1

A Run-Time Feedback Based Energy Estimation Model for

Embedded Systems

Selim GürünChandra Krintz

Department of Computer ScienceU.C. Santa Barbara

International Conference on Hardware/Software Codesign and System Synthesis (CODES-ISSS)

Seoul, KoreaOctober 22-25, 2006

CODES-ISSS’06 2

Power-Aware Execution: Big Picture

• Power-aware methods divide task execution into operations, and prepare an execution plan for each• Operation: smallest user-visible unit of

execution • Typical operation: Rendering a scene,

translating a sentence, calculating a shortest path in a map

• Need to know energy cost of each plan

Knowing future energy cost of operations requires profiling them at run-time

IdentifyOperations

Profile atRuntime

PredictFuture Costs

Develop Power-Aware Execution Strategy

CODES-ISSS’06 3

Outline

• Extant run-time power profiling techniques• Power profiling methodologies for

embedded computers

• Proposed model• Overview• Model construction• Capturing system dynamics

• Evaluation

• Summary and Conclusion

CODES-ISSS’06 4

Run-Time Energy Profiling: Overview

OS Interfaces like ACPI:

+ Provides simple API to battery voltage sensors+ Ok for different hw. power levels - Very coarse- Not precise

Execution Time:

+ Simple to measure

+ Fast and precise- Not correlated to power- Not suitable when hw.

power levels change: DVS, sleep

HPMs:

+ Fast access

+ Quite accurate- Architecture dependent- Not designed for power

estimation --many events missing

CODES-ISSS’06 5

Run Time Energy Profiling: HPMs

• CPU counters provide unparalleled insight into program behavior• Cache, TLB misses• Instructions executed per cycle (IPC)

• How can we accurately gather program energy consumption by monitoring key parameters?

Use HPMs as pseudo CPU component access counters:Energy Consumption = I Cache * a0 + D Cache * a1 +

ALU * a2 +…

CODES-ISSS’06 6

Run Time Energy Profiling: HPMs

• CPU counters provide unparalleled insight into program behavior• Cache, TLB misses• Instructions executed per cycle (IPC)

• How can we accurately gather program energy consumption by monitoring key parameters?

• Useful but not enough:• CPU consumes a portion of total energy; power-aware

strategies need to know full picture.• Fails when hardware changes its behavior: DVS, sleep states

• A different strategy needed!

Use HPMs as pseudo CPU component access counters:Energy Consumption = I Cache * a0 + D Cache * a1 +

ALU * a2 +…

CODES-ISSS’06 7

Proposed Energy Profiling Model

Construct Power Model

Update Model Coefficients

Measure Energy Consumption in Large Intervals and Compare to

HPM Model Estimates

Determine Model

Coefficients

Predict Energy Consumption using HPM Model

Power-Efficient Execution Plan

Offline Analysis

Continuous model improvement at run-time

Fine-Grain Energy Estimation

CODES-ISSS’06 8

Case Study: Intel XScale on Stargate

• 32 bit XScale – 400MHz• 64 MB RAM• Runs Familiar Linux• No Display• Wireless 802.11• Compact Flash

XScale Major HPM Events

Inst/Data cache misses

Data dependency stalls

Inst/Data TLB misses

Brach mispredicted

Instruction executedSCL

CODES-ISSS’06 9

Constructing Model

• Are there any correlations between HPM values and full system power consumption?• Absolutely! --but some challenges exist.

• Good correlation in memory/CPU subsystem• High IPC -> CPU intensive application • High cache misses/hits -> memory intensive application

• But I/O is the problem!• Some heuristics possible, e.g. Low memory activity and low IPC

-> possible I/O wait state • Better to use software counters embedded into drivers

CODES-ISSS’06 10

Model Coefficients

E = X1 a1 + X2 a2 + X3 a3 +…

XI : Independent VariablesaI : Coefficients

• Estimate coefficients using least squares linear regression (LSQ)• Stable and simple• Linearity assumption

Only Major All related

LSQ Model: Which variables?

Efficient, clear Easier to understand Less accurate

More accurate Run-time overhead Modeling difficulties

due to variable dependencies

CODES-ISSS’06 11

Parameter Selection & Dependencies

• Hard to include all variables: Too many parameters clutter model• Parameter dependencies unstable parameter estimations• E.g. Volume = a0 + a1 * pounds + a2 * grams

• Work-around is non-trivial; HPM characteristics e.g.:• TLB miss more CPU cycles & cache miss• Memory Stall Fewer instruction executed

Multicollinearity!

CODES-ISSS’06 12

Run-Time Energy Estimation

Computation Communication

SimpleCore Clock CyclesData Stalls

Core Clock CyclesBytes TransmittedBytes Received

Complex

Core Clock CyclesInstruction Cache MissesInstructions NDeliveredData StallsITLB MissesDTLB Misses

Core Clock CyclesBytes TransmittedBytes ReceivedPackets TransmittedPackets Received

CODES-ISSS’06 13

Run-Time Model Improvement

• Global coefficients• Compute using off-line model

• Continuously update coefficients• Improve using most recent

data• Gradually phase out previous

measurements

• Recursive least squares with exponential decay• Smaller decay factor-> more

agile

Global Coefficients

Measure Power

Update with RLS

Model Parameters: Decay factor Update period Measurement error

CODES-ISSS’06 14

Feedback Source: DS 2760

• Measures current flow in and out of battery

• Internally: A small A/D converter attached to a high precision internal resistor

Pros/Cons:+ Highly Available

e.g. iPAQs, sensor network gateways, cell phones

- Not precise enough for monitoring task energy consumption

0.25 mAh error in each reading

- Slow, one-wire serial interface

CODES-ISSS’06 15

Stargate and Our Evaluation Bench

PowerTool

VPerfmon

VPMonSCL High-precision

Data Acquisition Device

ProgrammablePower Supply

CODES-ISSS’06 16

Methodology

• Collect energy consumption every so often• Every 10 million instructions ( a so-called interval)

• Validate model accuracy on imprecise measurement data• Inject uniformly distributed random error • Evaluate various precision (error) levels: 1X – 8X• Predict energy consumption of each interval• Continuously improve model parameters every 10M * K

intervals

• Use a large group of workload• Computational benchmarks• Computational + communication oriented benchmarks

CODES-ISSS’06 17

Static vs. Adaptive Models

0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

bisor

t

em3d

gsm

deco

de

gsm

enco

de

jpegde

code

jpegen

code life

mpe

g2dec

ode

mpe

g2enc

ode

pvkx

pvkx

bpv

nx

treea

dd

AVERAGE

Err

or

%

0.9

Static

CODES-ISSS’06 18

Average Error Rates

Interval Size

1X 2X 4X 8X Best

100 53.3% 29.9% 14.5% 16.9% 3.8%

200 68.8% 24.1% 12.9% 7.3% 2.7%

400 33.0% 22.0% 9.1% 7.7% 2.8%

Error rates and Interval sizes –Simple Model

Measurement Precision

CODES-ISSS’06 19

Average Error Rates-Complex Model

Interval Size

8X 8X Best Best

100 16.9% 28.1% 3.8% 4.3%

200 7.3% 33.3% 2.7% 3.8%

400 7.7% 24.0% 2.8% 4.1%

Measurement imprecision reduce complex model quality more than the simple one!

Simple Model

CODES-ISSS’06 20

Related Work

• High-End CPU Power Models• Define CPU component access rate using HPM access heuristics• OS calls power consumption as a function of IPC

• Embedded CPU Power Models• Five HPM counters for XScale• Also evaluated memory model

• Memory models• UltraSparc memory subsystem

• All above are static models

• Power profiling setups • Powerscope

CODES-ISSS’06 21

Summary & Conclusions

• Our Goal: An accurate, efficient run-time power profiling system• Hardware counters are key

• Define software counters for I/O • Smart battery monitors expose dynamics in power behavior• We propose a hybrid system that combine both

• Lessons learned• Dynamic models are much better than static ones in power

modeling• Models should decay old measurements conservatively when

measurement errors are present• Measurement errors in the presence of multicollinearity can be

deadly

CODES-ISSS’06 22

Backup Slides

CODES-ISSS’06 23

Decay Factor vs. Accuracy

0%

20%

40%

60%

80%

100%

120%

bisor

t

em3d

gsm

deco

de

gsm

enco

de

jpegde

code

jpegen

code life

mpe

g2dec

ode

mpe

g2enc

ode

pvkx

pvkx

bpv

nx

treea

dd

AVERAGE

Err

or

%

0.5

0.7

0.9

1

CODES-ISSS’06 24

Execution Cost

0

5

10

15

20

25

30

35

40

compactcomputation

compact network complex network complexcomputation

Mil

liS

eco

nd

s

BMU access

Processing

CODES-ISSS’06 25

Benefit from an Offline Profiler

0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

bisor

t

em3d

gsm

deco

de

gsm

enco

de

jpegde

code

jpegen

code life

mpe

g2dec

ode

mpe

g2enc

ode

pvkx

pvkx

bpv

nx

treea

dd

AVERAGE

Err

or

%

with offline profiling

w/o offline profiling

CODES-ISSS’06 26

Power-Aware Execution: Case Study

0

20

40

60

80

100

120

140

160

En

erg

y (J

ou

les)

.

Speech Recognition Execution Plans

src: Flinn’01BaselineLocal ReducedRemote Reduced

Requires run-time power prediction of different

execution plans!