Uncertainty Based Scheduling: Energy-Efficient Ordering for Tasks with Variable Execution Time

Uncertainty Based Scheduling:Uncertainty Based Scheduling:Energy-Efficient OrderingEnergy-Efficient Ordering

forforTasks with Variable Execution Tasks with Variable Execution

TimeTime

Flavius Gruian and Krzysztof Kuchcinski

Embedded Systems DesignLaboratory

Lund Institute of TechnologySweden

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

2(11)

Presentation OutlinePresentation Outline

• Problem Set-up• A Motivation• Uncertainty Based Scheduling• Experiments

– comparison to FullSearch– measurements on EVB80200 platform

• Summary & Conclusions

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

3(11)

Problem Set-up Problem Set-up

• tasks: period=deadline, variable execution

• off-line (static) ordering but• run-time speed selection

– speed for the kth task

– energy for a period(clock energy e(s)=Ks)

– average energy

s(1, ,k1) WCEii1

NAfref Xjj1

k1

E(X1, ,N ) X1K1s0 XiKii2

N s(1, , i1)

E(X ) E(x )xN

x1

(x1)(xN )dx1dxN

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

4(11)

A Motivational ExampleA Motivational Example

Task Set: 3 tasks, uniform distribution (BCE,WCE) = {1:(12,20),2:(10,30),3:(24,40)}A = 100, K=1, fref=1, =2

Execution Type

<1, 3, 2> 42.094 41.839 134%

<2, 3, 1> 37.482 36.978 119%

Ideal: always mean

31.443 (speed 0.68)

100%

Offline WCE 55.080 (speed 0.90)

175%

E[X]

E[X ]

E[X]%Ideal

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

5(11)

UBS in a NutshellUBS in a Nutshell

• Main ideas:– achieve a low speed ASAP by ordering tasks wisely– approximate by

• Priority:

• Observations:– prioritize short tasks– prioritize tasks with large variation in execution– prioritize power efficient tasks– algorithmic complexity O(N2) for ordering N tasks

E[X]

E[X ]

pubs(k) KkX k

s sk

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

6(11)

UBS vs. Full SearchUBS vs. Full Search

• 300 sets of each size(3,4,5,6 tasks)

• used the “real” E formula (4)

• under 2% difference

aa

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

2 3 4 5 6 7

% of No Scaling Avg. Energy

Task set size

UBSFull Search

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

7(11)

The Test Platform: EVB80200The Test Platform: EVB80200

•Intel i80200 (XScale)•MAX1855 voltage regulator•32MB SDRAM, 4MB Flash•RS232, JTAG, 7segLED

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

8(11)

UBS example on i80200: m6UBS example on i80200: m6

aa

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0

Power (W)

Time (s)

UBS (21.11mJ)Reverse UBS (24.34mJ)

Random (22.78mJ)WCE-stretch (28.58mJ)

MAX (37.57mJ)

• 6 tasks2 LZ (K=770mW)2 QS (K=840mW)2 FOR (K=800mW)

•max speed time 49ms•variation 17ms•runtime rescheduling after every 5 x H

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

9(11)

More Experimental Results: More Experimental Results: m6m6

0

5

10

15

20

25

30

35

40

45

0% 25% 50% 75%

Deadline extension

Hyp

erp

eri

od

En

erg

y (

mJ)

Max WCE-S R-UBS Rand UBS

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

10(11)

Experimental Results: m15Experimental Results: m15

0

5

10

15

20

25

30

35

40

45

0% 25% 50% 75%

Deadline extension

Hyperp

eri

od E

nerg

y (

mJ)

Max WCE-S R-UBS Rand UBS

5x LZ5x QS5x FOR

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

11(11)

Summary & ConclusionsSummary & Conclusions

• use more information to derive better methods• UBS: runtime, non-intrusive ordering for tasks

with variable execution time• measurements on

– a real platform: EVB80200– realistic tasks: Lempel-Ziv codec & Quicksort

• execution order matters!• random reordering: OK• UBS strategy: BEST

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

Thank You!Thank You!

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

13(11)

Measuring the core PowerMeasuring the core Power

aa

+-

9V470nF

470nF

diff

trimoffset

to Oscilloscope

LF351

Vcc

Vreg

3

2

Icc

U = Icc x 1

Vin = 10/11 Vreg

trim

9V

0.1

1k

1k

1k

5.1k

4.3k

51

4

7

10k

GND

CPU

-

+

-

+

-

+

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

14(11)

i80200 I/O Poweri80200 I/O Power

aaa

1. Assume system driving one PC-100 DIMM and a companion chip with 10pF/pin capacitance.

Bus Speed vs. Power (Moderat e BusUtilization)

0.00

0.10

0.20

0.30

0.40

0.50

0.60

3.00 3.10 3.20 3.30 3.40 3.50 3.60

Vccp (Volts)

Pin Power (Watts)

66MHz bus

100MHz bus

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

15(11)

i80200 Core Poweri80200 Core Power

aa

Cor e P ower ( on Dhry st one 2 .1: h ig h c ore a c t iv it y )

0.000

0.100

0.200

0.300

0.400

0.500

0.600

0.700

0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6Vcc (Volts)

Core Power (Watts)

400MHz

600MHz

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

16(11)

UBO extension to EDFUBO extension to EDF

• Use preemption to extract regions• Push forward uncertain regions• Algorithm:

1. Start from the latest deadline2. Between two consecutive deadlines order

the regions according to the already given priorities

3. Preempt the task which does not fit entirely4. Proceed with the next consecutive deadlines

19/04/23 ISLPED 2003 --- Uncertainty Based Scheduling…

17(11)

An Example of UBO EDFAn Example of UBO EDF

m=3

mean=9Task 1

Task 2

WCE=10

WCE=6

T=16

D=15

D=16

Classic EDF10 6

Preemption for Reduced Energy

Reordering105 + 1

In the long run: 18% less energy than for the classic EDF!