture
• For large systems or design space exploration, we
  need higher efficiency maintaining high accuracy
  or fidelity
• Macro-Modeling

3
Related Work

• Instruction-level characterization
• Tiwari94
• Sciuto00, Sama00 perform measurements on a
  limited subset of instructions
• Chang00
• Simulation of the processor architecture
• Wattch, SimplePower

4
Overview

• Two approaches to energy macro-modeling
• Complexity-based macro-modeling
• For data-intensive functions
• Profiling-based macro-modeling
• For control-intensive functions
• 95 accuracy
• Speedup

5
General Approach

• Regression analysis
• Pj the parameters of the macro-model
• cj coefficients
• p the number of parameters
• Step 1 determine what parameters are needed Pj
• Step 2 find the corresponding coefficients cj

6
How to find cj

• Step 1 n typical input data SI1,I2, , In
• Pi,j is the j-th parameter value evaluated for
  input data Ii.
• Step 2 obtain the energy consumption for every
  Ii
• Energy vector E(E1 E2 En)T
• Step 3 E PC C(c1 c2 cp)T
• Step 4 C PTP1PTE
• Speedup Tsimulate / Tmodel

7
Complexity-based Energy Macro-modeling

• Many multimedia or data processing applications
  make use of algorithms with known average-case
  algorithmic complexities
• Insertion sort
• E c1 c2s c3s2 ( s is the size of the array
  )
• Ease of use
• Complexity of the function may not be known

8
Profiling-based Energy Macro-modeling

• Basic-block Profiling
• Tiwaris basic block profiling
• Regression-based macro-modeling
• bi is the execution counts
• Abstract away all the details
• Error can be large

9
Profiling-based Energy Macro-modeling

• Correlation Profiling
• Correlation a consecutive sequence of events
• Used branch prediction
• Rjs are the counters for the correlation events
• Basic-block correlation
• Ball-Larus Path correlation

R
c
R
c
E

2
2
1
1
10
Profiling-based Energy Macro-modeling

• Basic-block Correlations
• m-block correlation a sequence of execution of
  m basic block

3-block correlations ABC, ABD, AFG, AFH, BCE,
BDE, CEB, DEB, EFH, EFG,
Ri,j is the count for the j-th 3-block
correlation in trace i
Less effective when the paths through the
functions are generally long
11
Profiling-based Energy Macro-modeling

• Ball-Larus Path Correlations
• BL-path acyclic paths
• starting from either the ENTRY node or the nodes
  which are the targets of one or more back edges
• ending with either the EXIT or the nodes which
  are the sources of one or more back edges.

Trace 1 ABCEBCEBDEBCEFHI Trace 2
ABCEBCEBCEBDEFGI gtTrace 1 BL7 BL9 BL10 BL12
Trace 2 BL7 BL9 BL9 BL13 3-BL-path
BL7-BL9-BL10, BL9-BL10-BL12 BL7-BL9-BL9,
BL9-BL9-BL13
12
Profiling-based Energy Macro-modeling

• Selection of Profiling Methods
• Obtain q different energy macro-models using bb
  corr. and r different models using BL-path corr.
• Pareto-rank the qr different energy macro-models
  based on accuracy and speedup.
• Pareto-rank of other solutions which do not
  dominate it
• A solution dominates another solution if it is
  better than the second one in both accuracy and
  speedup.

13
Experimental Results

• Complexity-based Energy Macro-model

14
Experimental Results

• Profiling-based Energy Macro-model
• q5, r 5

15
Experimental Results

• Profiling-based Energy Macro-model

16
Conclusion

• Two kinds of macro-modeling techniques
• Complexity-based
• Use the algorithmic complexity
• Profiling-based
• Use bb corr. Or BL-path corr.
• Based on linear regression models
• Speedup over the low-level technique ranges from
  one to five orders of magnitude