0 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

KARLSRUHE INSTITUTE OF TECHNOLOGY (KIT)

Energy Efficient Task Partitioning based on the SingleFrequency Approximation Scheme

Santiago Pagani and Jian-Jia Chen

2013 IEEE 34th Real-Time Systems Symposium (RTSS)December 3-6 2013, Vancouver, Canada

KIT – University of the State of Baden-Wuerttemberg andNational Research Center of the Helmholtz Association www.kit.edu

Outline

1 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Outline

2 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Introduction

3 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Importance of Energy Efficiency:

Slow increases of battery capacity.Less Energy Consumption⇒ Prolong Battery Lifetime of EmbeddedSystems.

Increasing costs of energy.Less Energy Consumption⇒ Lower Power Bills for Servers.

Outcome for Computing Systems:

Motivated to move from single-core to multi-core.

Techniques for power management.

Introduction

3 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Importance of Energy Efficiency:

Slow increases of battery capacity.Less Energy Consumption⇒ Prolong Battery Lifetime of EmbeddedSystems.

Increasing costs of energy.Less Energy Consumption⇒ Lower Power Bills for Servers.

Outcome for Computing Systems:

Motivated to move from single-core to multi-core.

Techniques for power management.

Introduction

4 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Dynamic Power Management (DPM):

Technique for putting cores in a low-power mode: idle, sleep, off, etc.

Dynamic Voltage and Frequency Scaling (DVFS):

Technique for scaling the voltage and frequency of cores.

Per-core DVFS:Individual voltage and frequency for cores.

Optimal, but too expensive to manufacture.

Global DVFS:All cores share the same voltage.

Energy inefficient.

Introduction

4 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Dynamic Power Management (DPM):

Technique for putting cores in a low-power mode: idle, sleep, off, etc.

Dynamic Voltage and Frequency Scaling (DVFS):

Technique for scaling the voltage and frequency of cores.

Per-core DVFS:Individual voltage and frequency for cores.

Optimal, but too expensive to manufacture.

Global DVFS:All cores share the same voltage.

Energy inefficient.

Introduction

4 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Dynamic Power Management (DPM):

Technique for putting cores in a low-power mode: idle, sleep, off, etc.

Dynamic Voltage and Frequency Scaling (DVFS):

Technique for scaling the voltage and frequency of cores.

Per-core DVFS:Individual voltage and frequency for cores.

Optimal, but too expensive to manufacture.

Global DVFS:All cores share the same voltage.

Energy inefficient.

Introduction

4 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Dynamic Power Management (DPM):

Technique for putting cores in a low-power mode: idle, sleep, off, etc.

Dynamic Voltage and Frequency Scaling (DVFS):

Technique for scaling the voltage and frequency of cores.

Per-core DVFS:Individual voltage and frequency for cores.

Optimal, but too expensive to manufacture.

Global DVFS:All cores share the same voltage.

Energy inefficient.

Introduction

5 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Dynamic Voltage and Frequency Scaling (DVFS):

Multiple Voltage Islands:Compromise between Per-core DVFS and Global DVFS.

Cores are grouped into Voltage Islands.

Islands can have different voltages.

Figure: Intel’s SCC snapshot

Intel Corporation. Single-chip Cloud Computer (SCC). URL:http://www.intel.com/content/www/us/en/research/intel-labs-single-chip-cloud-computer.html

Introduction

6 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

CMOS-core Power Model

P (s) = Pdynamic (s) + Pstatic

Considering that:

Pdynamic (s) = CeffV2dds

s ∝(Vdd − Vt )

2

Vdd

We can approximate to:

P (s) = αsγ + β

Introduction

6 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

CMOS-core Power Model

P (s) = Pdynamic (s) + Pstatic

Considering that:

Pdynamic (s) = CeffV2dds

s ∝(Vdd − Vt )

2

Vdd

We can approximate to:

P (s) = αsγ + β

Introduction

6 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

CMOS-core Power Model

P (s) = Pdynamic (s) + Pstatic

Considering that:

Pdynamic (s) = CeffV2dds

s ∝(Vdd − Vt )

2

Vdd

We can approximate to:

P (s) = αsγ + β

Introduction

6 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

CMOS-core Power Model

P (s) = Pdynamic (s) + Pstatic

Considering that:

Pdynamic (s) = CeffV2dds

s ∝(Vdd − Vt )

2

Vdd

We can approximate to:

P (s) = αsγ + β

0 0.5 1 1.5 20

5

10

15

1.40

7.71

Frequency [GHz]

Pow

er [W

atts

]

Figure: α = 1.76 WattsGHz3 , γ = 3 and β = 0.5 Watts

Introduction

7 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Energy Consumption

E (s) = (αsγ + β)∆cs

Critical Frequency:

scrit =γ

√β

(γ− 1) α

0 0.5 1 1.5 20

2

4

6

8

10

Frequency [GHz]

Ene

rgy

[Jou

le]

Figure: α = 1.76 WattsGHz3 , γ = 3, β = 0.5 Watts and

∆c = 109 cycles

Outline

8 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Motivation

9 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

In each voltage island (or Global DVFS), for energy minimization:

Task partitioning and DVFS schedule:Play a major role in energy minimization.

Task Partitioning:

Convexity of E (s): For the same workload⇒ 2 · E (s) ≤ E (2 · s).

In general, load balancing reduces the dynamic energy consumption.

Balanced solution: with high complexity and not always feasible.

Good option: polynomial time algorithm based on load balancing, e.g.,Largest-Task-First (LTF)1 strategy.

1Chuan-Yue Yang, Jian-Jia Chen, and Tei-Wei Kuo. “An Approximation Algorithm for Energy-Efficient Scheduling on A Chip

Multiprocessor”. In: Conference on Design, Automation, and Test in Europe (DATE). 2005, pp. 468–473

Motivation

9 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

In each voltage island (or Global DVFS), for energy minimization:

Task partitioning and DVFS schedule:Play a major role in energy minimization.

Task Partitioning:

Convexity of E (s): For the same workload⇒ 2 · E (s) ≤ E (2 · s).

In general, load balancing reduces the dynamic energy consumption.

Balanced solution: with high complexity and not always feasible.

Good option: polynomial time algorithm based on load balancing, e.g.,Largest-Task-First (LTF)1 strategy.

1Chuan-Yue Yang, Jian-Jia Chen, and Tei-Wei Kuo. “An Approximation Algorithm for Energy-Efficient Scheduling on A Chip

Multiprocessor”. In: Conference on Design, Automation, and Test in Europe (DATE). 2005, pp. 468–473

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4τ5

τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1

τ2 τ3

τ4τ5

τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2

τ3

τ4τ5

τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4τ5

τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4

τ5τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4τ5

τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4τ5

τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4τ5

τ6

τ7

τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4τ5

τ6

τ7τ8

Motivation

10 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Largest-Task-First (LTF) Strategy Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 3 task sets: {T1,T2,T3}

τ1 τ2 τ3 τ4 τ5 τ6 τ7 τ8

T1 T2 T3

τ1 τ2 τ3

τ4τ5

τ6

τ7τ8

Motivation

11 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

DVFS Schedule→ Single Frequency Approximation (SFA)2 Scheme:

Use the lowest voltage/frequency, satisfying the timing constraints.

Linear time complexity. Is the simplest and most intuitive strategy.

Not optimal, but significantly reduces the management overhead.

No frequency alignment between cores =⇒ Any uni-core DPMtechnique can be adopted individually in each core.

s

tCore 1:

su

s

tCore 2:

su

s

tCore 3:

su

s

tCore 4:

su = max {scrit,w4}

2Santiago Pagani and Jian-Jia Chen. “Energy Efficiency Analysis for the Single Frequency Approximation (SFA) Scheme”. In: Proceedings

of the 19th IEEE International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA). 2013

Motivation

12 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Combining LTF and SFA for energy minimization:

Is a practical solution.

Very easy to implement.

What is the worst-case performancein terms of energy efficiency?

Problem Definition

13 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

For periodic real-time tasks, assigned on a voltage island.

Using Earliest-Deadline-First (EDF) on individual cores.

Contributions

Present a simple and practical solution for energy minimization.Task Partitioning: Double-Largest-Task-First (DLTF).

DVFS schedule: SFA.

Theoretically analyse such a solution:

AFDLTFSFA = max

EDLTFSFA

E∗OPT≤ max

EDLTFSFA

E∗↓

Problem Definition

13 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

For periodic real-time tasks, assigned on a voltage island.

Using Earliest-Deadline-First (EDF) on individual cores.

Contributions

Present a simple and practical solution for energy minimization.Task Partitioning: Double-Largest-Task-First (DLTF).

DVFS schedule: SFA.

Theoretically analyse such a solution:

AFDLTFSFA = max

EDLTFSFA

E∗OPT≤ max

EDLTFSFA

E∗↓

Problem Definition

13 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

For periodic real-time tasks, assigned on a voltage island.

Using Earliest-Deadline-First (EDF) on individual cores.

Contributions

Present a simple and practical solution for energy minimization.Task Partitioning: Double-Largest-Task-First (DLTF).

DVFS schedule: SFA.

Theoretically analyse such a solution:

AFDLTFSFA = max

EDLTFSFA

E∗OPT≤ max

EDLTFSFA

E∗↓

Outline

14 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Double-Largest-Task-First (DLTF)

15 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Initial solution: Task partitioning with LTF.

Tasks are regrouped, shutting down all possible cores.

This reduces the energy consumption for idling under SFA:

Example: LTF and SFAs

ts

ts

ts

ts

tbreak-even time

Core 1: τ1

Core 2: τ2

Core 3: τ3

Core 4: τ4

Core 5: τ5

Example: DLTF and SFAs

ts

ts

ts

ts

tbreak-even time

Core 1: τ1

Core 2: τ2

Core 3: τ3

Core 4:

τ4

Core 5:

τ5

Double-Largest-Task-First (DLTF)

15 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Initial solution: Task partitioning with LTF.

Tasks are regrouped, shutting down all possible cores.

This reduces the energy consumption for idling under SFA:

Example: LTF and SFAs

ts

ts

ts

ts

tbreak-even time

Core 1: τ1

Core 2: τ2

Core 3: τ3

Core 4: τ4

Core 5: τ5

Example: DLTF and SFAs

ts

ts

ts

ts

tbreak-even time

Core 1: τ1

Core 2: τ2

Core 3: τ3

Core 4:

τ4

Core 5:

τ5

Double-Largest-Task-First (DLTF)

15 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Initial solution: Task partitioning with LTF.

Tasks are regrouped, shutting down all possible cores.

This reduces the energy consumption for idling under SFA:

Example: LTF and SFAs

ts

ts

ts

ts

tbreak-even time

Core 1: τ1

Core 2: τ2

Core 3: τ3

Core 4: τ4

Core 5: τ5

Example: DLTF and SFAs

ts

ts

ts

ts

tbreak-even time

Core 1: τ1

Core 2: τ2

Core 3: τ3

Core 4:

τ4

Core 5:

τ5

Double-Largest-Task-First (DLTF)

16 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 4 task sets:{

TDLTF1 ,TDLTF

2 , . . . ,TDLTF4

}

TDLTF1 TDLTF

2 TDLTF3 TDLTF

4

τ1

τ2 τ3 τ4

τ5τ6τ7

τ8

τ4

τ5

τ5

τ6

τ5

Double-Largest-Task-First (DLTF)

16 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 4 task sets:{

TDLTF1 ,TDLTF

2 , . . . ,TDLTF4

}

TDLTF1 TDLTF

2 TDLTF3 TDLTF

4

τ1

τ2 τ3 τ4

τ5τ6τ7

τ8

τ4

τ5

τ5

τ6

τ5

Double-Largest-Task-First (DLTF)

16 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 4 task sets:{

TDLTF1 ,TDLTF

2 , . . . ,TDLTF4

}

TDLTF1 TDLTF

2 TDLTF3 TDLTF

4

τ1

τ2 τ3 τ4

τ5τ6τ7

τ8

τ4

τ5

τ5

τ6

τ5

Double-Largest-Task-First (DLTF)

16 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 4 task sets:{

TDLTF1 ,TDLTF

2 , . . . ,TDLTF4

}

TDLTF1 TDLTF

2 TDLTF3 TDLTF

4

τ1

τ2 τ3 τ4

τ5τ6τ7

τ8

τ4

τ5

τ5

τ6

τ5

Double-Largest-Task-First (DLTF)

16 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 4 task sets:{

TDLTF1 ,TDLTF

2 , . . . ,TDLTF4

}

TDLTF1 TDLTF

2 TDLTF3 TDLTF

4

τ1

τ2 τ3 τ4

τ5τ6τ7

τ8

τ4

τ5

τ5

τ6

τ5

Double-Largest-Task-First (DLTF)

16 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 4 task sets:{

TDLTF1 ,TDLTF

2 , . . . ,TDLTF4

}

TDLTF1 TDLTF

2 TDLTF3 TDLTF

4

τ1

τ2 τ3 τ4

τ5τ6τ7

τ8

τ4

τ5

τ5

τ6

τ5

Double-Largest-Task-First (DLTF)

16 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Example:

8 tasks: {τ1, τ2, . . . , τ8}

Partitioned into 4 task sets:{

TDLTF1 ,TDLTF

2 , . . . ,TDLTF4

}

TDLTF1 TDLTF

2 TDLTF3 TDLTF

4

τ1

τ2 τ3 τ4

τ5τ6τ7

τ8

τ4

τ5

τ5

τ6

τ5

Properties of Double-Largest-Task-First (DLTF)

17 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

If wLTFM < scrit, then:

wDLTF1 ≤ wDLTF

1 ≤ · · · ≤ wDLTFM < scrit

All cores run at the critical frequency: su = scrit

The energy consumption for execution is minimized.

EDLTFSFA = E∗OPT ⇒ AFDLTF

SFA = 1

If wLTFM ≥ scrit, then:

wDLTFM = wLTF

M

If there is only one task in TDLTFM ⇒ wDLTF

M ≤ w∗M

If there are at least two tasks in TDLTFM ⇒ wDLTF

Mw∗M≤ θLTF = 4

3 −1

3M

Properties of Double-Largest-Task-First (DLTF)

17 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

If wLTFM < scrit, then:

wDLTF1 ≤ wDLTF

1 ≤ · · · ≤ wDLTFM < scrit

All cores run at the critical frequency: su = scrit

The energy consumption for execution is minimized.

EDLTFSFA = E∗OPT ⇒ AFDLTF

SFA = 1

If wLTFM ≥ scrit, then:

wDLTFM = wLTF

M

If there is only one task in TDLTFM ⇒ wDLTF

M ≤ w∗M

If there are at least two tasks in TDLTFM ⇒ wDLTF

Mw∗M≤ θLTF = 4

3 −1

3M

Outline

18 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Negligible Leakage Power Consumption

19 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Energy Consumption for DLTF and SFA (when β = 0):

EDLTFSFA = P (su)

Lsu

M

∑i=1

wDLTFi ⇒ EDLTF

SFA = αL(

wDLTFM

γ−1) M

∑i=1

w∗i

Lower Bound Energy Consumption (when β = 0):

Unroll periodic tasks in a hyper-period⇒ frame-based tasks.

Use the results from the SFA analysis paper 3:

E∗↓(β=0) = αL

[M∑

i=1

(w∗i −w∗i−1

) γ√

M − i + 1]γ

3Santiago Pagani and Jian-Jia Chen. “Energy Efficiency Analysis for the Single Frequency Approximation (SFA) Scheme”. In: Proceedings

of the 19th IEEE International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA). 2013

Negligible Leakage Power Consumption

19 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Energy Consumption for DLTF and SFA (when β = 0):

EDLTFSFA = P (su)

Lsu

M

∑i=1

wDLTFi ⇒ EDLTF

SFA = αL(

wDLTFM

γ−1) M

∑i=1

w∗i

Lower Bound Energy Consumption (when β = 0):

Unroll periodic tasks in a hyper-period⇒ frame-based tasks.

Use the results from the SFA analysis paper 3:

E∗↓(β=0) = αL

[M∑

i=1

(w∗i −w∗i−1

) γ√

M − i + 1]γ

3Santiago Pagani and Jian-Jia Chen. “Energy Efficiency Analysis for the Single Frequency Approximation (SFA) Scheme”. In: Proceedings

of the 19th IEEE International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA). 2013

Negligible Leakage Power Consumption

20 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max

(wDLTF

M

)γ−1 M∑

i=1w∗i[

M∑

i=1

(w∗i −w∗i−1

)γ√

M − i + 1]γ

Critical Cycle Utilization Distribution: Minimizes E∗↓ , for a fixed

M∑

i=1w∗i

· · ·w∗1

w∗2w∗M-1

w∗M

· · ·w ′1 w ′2 w ′M-1

w ′M

w ′1 = w ′2 = · · · = w ′M−1 = Average(w∗1 ,w

∗2 , . . . ,w∗M−1

)Utilization Ratio: 0 ≤ δ =

Average(w∗1 ,w∗2 ,...,w

∗M−1)

w∗M≤ 1

Negligible Leakage Power Consumption

20 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max

(wDLTF

M

)γ−1 M∑

i=1w∗i[

M∑

i=1

(w∗i −w∗i−1

)γ√

M − i + 1]γ

Critical Cycle Utilization Distribution: Minimizes E∗↓ , for a fixed

M∑

i=1w∗i

· · ·w∗1

w∗2w∗M-1

w∗M

· · ·w ′1 w ′2 w ′M-1

w ′M

w ′1 = w ′2 = · · · = w ′M−1 = Average(w∗1 ,w

∗2 , . . . ,w∗M−1

)Utilization Ratio: 0 ≤ δ =

Average(w∗1 ,w∗2 ,...,w

∗M−1)

w∗M≤ 1

Negligible Leakage Power Consumption

21 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max

(

wDLTFMw∗M

)γ−1

· h (δ)

where

h (δ) =1− δ + δM(

1− δ + δγ√

M)γ ≤ h (δmax)

and

δmax =

1γ−1

(γ− 1 + M − γ

γ√

M)

(M γ√

M −M − γ√

M + 1)

Negligible Leakage Power Consumption

21 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max

(

wDLTFMw∗M

)γ−1

· h (δ)

where

h (δ) =1− δ + δM(

1− δ + δγ√

M)γ ≤ h (δmax)

and

δmax =

1γ−1

(γ− 1 + M − γ

γ√

M)

(M γ√

M −M − γ√

M + 1)

Negligible Leakage Power Consumption

21 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max

(

wDLTFMw∗M

)γ−1

· h (δ)

where

h (δ) =1− δ + δM(

1− δ + δγ√

M)γ ≤ h (δmax)

and

δmax =

1γ−1

(γ− 1 + M − γ

γ√

M)

(M γ√

M −M − γ√

M + 1)

h (δ) for γ = 3:

0 0.5 11

1.5

2

2.5 M=32

δ

h(δ)

Negligible Leakage Power Consumption

21 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max

(

wDLTFMw∗M

)γ−1

· h (δ)

where

h (δ) =1− δ + δM(

1− δ + δγ√

M)γ ≤ h (δmax)

and

δmax =

1γ−1

(γ− 1 + M − γ

γ√

M)

(M γ√

M −M − γ√

M + 1)

h (δ) for γ = 3:

0 0.5 11

1.5

2

2.5 M=32

δ

h(δ)

δmax

Negligible Leakage Power Consumption

21 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max

(

wDLTFMw∗M

)γ−1

· h (δ)

where

h (δ) =1− δ + δM(

1− δ + δγ√

M)γ ≤ h (δmax)

and

δmax =

1γ−1

(γ− 1 + M − γ

γ√

M)

(M γ√

M −M − γ√

M + 1)

h (δ) for γ = 3:

0 0.5 11

1.5

2

2.5

δmax

M=2M=4

M=8

M=16

M=32

δ

h(δ)

Negligible Leakage Power Consumption

22 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

If w∗M ≥ wDLTFM ⇒ w ′M = wDLTF

M → Same AF as for fixed task sets

0 8 16 24 321

2

3

M

AF

n.p.

SFA

(β=0)

1.81

1.41

1.17

AFSFAn.p. (β = 0) (γ=3)

AFSFAn.p. (β = 0) (γ=2)

Note: AFn.p.SFA

(β=0)only depends on the values of γ and M.

Negligible Leakage Power Consumption

23 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

If w∗M < wDLTFM ⇒

If δ > δmax ⇒ h (δ) < h (δmax):

0 0.5 11

1.5

2

2.5

M=2M=4

M=8

M=16

M=32

δ

h(δ)

We prove that:

δ ≥w ′1w ′M≥ 4M + 1

6M≥ 0.66

The AF for this case is ≤ θγ−1LTF · h

(4M+1

6M

)

Negligible Leakage Power Consumption

23 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

If w∗M < wDLTFM ⇒

If δ > δmax ⇒ h (δ) < h (δmax):

0 0.5 11

1.5

2

2.5

M=2M=4

M=8

M=16

M=32

δ

h(δ)

We prove that:

δ ≥w ′1w ′M≥ 4M + 1

6M≥ 0.66

The AF for this case is ≤ θγ−1LTF · h

(4M+1

6M

)

Negligible Leakage Power Consumption

23 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Approximation Factor for DLTF and SFA (when β = 0):

If w∗M < wDLTFM ⇒

If δ > δmax ⇒ h (δ) < h (δmax):

0 0.5 11

1.5

2

2.5

M=2M=4

M=8

M=16

M=32

δ

h(δ)

We prove that:

δ ≥w ′1w ′M≥ 4M + 1

6M≥ 0.66

The AF for this case is ≤ θγ−1LTF · h

(4M+1

6M

)

Negligible Leakage Power Consumption

24 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Worst-case Approximation Factor for DLTF and SFA (when β = 0):

AFDLTFSFA

(β=0) ≤ max{

h (δmax) , θγ−1LTF · h

(4M + 1

6M

)}

0 8 16 24 32 40 481

2

3

4

M

AFDLTF

SFA

(β=0)

AFDLTFSFA

(β=0 )whenγ = 3

θγ−1LTF · h(

4M +16M ) when γ = 3

h(δmax ) when γ = 3

AFDLTFSFA

(β=0 )whenγ = 2

θγ−1LTF · h(

4M +16M ) when γ = 2

h(δmax ) when γ = 2

Note: AFDLTFSFA

(β=0)only depends on the values of γ and M.

Outline

25 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Non-negligible Leakage Power Consumption

26 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

We approximate the Lower Bound Energy Consumption:

0 0.5 1 1.5 2 2.50

20

40

60

80 if wM

* > sdyn

:

P(s) = α sγ

if wM

* ≤ sdyn

:

P(s) = α scrit

γ + β

smin

scrit

sdyn

smax

w∗

M[GHz]

E∗ ↓(w

∗ M)[Joule]

Non-negligible Leakage Power Consumption

26 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

We approximate the Lower Bound Energy Consumption:

0 0.5 1 1.5 2 2.50

20

40

60

80 if wM

* > sdyn

:

P(s) = α sγ

if wM

* ≤ sdyn

:

P(s) = α scrit

γ + β

smin

scrit

sdyn

smax

w∗

M[GHz]

E∗ ↓(w

∗ M)[Joule]

Non-negligible Leakage Power Consumption

26 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

We approximate the Lower Bound Energy Consumption:

0 0.5 1 1.5 2 2.50

20

40

60

80 if wM

* > sdyn

:

P(s) = α sγ

if wM

* ≤ sdyn

:

P(s) = α scrit

γ + β

smin

scrit

sdyn

smax

w∗

M[GHz]

E∗ ↓(w

∗ M)[Joule]

Negligible Leakage Power Consumption

27 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Worst-case Approximation Factor for DLTF and SFA (when β 6= 0):

AFDLTFSFA ≤ max

γ− 1

[γγh (δmax)]1

γ−1

+ h (δmax) ,γ− 1

θLTF

[γγh

(4M+1

6M

)] 1γ−1

+ θγ−1LTF h

(4M + 1

6M

)

0 8 16 24 32 40 481

2

3

4

M

AFDLTF

SFA

AFDLTFSFA when γ = 3

γ−1

θL T F[γ γh( 4M +16M )]

1γ− 1

+ θγ−1LTFh(

4M +16M ) when γ = 3

γ−1

[γ γh(δmax)]1

γ− 1+ h(δmax) whenγ = 3

AFDLTFSFA when γ = 2

γ−1

θL T F[γ γh( 4M +16M )]

1γ− 1

+ θγ−1LTFh(

4M +16M ) when γ = 2

γ−1

[γ γh(δmax)]1

γ− 1+ h(δmax) whenγ = 2

Note: AFDLTFSFA only depends on the values of γ and M.

Outline

28 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Non-negligible Sleeping Overhead

29 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Our strategy can be combined with any DPM schemes.

Two cases:

∑Mi=1 w∗i < scrit:

DLTF assigns all tasks on one core and SFA executes at scrit.

Uniprocessor scheduling problem.

For example, using Left-To-Right (LTR) algorithm 4 → 2-approximation.

∑Mi=1 w∗i ≥ scrit:

Using any DPM scheme, against the optimal solution:

AFDLTFSFA-DPM ≤ AFDLTF

SFA +γ− 1

γ

4Sandy Irani, Sandeep Shukla, and Rajesh Gupta. “Algorithms for power savings”. In: Symposium on Discrete Algorithms (SODA).

Baltimore, Maryland, 2003, pp. 37–46

Non-negligible Sleeping Overhead

29 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Our strategy can be combined with any DPM schemes.

Two cases:

∑Mi=1 w∗i < scrit:

DLTF assigns all tasks on one core and SFA executes at scrit.

Uniprocessor scheduling problem.

For example, using Left-To-Right (LTR) algorithm 4 → 2-approximation.

∑Mi=1 w∗i ≥ scrit:

Using any DPM scheme, against the optimal solution:

AFDLTFSFA-DPM ≤ AFDLTF

SFA +γ− 1

γ

4Sandy Irani, Sandeep Shukla, and Rajesh Gupta. “Algorithms for power savings”. In: Symposium on Discrete Algorithms (SODA).

Baltimore, Maryland, 2003, pp. 37–46

Non-negligible Sleeping Overhead

29 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Our strategy can be combined with any DPM schemes.

Two cases:

∑Mi=1 w∗i < scrit:

DLTF assigns all tasks on one core and SFA executes at scrit.

Uniprocessor scheduling problem.

For example, using Left-To-Right (LTR) algorithm 4 → 2-approximation.

∑Mi=1 w∗i ≥ scrit:

Using any DPM scheme, against the optimal solution:

AFDLTFSFA-DPM ≤ AFDLTF

SFA +γ− 1

γ

4Sandy Irani, Sandeep Shukla, and Rajesh Gupta. “Algorithms for power savings”. In: Symposium on Discrete Algorithms (SODA).

Baltimore, Maryland, 2003, pp. 37–46

Outline

30 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Simulation Results

31 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

For negligible overhead for sleeping:

Power parametersmodelled from SCC.

150 cases of synthetictasks for every M, withdifferent:

Amount of tasks.

Cycle utilizations.

Resulting hyper-periods.

Task Partitioning:DLTF for SFA.

Lower bound for theoptimal solution.

Simulation Results

31 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

For negligible overhead for sleeping:

Power parametersmodelled from SCC.

150 cases of synthetictasks for every M, withdifferent:

Amount of tasks.

Cycle utilizations.

Resulting hyper-periods.

Task Partitioning:DLTF for SFA.

Lower bound for theoptimal solution.

0 8 16 24 32 40 481

2

3

4

M

EDLTF

SFA

(wDLTF

M)

E∗ ↓(w

∗ M)

pea

k

AFDLTFSFA

EDLTFSFA (wDLTF

M)

E∗

↓(w∗

M)

peak

Outline

32 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Introduction

Motivation and Problem Definition

Task Partitioning Scheme: Double-Largest-Task-First (DLTF)

Approximation Factor Analysis (energy consumption): DLTF and SFANegligible Leakage Power Consumption

Non-negligible Leakage Power Consumption

Non-negligible Sleeping Overhead

Simulations

Conclusions

Conclusions

33 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

For periodic tasks, combining DLTF and SFA is a practical solution forenergy minimization.

Approximation factor of DLTF and SFA for energy efficiency:Considered cases: negligible leakage, non-negligible leakage, andcombinations with DPM.

Bounded by γ and M (for all cases).

Simulations show a small gap compared with our analysis (for theworst-case).

Combining DLTF and SFA is an acceptable scheme based on theworst-case analysis.

34 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Thank you!

Questions?

34 Dec. 2013 Santiago Pagani and Jian-Jia ChenEnergy Efficient Task Partitioning based on the Single Frequency Approximation Scheme

KIT

Thank you!

Questions?