ProfileDriven Code Execution for Low Power Dissipation

1
Profile-Driven Code Execution for Low Power
Dissipation

• Diana Marculescu
• ISLPED00

2
Introduction

• ? an optimal level of parallelism for energy
  consumption
• ? a compiler-assisted technique for code
  annotation that select run-time the optimal
  number of instructions to be fetched or executed
  in parallel as far as energy is concerned.
• Software-driven and a fine grain energy
  characterization on a per basic block, per
  process basis.

3
Introduction

• In superscalar processors, the hardware may
  execute from one to eight instruction per cycle.
• If a certain instruction is dependent or doesnt
  meet the constraint,
• only the instruction preceding it in the
  sequence are issues,
• hence the variability in issue rate.

4
Example

• Four 16-bit multipliers able to perform the
  multiplication

Total energy per operation
5
Profile-Driven Instruction Execution for energy
Optimization

• Analytic model
• The effect of speculative execution

6
Analytic model

• Power consumption is a function of the switched
  capacitance of a module.
• So, use the average Hamming distance.

7
Analytic model

• Proposition 1
• If an input sequence Xnn gt 0 is modeled by a
  lab-one Matkov chain, p(x) is the probability of
  occurrence of vector x and pk(xy) is the
  conditional probability of making a transition
  from y to x in exactly k steps, then the average
  Hamming k-distance denoted
• 
  converges when k -gt 8 and the limit is

8
Analytic model

• dk is the average Hamming distance between input
  vectors that are exactly k steps apart.
• Thus, dk is a measure of the average switching
  activity when exactly k units are used to process
  the input sequence I parallel.

9
Analytic model

• If a data trace Xnn gt 0 is modeled by a
  lab-one Matkov chain, and Ek Em(dk) is the
  energy consumed per operation when k identical
  functional units are available, then
• where Em is the energy macromodel and D is as in
  Proposition 1.

10
Analytic model

• The results presented above offers an answer to
  what is the optimal number of instructions to be
  executed in parallel?.
• The energy cost per operation for k instructions
  issued in parallel is proportional to dk..

11
The effect of speculative execution
If the branch Is mispredicted
12
Profile-Driven Code-Annotation for Low-Energy
Code Execution

• In practice, computing dk using proposition(1) is
  so expensive since it implies repeated matrix
  multiplications.
• The result do not capture the effect of
  speculative execution.
• Thus, propose profile-driven methodology to find
  the optimal number of instructions to be executed
  in parallel for each basic block.

13
Profile-Driven Code-Annotation for Low-Energy
Code Execution

• Assume that the input program is simulated for a
  typical input stream with an execution rate
  varying between 1 and K.

14
Profile-Driven Code-Annotation for Low-Energy
Code Execution

• After collecting the statistics for each basic
  block,
• the value of k which gives the minimum value of
  switched capacitance for the execution stage is
  used to annotate all instructions in the basic
  block under consideration.

15
Experimental Results
16
Conclusion

• Present a novel technique for code optimization
  for low power based on profile driven
  methodology.
• The technique can be used as means
• for prolonging the battery life by decreasing
  the energy consumption,
• for the thermal management by achieving
  significant average power reduction in the
  execution stage.