Você está na página 1de 4

2010 3rd International Conference on Advanced Computer Theory and Engineering(ICACTE)

A Dynamic Voltage Scaling Algorithm for Wireless Sensor Networks

Wei Tuming,Yu Sijia, Wei Hailong


School of Mechatronics Engineering University of Electronic Science and Technology of China
Chengdu, China
e-mail: {wtm71.sing_common.weihailong2008}@163.com

Abstract-A wireless sensor node is often powered by battery and clock frequency depending on the workloads can reduce
which is not easily replaced, so researching how to use its the sensor node's power consumption effectively.
limited energy effectively is the meaningful thing for wireless Prior DVS techniques had been demonstrated how to
sensor networks (WSNs). Dynamic voltage scaling (DVS) has obtain significant energy savings for time constrained
become a promising way for wireless sensor networks to systems [3,4,6,9,10,11,12]. However, these algorithms did
exploit multiple voltage and frequency levels and prolong the
not perform well in wireless sensor networks for these
sensor node's life. However, pure DVS for embedded systems
features in which the multihop routing and topology changed
did not perform well without considering the feature of
easily. So the objective of this paper is to develop a novel
wireless sensor networks' workloads. This paper proposed a
DVS technique according to the dynamic changing
task-driven feedback dynamic voltage scaling algorithm based
workloads of wireless sensor networks.
on the multihop routing and topology changed easily could
scale the working frequency and voltage levels dynamically
This paper is structured as follows. In Section II, sensor
according to the workloads of sensor node, fix the errors node task model, execution time model, energy consumption
through feedback scheme and reduce the node's power model and PID feedback control model are discussed. In
consumption at the premise of real-time tasks. The results of Section III, our DVS algorithm is detailed after describing
simulation indicated that the improved algorithm could the features of wireless sensor networks, and an example is
effectively reduce almost 30% more energy consumption than showing how our scheme works on practical task sets.
previous dynamic voltage scaling algorithm and prolong the Section IV presents the experimental results of different
life of wireless sensor networks significantly.
algorithms and sub-tasks in task sets. Finally, Section V
Keywords- dynamic voltage scaling; energy efficiency; power presents our conclusions.
optimization; wireless sensor networks
II. MATHEMATICAL MODELING

In order to describe the algorithm for the constituent


I. INTRODUCTION
elements better, there is a mathematical modeling for our
A large number of wireless sensor nodes distributed in DVS. Before modeling, the processor's frequency levels are
certain regions are the components of wireless sensor normalized: sensor node processor frequency levels range is
networks ( WSNs); each sensor node consists of the data
collecting, processing and transferring modules. Since
! fm'" , fm., ] and the scaling parameter is a
mox
' then the
= ;;.
wireless sensor node is generally powered by battery of frequency levels range is [ amin,amox ] in
which the energy is limited, especially in the harsh
environment the battery change is very difficult, the node's which amin = f
m ;;'
,",
x ,a""" = 1. Specific models are as follows:
power consumption determines the life-cycle of the wireless
sensor networks. Thus the energy consumption of WSNs has A. Sensor Node Task Model
become one of the main research topics. The task of a WSNs' single senor node can be expressed
To prolong the life-cycle of the WSNs, the dynamic as 'F;(/3"CpP;,y,) in which /3, denotes task level. When
power management (DPM)[1,2] and dynamic voltage scaling
(DVS)[3,4] can be implemented to save energy. DPM
/3, = I , the task 'F; is an urgent task that requires processing at
reduces the energy consumption by shutting down most of the maximum frequency; on the contrary, task 'F; needs
the free components of the system when there is no task and dynamic voltage scaling for its /3, = 0 . The C, is the worst­
waking up at proper time. This is commonly used within case execution time of task 'F; , the P; is the period of task 'F;
non-real-time system for its probabilistic characteristics. And
DVS, which can solve the embedded real-time system and the y, means the distance from the sink node. According
performance and meet power consumption requirements, is to the characteristics of wireless sensor networks, the
saving power in the working state without sacrificing the following hypothesis can be drawn: the closer a sensor node
performance of sensor node. As the power consumption of is from the sink node, the heavier the workloads are.
the CMOS integrated circuit [5] is proportional to clock Relatively these is a shorter period to sleep and the time
frequency and supply voltage, the scaling of supply voltage scaled by DVS adjustment is longer; whereas the task's

978-1-4244-6542-2/$26.00 © 2010 IEEE VI-554


2010 3rd International Conference on Advanced Computer Theory and Engineering(ICACTE)

workloads are lighter and the sensor node sleeps longer and So in practical application it is necessary to balance
DVS adjusts it in a shorter time period. between dynamic voltage scaling to reduce energy
consumption and the threshold delay.
B. Execution Time Model
To meet the real-time demand of the wireless sensor D. PID Feedback Control Model
networks, the model uses the Earliest Deadline First This model reduces the error between the assigned
scheduling ( EDF) [6] which means the highest priority job is execution time and the actual execution time by using the
the one with the earliest deadline. The scaling parameter a is PID feedback control. In order to use resources efficiently,
considered and the modified EDF scheduling policy is the PID control revises the error periodically. As the
workloads are inversely proportional to the y, , the period
a-ICk + L :5
1C'
(1)
It 'E{I,.,nl'*k} P, p (y,) can be set according to y, . The system error is defined
Only the highest priority task is scaled and the remaining as
tasks are modeled to execute at the maximum frequency 1m . e(t)=cij -CiA (10)
Let task Ck splits into CkA and CkB in which CkA is the And the PID control model is
execution time scaled by DVS and CkE is the execution time de(t)
�e(t)= Kpe(t)+� f e(t)dt+ KD (11)
at the maximum frequency 1m' The Sk is the slack available K] dt
to task T" .From The next period assigned execution time of the same task is
(2) CA'(j+l)=CAij+�e(t) (12)

(3) III. ESTABLISHM ENT O F DVS A LGORI THM FOR WSNS


we can derive that:
A. The Features of Wireless Wensor Networks
CkA
a= __ _ (4) The one of the important features of wireless sensor
CkA +Sk
networks is multi-hop routing [8]. In general, one sensor
Equation (3) shows that the scaling factor a depends on node communicates with its neighbor node directly in
the amount of available slack Sk and the execution time CkA is networks. If the node wants to communicate with the node
assigned to task Tk• To make sure the deadline for the task, outside the coverage of sensor node's radio frequency, it
when the current task is interrupted the most conservative needs an intermediate node for routing. Another feature is
time for the task is the networks variable topology. Environmental factors,
(5) energy depletion, the sensor nodes end or sleep or the new
nodes join may lead to the network topology change. It is
where the cy is the actual execution time in the above more likely for the nodes from the sink node closer be
equation. The available slack time of the next task from the chosen as the intermediate node for routing and so their
previous one is workloads would be heavier.
Sk" =Sk - Er;j (6 )
B. The DVS Algorithnm for WSNs
We suppose that the idle state of CPU is the
The core idea of our scheme proposed in this paper is to
implementation of a space mission, so t'dle ( dk,dk+l) can be
scale the feedback interval in accordance with the
seen as the total time spent on space missions between heterogeneity of the WSNs sensor node's workloads and fix
deadline dk and dk+1• The available slack for the next task is the frequency and voltage levels in the premise of the real­
Sk+1 =Sk" +tidle (dk,dk+1 ) (7) time task and modify the assigned execution time through
the feedback scheme.
C. Energy Consumption Model The task set in our DVS consists of n-tasks (7;,7;", Tn )
COMS integrated circuit energy consumption [7] which and the task r; is r; (fJ"C"P"y,) which was detailed in the
consists of dynamic and static power consumption is
sensor node task model and the deadline of task r; is d,. The
Energy CVd� + Vddileakllt (8)
Tij is the j-th of task r; , its release time is P, x (J -I) and the
oc

where the Vdd is ( the) supply voltage and the Ileak is the
deadline is P, x j . Defining the previous task before task
leakage current. Compared to dynamic power, static power is
usually very small and can be neglected, so the equation ( 8) Tij is Tp k ' In the initial state, CkA =Ck/ 2 , ETkO =Ck, Sk =0, the
is modified as Energy C V� . We can see that the energy
oc utilization is U =CtiP'+C2/Pz'" +C./p". In fact, the initial
consumption is proportional to the voltage square. When CkA can be any values in [o,ckl because the later one the
reducing the supply voltage to save sensor node energy, the
feedback scheme will adjust CkA to make it approximate to
threshold delay is extended:
Vdd 2 the task's actual execution time. Generally the CkA can be
Delay =k (9) initialized as 50% of the worst initial execution time.
( dd -v,)
V
Detailed algorithm is as follows:

VI-555
2010 3rd International Conference on Advanced Computer Theory and Engineering(ICACTE)

Firstly detect the value of /3, . If p, = I , then a f- am . low frequency level and the energy consumption is
Generally the maxmum scaling factor am = I . Then, after the minimized.

EDF scheduling, the task sets are proceeded as follows: 100010


75%
if processor was idle for t'd/,then 50%
25%

if Tp, interrupted then


El'y f-Clj -clj xa 2 idle for Tl 1 idle forTI

s", f-S, -t'd/, (dk ' d'+I ) Tl T2 TI.A T1.1l T3 1'2 TI T3

if El'y < s, then


100%
75%

s, f-S, -El'y
50%

else s, f-0
25%
i Jl•••IJ••=iI••L.IIL
_
IO IS 20 25 30

if now> dp, then Figure 1. A set of 3-tasks scaled by our DVS

s, f-S, +tid!, (d" dk+l)


In order to save as much energy as possible, the scaling

{iL, ?:�}
IV. SIMULATION AND ANALYSIS

factor a is set by min ... ,1m I f. and the In order to verify the DVS scheme we proposed energy
1m 1m 1m CAy +S, saving effect, the experiments are designed and simulated on
assigned execution time CAlf is skxa/(I-a) ; but if a=l, MATLAB. The WSNs' sensor node works at the normalized
frequency and voltage level. We assume a processor model
the CAlf should be set as 0 . Operating frequency f is
capable of operating at different voltage, frequency and
determined by aXlm . When some task is completed, corresponding energy consumption depicted in Table 1.
feedback scheme is used and some settings are needed for WSNs sensor nodes distribute in a circular area whose centre
the nest task processing: is a sink node showed in Figure 2. We select randomly
s, f-S, -clf +C, sensor nodes in simulation according different radius and set
the feedback coefficient as Kp 0.9, K[ 0.08, KD 0.1.= = =
t: f-clj -CAlj
TABLE I Processor model
I t:(t)-t:(
'
tI -DW)
!l.t:f-Kt:(
P
tI ) +-I IWt:(t)
'
+KD
K[ DW Nonnalized frequency Voltage(V) Energy consumption
CAy f-CAlf +!:It:
LOOO L65 LOOO
C. Example
0857 160 0806
We tum to a description of the entire DVS algorithm
proposed in the previous section by combining all the 0714 1 SO o S90
techniques illustrated above. The set consists of three tasks o S71 140 0411
T, {3,8}, T, {3,1O}, T, {I,14} and a blank task 1 {1, 4} ,
= = = =

0.428 1.25 0.246


where we consider this set is the workloads in one WSNs
senor node and the deadline of each task is equal to its period 0.286 110 0.127
as dij P, . All these task levels are set as zero, which means
=

these tasks are scaled by DVS.


0.6 .
As shown in Figure 1, as T,.o is first scheduled, no
previous slack is passed to T,.o and the available idle slacks
are two. The T"o 's initial CA is set to 1.5 and the frequency
scaling factor a is 0.428 according to the equation (4). Then
the frequency scaled as 0.428 1m' Since T"o is completed at
time 4, it passes on one unused slack to the next task T"o . Its
CA is also updated at that time according to the feedback
scheme. Besides of the slack from the idle task, T"o also gets
one slack from and its frequency level is determined in the .1 '-_----,.�=�____=---"t=--___'
·1 .() . 5
same way as the first task. For the following tasks, the
feedback scheme sets to approximate the task's actual
execution time. By this way the entire task set is scaled at a Figure 2. WSNs sensor nodes distribution

VI-556
2010 3rd International Conference on Advanced Computer Theory and Engineering(ICACTE)

The Look ahead RT-DVS proposed by Pillai [9] and Shin V. CONCLUSIONS
meets the requirement of real-time system of periodic task
This paper presents a novel DVS technique with
based on the EDF algorithm for dynamic voltage scaling and feedback scheme and EDF scheduling which adjusts the
saves energy efficiently. The DVSST [10] was the voltage
sensor node's frequency and voltage levels in accordance
scaling algorithm designed by Qadi etc for sporadic tasks
with the sensor node's workloads after we establish the
dynamic, which assumed that the interoperability among
mathematical models for our DVS algorithm. Moreover, the
tasks resources and could be used for wireless sensor
relevant experiments are designed and the result of
networks under non-intensive tasks. Figure 3 compares our
simulations shows that our DVS scheme is able to adapt to
DVS energy consumption with the Look ahead RT-DVS and
dynamically fluctuating workloads better than previous
DVSST. When the task set utilization is less than 0.3, it is
DVS algorithms in WSNs, saving up to about 30%
showed that three schemes consume the same amount of
additional energy. Also it can handle with simple and
power because the set with low utilization usually has
complex task sets efficiently. However, this scheme is
enough slack and idle slack for all tasks complete at the low
suitable for WSNs with real-time requirement but is not
frequency level before their deadline. With the increase of
suitable for non-real time system in WSNs.
the utilization, our DVS scheme saves 8%-24% more energy
than the Look ahead RT-DVS and DVSST.
REFERENCES

[I] A Zuquim. et al,"Effcient power management in real time embedded


systems," IEEE International Conference on Emerging Technologies
and Factory Automation- ETFA'03, 2003.
[2] A Sinhua, A Chandrakasan, "Dynamic power management in wireless
sensor network, " IEEE Design and Test of Computer,2001,18(2):62-
74.
[3] F Gruian, "Hard real-time scheduling for low-energy using stochastic
data and DVS processors," The International Symposium on Low­
Power Electronics and Design, Aug. 200I, pp.46-51.
[4] D Kang, S Crago, J Suh, "A Fast Resource Synthesis Technique for
Energy-Efficient Real-Time Systems, " IEEE Reak-Time Systems
Utilization Symposium,2002.

Figure 3. Comparation among different schemes [5] A Chandrakasan, S Sheng, R Brodersen, "Low power cmos digital
design, " IEEE Journal of Solid-State Circuits, 1992, 27: 473-484
Figure 4 illustrates the performance of our DVS under [6] Y Zhu,F Mueller, "Feedback EDF Scheduling Exploiting Dynamic
the different number of sub-tasks in the task set. The task set Voltage Scaling," Proceedings of the 10th IEEE Real-Time and
consists of the sub-tasks in a certain sequence and it is clear Embedded Technology and Applications Symposium, Washington
DC, 2004:84-93.
that the more sub-tasks the sensor node has, the more
[7] A Sinhua, A Chandrakasan, "Dynamic power management in wireless
complex workloads it need to process. In this experiment we
sensor network," IEEE Design and Test of Computer,2001,18(2):62-
compare the energy consumption of3-tasks, 1O-tasks and 15- 74.
tasks in the task set. It can be seen that 3-tasks consumes less [8] M Ilyas, I Mahgoub, "Handbook of sensor network:compact wireless
energy than IO-tasks and IS-tasks and in these cases the and wired sensing systems, " New York,CRC Press,2005.
energy using is small when the utilization is less than 0.3, up [9] D. P Pillai, K Shin, "Real-Time Dynamic Voltage Scaling for Low­
to about 90% energy saving; The 1O-tasks and IS-tasks Power Embedded Operating Systems," Proceedings of the eighteenth
energy consumption are almost equal to nearly 70% energy ACM symposium on Operating systems principles, 2001:89-102.
savings. It can be concluded that the DVS algorithm we [10] A Qadi, S Goddard, S Farritor, "A Dynamic Voltage Scaling
proposed can reduce the energy consumption of WSNs Algorithm for Sporadic Tasks," Proceedings of the 24th IEEE
International Real-Time Systems Symposium, 2003:52.
sensor node in the cases of simple and complex task sets.
[11] J Zhuo, C Chakrabarti, "Energy-efficient dynamic task scheduling
algorithms for DVS systems," ACM Trans Embedded Comput
Syst,vol:7(2),2008.
[12] YC Lee, AY Zomaya,"Minimizing energy consumption for
precedence-constrained applications using dynamic voltage scaling,"
2009 9th IEEE/ACM International Symposium on Cluster Computing
1"O.4 and the Grid, 2009:92-99.
j

\Jtilizalioll

Figure 4. Energy consumption under different subtasks

VI-557

Você também pode gostar