Jiwon HahnECE295 SeminarDecember 2, 2002

Integrated Management of Power Aware Computing & Communication Technologies

IMPACCT Project People

Faculty: Pai Chou, Nader Bagherzadeh

Students: Jinfeng Liu, Dexin Li, Bita Gorji-Ara, Duan Tran, and

Jiwon Hahn

Collaborator NASA JPL, Rockwell Collins, ISI

Sponsors DARPA PAC/C, Broadcom, HP

Outline What is IMPACCT? Motivation How it works Tools Experiments and Results Conclusions and Future Works

What is IMPACCT? A CAD tool for exploring

power/performance tradeoffs A new technique that performs component,

system, and mission-level integrated power management

Target applications Mars pathfinder, ATR, UCAV,…

Motivation Embedded Systems

Computers inside devices PDA, cellphone, camera, vehicles, robots,…

Power management Power-Aware vs. Low-power Mission-Aware: meet the constraints

High-level approach Amdahl’s law: Power saving of a component must be

scaled by its percentage contribution to entire system Evaluate combined effects in the context of system Higher abstraction level enables global optimum

How it works Hierarchical power management System-level power scheduling

Power-aware scheduling Mode selection

Mission-level power scheduling Schedule Selection

Scheduling & Mode Selection Power Aware Scheduling

Schedules tasks, meeting timing and power constraints Output: initial schedule Static scheduling/planning [DAC'01]

Mode Selection Selects resource modes of each task considering mode

dependency Minimize energy consumption Output: mode schedule

Mode selection/modeling [ASPDAC’02] Winner Best Student Paper Award

Schedule Selection Goal

To generate a mission level schedule which Adapts to variable power constraint Considers higher-level context change overhead Meets the global deadline

Assumptions A mission contains one or more applications It is static

Schedule Selection Problem:

Select N schedules among M different

schedules, by deadline D, under the maximum

power curve P. Minimize energy

considering overhead

0 1 2 … M…

Schedule Set

D

P

Put N schedules here!

0 1 … M

0

1

…

M

Overhead MatrixCurrent schedule

Previous Schedule

Schedule Selection(cont.) Parameters: t,n,k,m

t: timestamp. (discrete value) 0tD

n: schedule count excluding S0 1nN k: schedule count including S0.

(for tracking the selected path) 1kD m:current schedule 0mM

Schedule Selection(cont.) Algorithm: 4D Dynamic Programming

n

k

m

t = D

t =2

t =1

n

m

t

Idea: Reach the global optimum by keeping track of optimal solutions of subproblems

Optimal substructure For some(k,m), min{E(t,n,m)} contains the optimal value.

Space D3 for keeping optimal Energy

D4 for bookkeeping indices

Speed O(D3) – polynomial!

Could be optimized for speed-up

Energy CubeBookkeeping

Cubes

Schedule Selection(cont.) Notation

Set of schedule, S= {S0, S1, … ,SM} Time period of each schedule: Ts[0…M] Power level of each schedule: Ps[0…M] Energy of each schedule: Ts x Ps = Es[0…M] Schedule-switch overheads from Si to Sj:

Po(i,j), To(i,j), Eo(i,j)

Schedule Selection(cont.) Algorithm

Initialization If (t,n,k,m) is the first possible selection,

E(t,n,k,m) = directly calculated energy Else

E(t,n,k,m) = Process

if m!=0, E(t,n,k,m)=E(t’,n-1,k-1,m’)+Eo(m’,m)+Ps(m)Ts(m) t = t’+To(m’,m)+Ts(m)

if m=0, E(t,n,k,m)= E(t’,n,k-1,m’)+Eo(m’,m)+Ps(m)Ts(m)

System-level Input:

Application Model ports, channels,…

Architecture Model System architecture template Component library + mode dependency model(MDM)

Constraints Power and Timing

Output: Mode Schedule

Mission-Level

Mission-level Constraints: Power and Deadline

Mission Schedule

System-level1

App. Model Constraints

Sys. Arch. Template

Component Library+ MDM

Mode Schedule

1)

Schedule Selector

Mission-Level2

Mission Constraints

Overhead Calculator

Mission Schedule

MS MS MS MS MS …Schedule Collector

2

Scheduler

Initial Schedule

Mode Selector

1

System-level• Initial Schedule:

• Mode Schedule:

Mission-Level

Mission Power Constraint Curve

Mission Deadline

• Mission Schedule:

Tools Scheduler (Jinfeng) Mode Selector (Dexin) Schedule Selector (Jiwon) Etc.. Programs and tutorial are available

Tools(I): Scheduler

Tools(II): Mode Selector

Tools(III): Schedule Selector

Mars Rover Comparison over 3 scenarios

Overall mission 3 power scenarios: best, typical, worst, 10 min each 48 steps

Power-aware schedules Accelerated speed by tracking available power Finished earlier before working in the worst case 33% faster, 32.7% less energy cost

Experiments and Results I

Experiments and Results I(cont.)

Experiments and Results II Mars Rover

Behaviors and tasks Moving around on Mars surface Communicating with the Lander Taking pictures Performing scientific experiments

Components in the entire system Hazard detector, Driving motor, Steer motor,

Radio frequency modem (RF), Camera (CAM), Microprocessor (PPC), Micro-controller

Experiments and Results II(cont.) On/off only

Relaxed constraints Mode change overhead No max power constraint

Mode selection Energy saving:

From 6.9% to 49.3% average 26.5%

Meets max power

0

100

200

300

400

500

600

700

800

900

1000

5 6 7 8 9 10

Mode SelectionOn&Off

Pmin (W)

Energy (J)

Experiments and Results III Example of Schedule Selection

0 1 2 3

0

1s 1s 3s 4s

0J 4J 7J 6J

0W 4W 21W 24W

1 ?

1s

? ?4J

4W

2 ?

2s

? ?5J

10W

3 ?

3s

? ?10J

30W

20s

30-3 13-1

10s, 5J, 50W 5s, 10J,

50W

3s, 15J, 45W

0 1 2 3

• Schedule Set:• Overhead Matrix

Experiments and Results III(cont.)

30-3 13-1

20-2 12-1

1 120s

20s

20s

Low Power

Greedy

Dynamic Programming

Exceed deadline

Energy = 149W Nonoptimal solution

Energy = 131W Optimal solution!

Experiments and Results III(cont.)

Run-time Deadline Guarantee

Energy optimization

Manual N/A X X

Low Power constant X X

Greedy linear X X

Enumeration exponential O O

Dynamic Programming

polynomial O O

Conclusions IMPACCT

greatly expands the range of power/performance trade-offs

effectively integrates existing power management techniques

models system-level dependencies saves great amount of energy consumption while

meeting all constraints proposes novel hierarchical power management

technique

Current & Future Works Ongoing work

Architecture Modeling(Dexin) Mission-level Power Management(Jiwon) Extended experiments on broadcom and itsy

board(Jinfeng) Applying to different application: SDR(Bita)

Future work Dynamic Power Management Mixed application schedule selection More applications

Reference Pai H. Chou, Jinfeng Liu, Dexin Li, and Nader Bagherzadeh, “IMPACCT: Methodology and Tools for Power-

Aware Embedded Systems”, Design Automation for Embedded Systems J. Liu, P. Chou, N. Bagherzadeh, and F. Kurdahi. “Power-aware scheduling under timing constraints for

mission-critical embedded systems”. In Proc. 38th Design Automation Conference, pages 840–845, June 2001 D. Li, P. Chou, and N. Bagherzadeh. Mode selection and mode-dependency modeling for power-aware

embedded systems. In Proc. 7th Asia South Pacific Design Automation Conference, pages 697–704, January 2002

J. Hahn, P. Chou, and N. Bagherzadeh, “Tutorial: IMPACCT Tool v1.0”, University of California at Irvine, August, 2002

THANK YOU!

THANK YOU!