Transcript Slide 1

Battery Aware Dynamic
Scheduling for Periodic
Task Graphs
Venkat Rao #, Nicolas Navet #, Gaurav Singhal *, Anshul
Kumar, GS Visweswaran
#TRIO
Group, INRIA-Lorraine /LORIA.
* Dept of ECE, UT Austin, Dept of CSE, IIT Delhi
Dept of EE, IIT Delhi
Introduction
Mobile Embedded Systems Design :



Battery lifetime is major constraint
Slow growth in energy densities not keeping up with
increase in power consumption
Extension of battery lifetime and not just low
energy design the REAL GOAL
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Traditional approaches to energy optimization



CMOS Energy and power
Energy α V2
Power α V2.f
fmax α V
Dynamic Voltage Scaling (DVS):
busy system => increase Vdd, frequency
idle system => decrease Vdd, frequency
Potential to achieve quadratic energy and cubic power
savings.
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Variable-supply Architectures
 High-efficiency
adjustable DC-DC converter
 View from battery side
Vbat is constant and depends on battery technology( 1.2 V
for NiMh, 3.6-4.2 V for Li ion)
 High Vdd translates to high Ibat `

SoC
Vsys X Isys = µ X Vbat X Ibat
Power Manager
Clkgen
Vsy
s
WK
to f
f to
Vdd
Isy
s
Switching
DCDC
regulator
Vset
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Vbat
Ibat
Battery
Battery Basics
 Battery characterized by Voc and Vcut.
Electron Flow
_
Load
A
n
o
d
e
Positive
Ions
Electrolyte
+
C
a
t
h
o
d
e
 Battery lifetime governed by active species
concentration at electrode-electrolyte
interface.
 Phenomenon governing battery lifetime:
 “Rate Capacity Effect”
“high load current implies lower charge
delivered.”
 “Recovery Effect”
“charge recovered by giving idle slots”
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Diffusion Model
- Rakhmatov, Vrudula et al.
After a recent
discharge
Fully charged
battery
Fully discharged
After
Recovery
Electro-active
species


Analytically very sound but computationally intensive
Cannot be used for online scheduling decisions.
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Battery Aware Scheduling
Guideline 1: For a set of schedulable tasks (t0,
t1……tN) having corresponding currents costs (I0,
I1……IN) scheduling them in decreasing order of
current costs is the optimum battery
solution.[Rakhmatov03]

Ibat
time
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Battery Aware Scheduling
Guideline 2: For a given task t to be executed before a
given deadline d its better to lower the frequency and
run without giving an idle slot than give an idle slot and
run at a higher frequency.[Rakhmatov03]
freq

freq
idle
time
WDPRTS, Rhode Island
Greece 25th April 2006
d
time
Venkat Rao – INRIA Lorraine /LORIA
d
Problem Definition
To find a battery efficient schedule for a given a set of periodic tasks
graphs (T1, T2, ....Tn) which have corresponding deadlines (D1,D2,
.....Dn) equal to their periods, where a taskgraph Ti comprises of any
m interdependent nodes, each of which are in themselves tasks with
given worst case computations (wci1, wci2, ......wcim).
Precendence
constraint
WDPRTS, Rhode Island
Greece 25th April 2006
wci
T1
D1
T2
D2
Venkat Rao – INRIA Lorraine /LORIA
T3
D3
Our Methodology

There are 2 aspects to the problem
Global Frequency Setting
 Local order of execution of nodes

Task Graphs
WCi’s
nodes
Di’s
DVS Algorithm
Priority function for
max slack recovery
WDPRTS, Rhode Island
Greece 25th April 2006
Ready list
Frequency Setting
Local Task Order
Venkat Rao – INRIA Lorraine /LORIA
fcurr
next
node
Global Frequency Setting

To calculate the min frequency that can ensure all subsequent
deadlines are met.
upon release( Taskgraph Ti )
Modified ccEDF
algorithm from
[pillai01]
1: WCi =  wcij
2: select_frequency( )
upon end_of_node( τij )
1: WCi = WCi + acij − wcij
wcij
WCET of the jth node of
the ith task graph at fmax
2: select frequency( )
acij
Actual exec time for jth node
of the ith task graph at fmax
Di
Deadline for the ith task
graph
τij
The jth node of the ith task
graph whose execution just
ended.
select_frequency ( )
1: U =  WCi/Di
2: fref = U × Fmax , return fref
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Global Frequency Setting



Follows EDF so works up to U= 100%
Ensures all deadlines are met.
Ensures a Non Increasing discharge profile for
set of jobs (set of instances of periodic tasks)
re-computing speed
freq
time
WDPRTS, Rhode Island
Greece 25th April 2006
d
Venkat Rao – INRIA Lorraine /LORIA
Local order of execution

Slack Recovery maximization.
Worst case seldom arrives leading to dynamic slack
 Order of execution effects dynamic slack recovery
 Important to choose the order optimally
 A priority function needs to be chosen
 Heuristics like LTF and STF work well in specific
cases
 pUBS : a near optimal priority function from
[Gruian02]

WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Ready List

Ready list comprising of nodes from current(EDF)
Task graph only.
D2
D1
D3
Ready list
D1 < D2 < D3
Execute
WDPRTS, Rhode Island
Greece 25th April 2006
Priority function
Venkat Rao – INRIA Lorraine /LORIA
Ready list comprising of nodes
from current Task graph only

Advantages :
Follows EDF so ensures meeting of deadlines
 Simple to implement


Disadvantages :
Limited choice for the priority function.
 Limited slack recovery.

WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Ready List

Ready list comprising of nodes from all released Task
graphs.
D1
D2
D3
Ready list
D1 < D2 < D3
Execute
WDPRTS, Rhode Island
Greece 25th April 2006
Priority function
Venkat Rao – INRIA Lorraine /LORIA
Ready list comprising of nodes
from all released Task graphs

Advantages :
More choice for the priority function.
 Better slack recovery hence lower energy
consumption


Disadvantages :

Out of EDF execution hence deadline can be
missed
Need For additional feasibility check
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Ready List

Ready list comprising of nodes from all released Task
graphs.
D1
D2
D3
Ready list
D1 < D2 < D3
Execute
WDPRTS, Rhode Island
Greece 25th April 2006
Feasibility check
Priority function
Venkat Rao – INRIA Lorraine /LORIA
Feasibility check



Check to ensure that an out of EDF execution will not cause a
deadline miss
Or more stringently will not cause the raising of frequency later
for meeting deadlines
For task belonging to EDF order k, k-1 checks are required.
Feasibility Check ( tij )
flag= 1;
for (k=1 to j-1)
{
if (WCk +wcij > fcurr X Dk – Tcurr )
Flag =0;
}
return flag
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Simulations






C simulations were conducted to test our methodology
The DVS enabled processor simulated supports the following 3
frequency-voltage tuples [(0.5GHz,3 V), (0.75GHz,4V),
(1.0GHz,5V)].
Task graphs were generated from TGFF with random
dependencies
Utilization of the system was kept to 70%
Stochastic battery model from [G.Singhal05] was used to
estimate battery life for the profiles generated by various
scheduling algorithms
Simulated for NiMH AAA Panasonic batteries with max capacity
of 2000mAh and nominal capacity of 1600mAh
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Simulation Results :
Battery lifetime and charge delivered.


Results were obtained by averaging performance of the various algorithms over
100 random taskgraph sets
Battery Aware Schedule 2 delivers maximum battery life amongst the schemes
compared
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Conclusion



We have presented a Battery-aware Scheduling
Methodology that facilitates the combining of a good
DVS algorithm with a heuristic based priority function
for scheduling of taskgraphs.
Simulations suggest that our methodology performs up
to 47% better than ccEDF and upto 23.3% better than
laEDF scheduling schemes in terms of battery lifetime.
It can result in up to 100% improvement in battery
lifetime over systems with no DVS.
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
References and Credits
[1] V. Rao and G. Singhal. Integrated power management for embedded systems. Bachelors Thesis, Indian
Institute of Technology, Delhi, 2005.
[2] F.Yao, A Demers and S Shenkers. A Scheduling Model for Reduced CPU energy. IEEE 1995.
[3] P. Pillai and K. G.Shin. Real time dynamic voltage scaling for low powered embedded systems.
Operating Systems Review, 35:89–102, October 2001.
[4] S. Vrudhula and D. Rakhmatov. Energy management for battery powered embedded systems.
ACM Transactions on Embedded Computing Systems, pages 277– 324, August 2003.
[5] J. Luo and N. K. Jha. Battery-aware static scheduling for distributed real-time embedded systems.
In DAC’01: Proceedings of the 38th conference on Design automation, 2001.
[6] Gruian F., Energy-Centric Scheduling for Real-Time Systems, PhD thesis, Lund Institute of
Technology, 2002.
[7] V. Rao, G. Singhal, A. Kumar, and N. Navet. Battery model for embedded systems. In Proceedings of
International Conference on VLSI Design, pages 105–110, January 2005.
[8] V. Rao, G. Singhal, and A. Kumar. Real Time Dynamic Voltage Scaling for Embedded Systems. In
Proceedings of International Conference on VLSI Design, pages 650–653, January 2004.
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Thank You
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Battery Models
Advantages
Disadvantages
PDE
Accurate
Slow, involves a large
number of parameters
(higher forms
of KiBaM) Still Too computationally
intensive
Use
capacitorfor use
Notataccurate, elements
runtime
and resistors
to
change value
Circuit
represent battery depending conditions
Stochastic
WDPRTS, Rhode Island
Greece 25th April 2006
Relatively
Still in the process of
accurate and fast. development.
Venkat Rao – INRIA Lorraine /LORIA
Rate Capacity Effect
 Total charge delivered by the
battery goes down with the
increase in load current.
 Concentration of active species
at interface falls rapidly with
increasing load current.
Rate Capacity Effect
 Battery seems discharged when
the concentration at interface
becomes zero.
back
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Cell Voltage
Recovery Effect
Intermittent Discharge
Continuous
discharge
Elapsed time of discharge
Recovery Effect
 Battery recovers capacity if
given idle slots in between
discharges.
 Diffusion process
compensates for the low
concentration near the
electrode.
 Battery can support further
discharge.
back
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Simulation Results:
Effect of ready list on energy consumption
At Utilization
70% and actual
computation
times varying
from 20% to
70%
Energy consumption (normalized w.r.t optimal schedule) by various
scheduling policies for different number of tasks in a taskgraph
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Simulation Results:
Effect of priority function on energy consumption
At Utilization
70% and actual
computation
times varying
from 20% to
70%. Ready list
comprises of
most imminent.
Energy consumption (normalized w.r.t optimal schedule) by various
scheduling policies for different number of tasks in a taskgraph
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA
Kinetic Battery Model



Simplest PDE model to explain both recovery and rate
capacity.
Available and Bound charge wells
Dynamic transfer of charges governed by a rate constant and
difference in heights.
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA


Introduction
Battery Basics
1.
2.





Rate Capacity Effect
Recovery Effect
Related Work : Review of relevant models
Scheduling Problem
Our Methodology.
Simulation and Results
Conclusion
WDPRTS, Rhode Island
Greece 25th April 2006
Venkat Rao – INRIA Lorraine /LORIA