A Holistic Approach for Optimizing Lifetime of

IEEE 802.15.4/ZigBee Networks with

Deterministic Guarantee of Real-Time Flows
Myung-Gon Park, Kang-Wook Kim, and Chang-Gun Lee
School of Computer Science and Engineering
Seoul National University
Seoul, Korea
Email: {mgpark, kwkim}@rubis.snu.ac.kr, cglee@snu.ac.kr
3
4
2
5
1
3
4
6
2
5
1
6
GTS for 3
GTS for 2
3
2
1
3
4
6
5
4
6
5
3
2
1
(2) power configuration
routing decision
(1) clustertree construction
(3) dutycycle scheduling
endtoend delay
network lifetime
GTS for 3
GTS for 5
GTS for 4
f1
f1
f2
f2
cluster3
cluster3
cluster4
cluster4
SD3
SD4
BI3
BI4
Figure 1. A holistic optimization of a ZigBee network
AbstractIEEE 802.15.4 is a global standard designed
for emerging applications in low-rate wireless personal area
networks (LR-WPANs). The standard provides nice features
such as a beacon-enabled mode and guaranteed time slots
for real-time data delivery. However, how to optimally operate
those features is still an open issue. For the optimal operation
of the features, this paper proposes a holistic optimization
method that jointly optimizes three cross-related problems:
(1) cluster-tree construction, (2) nodes power conguration,
and (3) duty-cycle scheduling. Our holistic optimization method
nds the solution for the three problems such that all the real-
time packets can be delivered within their deadlines in the
most energy-efcient way. Our simulation study shows that,
comparing with existing methods, our holistic optimization
can guarantee the on-time delivery of all real-time packets
while signicantly saving the energy and hence signicantly
increasing the network lifetime.
Keywords-real-time; sensor network; holistic optimization
I. INTRODUCTION
IEEE 802.15.4 is a global standard for emerging ap-
plications in low-rate wireless sensor networks. Its target-
ing applications include health monitoring, disaster detec-
tion/reporting, target tracking, and factory automation. In
those applications, a number of real-time data ows are
ongoing and their time-sensitive packets need to be deter-
ministically delivered on-time. This real time guarantee must
be provided with an energy-efcient manner such that the
network lifetime can be maximized. Such energy-efcient
real-time guarantee is important because the replacement of
batteries of sensor nodes is not only very cumbersome but
also practically impossible in some applications such as a
densely deployed large-scale sensor network.
For the energy efciency and the real-time guarantee, the
beacon-enabled mode of IEEE 802.15.4 provides nice fea-
tures such as synchronized operations with small duty cycles
and guaranteed time slots for collision-free transmissions.
However, how to optimally utilize those features is still an
open issue. In this paper, we propose a holistic approach to
optimally conguring the IEEE 802.15.4/ZigBee cluster-tree
network jointly addressing the three cross-related problems:
(1) logical cluster-tree construction, (2) power congura-
tion for nodes, and (3) duty-cycle scheduling of clusters.
Consider a sensor network that has six ZigBee nodes and
two real-time data ows as shown in Figure 1. In order to
guarantee the end-to-end deadline of every data packet of the
given ows, we rst have to construct the logical cluster-tree
such that packets can be routed along the tree structure. This
problem of logical cluster-tree construction is directly related
to the power conguration problem of all nodes since the
powers of child and parent nodes need to be properly cong-
ured such that their RF signals are bi-directionally reachable.
Once a cluster-tree is determined, we also have to determine
the duty cycle scheduling of all the clusters such that each
node in a cluster can send a packet using its dedicated
GTS (guaranteed time slot) within the clusters active period
called a SD (superframe duration), which periodically comes
at every BI (beacon interval). This duty-cycle scheduling
problem should also determine the lengths of SDs and BIs
of all the clusters and allocations of GTSs to all the nodes.
Thus, the duty-cycle scheduling problem is inter-dependent
with the cluster-tree construction. Furthermore, the resulting
cluster-tree and duty-cycle scheduling affect the end-to-
end delay of all the real-time ows. Also, the resulting
power-conguration and duty-cycle scheduling affect the
overall energy consumption and in turn the network lifetime.
Therefore, we need to jointly address all the three problems
together in order to guarantee all the real-time ows in
the most energy efcient way, as we propose in this paper.
Existing works try to solve only a part of these three cross-
related problems. For example, Han addresses only the duty-
cycle scheduling problem assuming that the logical cluster-
tree and power congurations are already given [1]. Ergen
et al. presents an energy efcient routing method with delay
guarantee for wireless sensor networks assuming that duty-
cycle scheduling and power congurations are already given
[2]. However, none of existing works propose a holistic
solution for optimally solving all the three cross-related
problems together.
The rest of this paper is organized as follows: The next
section surveys the related work. Then, Section III overviews
the IEEE 802.15.4/ZigBee standard and formally denes the
problem to be addressed in this paper. Section IV proposes
our holistic optimization framework. Section V presents our
experiments. Finally, Section VI concludes this paper.
II. RELATED WORK
Recently, the issues for real-time data delivery has been
extensively addressed in different settings of wireless sensor
networks. In a broad scope, real-time MAC protocols, real-
time routing protocols, and real-time MAC/routing cross-
layer protocols are proposed. For example, the implicit
EDF [3] is a hard real-time MAC protocol that provides a
collision-free real-time packet scheduling scheme exploiting
the periodic nature of real-time data ows. The dual-mode
real-time MAC protocol [4] provides both bounded worst
case delays for real-time packets and good average delays
for best-effort packets by switching between protected and
unprotected modes. On the other hand, RPAR [5] is an exam-
ple of real-time routing protocols that meets the application
specied delay bound requirements at a low energy cost
by dynamically conguring transmission power and routing
decisions. Ergen et al. [2] propose another routing method
that nds the optimal routing paths for real-time data ows
in a way of maximizing the network lifetime. The method
uses a linear programming formulation to nd the optimal
routing solution assuming that the power conguration and
the network topology are given as inputs. SPEED [6] is an
example of MAC/routing cross-layer protocols that support
real-time packet deliveries. SPEED is designed to provide
soft end-to-end deadline guarantees by enforcing a uniform
packet delivery speeds all over the entire network through
feedback control in the MAC layer and geographic packet
forwarding in the routing layer. MMSPEED [7] extends
the SPEED protocol by providing multiple speeds in order
to provide service differentiations for different classes of
real-time ows. All the above methods, however, are de-
signed for sensor network settings far different from IEEE
802.15.4/ZigBee standard. Thus, they cannot be applied to
build the optimal IEEE 802.15.4/ZigBee network. More
importantly, even in this broad scope, there is no exist-
ing work that holistically optimizes topology construction,
power conguration, and packet scheduling, all together.
In the specic scope of the IEEE 802.15.4/ZigBee stan-
dard [8], Koub aa et al. propose algorithms to schedule SDs
of all the clusters in a collision-free way for the given
values of SDs and BIs for all the clusters [9]. However,
they do not address how to optimally determine the SD
and BI values for all the cluster. In [1], Han addresses the
optimal duty-cycle scheduling problem, that is, optimally
nding the SD, BI values and also GTS allocations for
guaranteeing all the real-time ows while maximizing the
network lifetime. Hans approach, however, addresses only
the duty-cycle scheduling problem assuming that the cluster-
tree topology and power congurations are given as inputs.
None of the IEEE 802.15.4/ZigBee related work addresses
the holistic optimization problem that considers cluster-tree
construction, power conguration, and duty-cycle schedul-
ing, all together.
III. IEEE 802.15.4/ZIGBEE OVERVIEW AND PROBLEM
DESCRIPTION
The IEEE 802.15.4/ZigBee standard provides a beacon-
enabled mode for energy-efcient delivery of real-time
packets. In the beacon-enabled mode, all the ZigBee nodes
rst need to form a logical cluster-tree network. As an
example, Figure 2(a) shows physical deployments of ZigBee
nodes, denoted by N
1
, N
2
, , N
8
. The nodes form a logical
cluster-tree as in Figure 2(b) where N
1
, N
2
, and N
3
form
a cluster with N
3
as the cluster head, N
7
and N
8
form
another cluster with the head of N
7
, and N
3
, N
4
, N
5
, N
6
,
and N
7
form an upper level cluster with the head of N
4
.
This tree formation is necessary in two folds: (1) routing
of packets and (2) scheduling of packets. First, the routing
path of a packet from a source node N
s
and a destination
node N
d
is simply determined along the structure of the
cluster-tree. For example, the routing path from N
5
to N
2
is N
5
N
4
N
3
N
2
. Second, the packet transmission
times are scheduled in a cluster-based collision-free manner.
More specically, each cluster periodically has its dedicated
active duration called SD (Superframe Duration) which is
disjoint with other clusters SDs as shown in Figure 2(c).
Only within the clusters SD, all the nodes of the cluster
wake up and communicate each other. For this, the head of
each cluster sends a beacon frame at every beacon period
BI as shown in Figure 2(c). With the beacon frame, all
the nodes belong to the cluster synchronize. Each beacon
period BI
k
of a cluster k is composed of the clusters active
period SD
k
and its inactive period as shown in Figure 3. The
SD
k
contains a contention access period (CAP
k
), in which
nodes compete in a slotted CSMA/CA manner for non real-
time packets, and a contention-free period (CFP
k
), in which
nodes transmit their real-time packets with their dedicated
guaranteed time slots (GTS
k
s).
As such, IEEE 802.15.4/ZigBee standard provides base-
line features for routing and scheduling real-time packets.
However, for the proper operation of the ZigBee network,
we still have to nd the optimal conguration of all the
operating parameters for
cluster-tree construction
BIs, SDs for all the clusters and GTS allocations for
the member nodes and
power conguration of all the nodes such that the
desired cluster-tree can be formed.
In order to formally dene this holistic optimization
problem, we give necessary denitions and notations in the
following:
cluster-tree construction related denitions: A cluster tree
is represented by the parent-child relation
ij
for every pair
of two nodes N
i
and N
j
, which is dened as follows:

ij
=
_
1 if N
i
is the parent of N
j
0 otherwise
.
beacon frames
N1
N1
N2
N2
N3
N3
N4
N4
N5
N5
N6
N6
N7
N7
N8
N8
fa
fa
cluster3
cluster3
cluster7
cluster7
cluster4
cluster4
SD3
BI3
(a) physical placement of nodes
(b) logical cluster-tree view
(c) clusters SD scheduling
Figure 2. An example ZigBee network
Inactive period
CAP
k
Beacon frame
Superframe duration (SD
k
)
Beacon interval (BI
k
)
CFP
k
GTS
k
1
GTS
k
2
GTS
k
3
Figure 3. Beacon interval (BI) and superframe duration (SD) concepts
In the example of Figure 2(b),
31
and
45
are 1 while

13
,
25
and
52
are 0. With this
ij
notation, a node
N
i
and all its direct child nodes N
j
s for which
ij
= 1
form a cluster. For the cluster, N
i
is called the head while
all its child nodes N
j
s are called members. Thus, we use
the cluster heads id as the clusters id. For the example of
Figure 2(b), we have three clusters, that is, cluster
3
with
head N
3
, cluster
7
with head N
7
, and cluster
4
with head
N
4
.
duty-cycle scheduling related denitions: As mentioned in
Figure 2(c), each cluster
k
s activity window is characterized
by the superframe duration SD
k
and the beacon period
BI
k
. The length of BI
k
and SD
k
are determined by two
parameters, beacon order (BO) and superframe order (SO),
respectively, as follows:
BI
k
= aBaseSuperframeDuration 2
BO
k
SD
k
= aBaseSuperframeDuration 2
SO
k
_
for 0 SO
k
BO
k
14,
where aBaseSuperframeDuration denotes the minimum
length of the superframe, corresponding to SO
k
= 0.
This duration is xed to 960 symbols (a symbol cor-
responds to 4 bits) by the standard. This value corre-
sponds to the xed duration of 15.36 ms, assuming a 250
kbps in the 2.4 GHz frequency band. We call this xed
aBaseSuperframeDuration a slot. Thus, the problem of
determining the lengths of BI
k
and SD
k
for every cluster
k
is to determine the integer values of BO
k
and SO
k
in the
range from 0 to 14.
The head node and all the member nodes of cluster
k
are allowed to transmit their packets only for the duration
of SD
k
. The slots in a part of SD
k
, denoted by CAP
k
in
Figure 3, are accessed by the clusters nodes in a contention
based manner for non real-time packets. For the non real-
time packets, it is required to provide the minimum capacity,
denoted by MIN
nonRealTime
, that is,
CAP
k
BI
k
MIN
nonRealTime
,
where CAP
k
is an integer multiple of a slot. In this paper,
we consider only real-time packets and hence assume that
MIN
nonRealTime
= 0. However, our proposed optimization
framework works for any value of MIN
nonRealTime
.
For the nodes that deliver real-time data ows, we have
to allocate a GTS
k
s (guarantee time slots) as in Figure 3.
The length of a GTS allocated to a node N
i
in cluster
k
is
denoted by GTS
i
k
, which is an integer multiple of a slot.
In case that N
i
is the cluster
k
s head, that is, N
i
= N
k
,
its GTS, that is, GTS
k
k
is used for N
k
to transmit real-
time data to its member nodes N
j
for all j where
kj
=
1. The portion of GTS
k
k
used for transmitting data from
N
k
to N
j
is denoted by GTS
kj
k
. In other words, GTS
k
k
=

j,
kj
=1
GTS
kj
k
.
In case that N
i
is a member of cluster
k
, that is, N
i
=
N
k
, its GTS, that is, GTS
i
k
is used for N
i
to transmit real-
time data only to its head N
k
. That is, if we use the same
notational meaning GTS
ik
k
, that is, the portion used for N
i
to transmit real-time data to N
k
, we note GTS
i
k
= GTS
ik
k
.
nodes power related denitions: For the successful trans-
mission in between two nodes, their powers need to be
properly congured. For this, we control the transmission
power denoted by PW
i
of each node N
i
assuming that every
node in a receive mode uses the same power PW
recv
.
With these denitions, our problem is described as fol-
lows.
Problem Description: We are given a set of sensor nodes
V = {N
1
, N
2
, , N
m
} whose locations are pre-xed.
Thus, we know the geometric distance between every pair
of two nodes, N
i
and N
j
. The distance is denoted by d
ij
.
We are also given a set of n periodic real-time data
ows f
1
, f
2
, , f
n
, i.e.,
= {f
1
, f
2
, , f
n
}.
A periodic ow f
a
is characterized by a 5-tuple as follows.
f
a
= (src
a
, dst
a
, size
a
, p
a
, D
a
),
where src
a
, dst
a
, size
a
, p
a
, and D
a
are the source, desti-
nation, packet size (bits), period, and end-to-end deadline,
respectively.
For these given set V of nodes and set of real-time
ows, our problem is to nd all the following three domain
parameters:
cluster-tree construction:
ij
for every pair of N
i
and
N
j
,
duty-cycle scheduling: BI
k
and SD
k
for every
cluster
k
and GTS
i
k
for each node N
i
of cluster
k
, and
nodes power conguration: PW
i
for every node N
i
,
such that the network lifetime can be maximized while
guaranteeing the delivery of all the packets of each ow
f
a
within their end-to-end deadlines.
IV. PROPOSED HOLISTIC OPTIMIZATION FRAMEWORK
The objective of our optimization problem is to maximize
the lifetime of the network. The network lifetime is the time
until the most energy consuming node dies. As a measure
of the time, let us consider the long-run average power
avgPW(N
i
) of a node N
i
. With notations introduced in
Section III, avgPW(N
i
) can be formulated as follows:
avgPW(N
i
) = (1)
m

j=1

ij
_
PW
i
GTS
ij
i
+ PW
recv
GTS
ji
i
BI
i
_
+
m

j=1

ji
_
PW
i
GTS
ij
j
+ PW
recv
GTS
ji
j
BI
j
_
.
This formula consists of two parts. The rst part, i.e.,

m
j=1

ij
_
PWiGTS
ij
i
+PWrecvGTS
ji
i
BIi
_
represents the aver-
age power for the case when N
i
works as the cluster head
of cluster
i
. For this case, N
i
consumes the energy amount
of

m
j=1

ij
(PW
i
GTS
ij
i
+PW
recv
GTS
ji
i
) for sending
to and receiving from its child nodes N
j
at every its beacon
interval BI
i
. Thus, the average power is given as in the rst
part of Eq. (1).
On the other hand, the second part, i.e.,

m
j=1

ji
_
PWiGTS
ij
j
+PWrecvGTS
ji
j
BIj
_
represents the
average power for the case when N
i
works as a member
node of a cluster, say cluster
j
, that is, for the solely node N
j
for which
ji
= 1. For this case, N
i
consumes the energy
amount of

m
j=1

ji
(PW
i
GTS
ij
j
+PW
recv
GTS
ji
j
) for
send to and receive from its head node N
j
at every N
j
s
beacon interval BI
j
. Thus, the average power is given as
in the second part of Eq. (1).
With this formulation of avgPW(N
i
), the network life-
time is limited by the node that consumes the energy the
most, that is, for the node whose long-run average power
avgPW(N
i
) is the largest. Thus, our objective is to min-
imize the maximum avgPW(N
i
) for all i = 1, 2, , m,
that is,
Minimize max
1im
avgPW(N
i
). (2)
This Min-Max problem needs to be solved under a number
of constraints.
Valid tree construction: We use a set of
ij
s, 1 i m,
1 j m to represent a tree. For the set to represent a
valid tree,
ij
s need to have certain properties. First, every
node N
i
has at most one parent node. This property can be
formulated as the following constraint:
m

j=1

ji
1, i {1, 2, , m}. (3)
Second, if we add up the numbers of child nodes for all
the nodes, it should always be m1. This property can be
formulated as the following constraint:
m

i=1
m

j=1

ij
= m1. (4)
Third, we also have the following obvious constraint,
since N
i
is not the parent of itself.

ii
= 0, i {1, 2, , m}. (5)
Finally, if N
i
is the parent of N
j
, then the other way
around is not true. That is,

ij
+
ji
1, i, j {1, 2, , m}. (6)
Any set of
ij
s that satises the above constraints denes
one valid tree.
Valid power conguration: The transmission power of
each node N
i
needs to be congured such that its transmitted
packets can reach to its parent and all its child nodes.
Thus, N
i
s transmission power should be determined by the
maximum of the distances to its parent and child nodes, i.e.,
max
1jm
(
ij
+
ji
) d
ij
. Thus, we have the following
constraint:
RF(PW
i
) RF
min
recv
+ (7)
10 log
10
_
max
1jm
(
ij
+
ji
) d
ij
_
+ C,
i {1, 2, , m},
where RF(PW
i
) denotes the emitted RF signal strength
with the nodes transmission power PW
i
, and RF
min
recv
is
the minimum received RF strength that the signal must have
to achieve a certain bit rate, is the path loss exponent
whose value is normally in the range of 2 to 4, and C is
a constant which accounts for system losses. This is the
simplest form of the log-distance path loss model to ensure
successful transmission from N
i
to all its parent and child
nodes.
Valid duty-cycle scheduling: The given solution of
(BI
k
, SD
k
)s for all the clusters is valid only if a non-
overlapping SD schedule is possible. Fortunately, we can
use Koub aa et al.s algorithm [9] that checks if there exists
a non-overlapping SD schedule with the set of (BI
k
, SD
k
)s
of all the clusters. If any, Koub aa et al.s algorithm nd a
non-overlapping SD schedule for the given (BI
k
, SD
k
)s of
all the clusters.
In addition, SD
k
and GTS
j
k
are related as follows:
SD
k
CAP
k
+ GTS
k
k
+
m

j=1

kj
GTS
j
k
(8)
= BI
k
MIN
nonRealTime
+ GTS
k
k
+
m

j=1

kj
GTS
j
k
.
Also, there is a restriction on validity of the GTS allo-
cation by the IEEE 802.15.4 standard. Specically, IEEE
802.15.4 specication limits the number of GTSs within a
SD as 7. Thus, we have the following constraint:
(GTS
i
i
) +
m

j=1

ij
(GTS
j
i
) 7, i {1, 2, , m}, (9)
where (x) is one if x > 0 and zero otherwise.
End-to-end deadline guarantee: For the above valid so-
lution of cluster tree, i.e.,
ij
s, power conguration, i.e.,
PW
i
s, and duty-cycle scheduling, i.e., BI
k
, SD
k
, and
GTS
j
k
, we have to nally check if the solution can guar-
antee the end-to-end deadlines of all the real-time ows
= {f
1
, f
2
, , f
n
}. For this check, let us explain how
we can calculate the worst case end-to-end delay for each
ow f
a
from the source node src
a
to the destination node
dst
a
.
Thanks to the tree-based routing, once a cluster-tree
topology is given, we can determine the routing path for
each ow f
a
. In the example of Figure 2(b), the routing
path of ow f
a
from N
5
to N
2
is N
5
N
4
N
3
N
2
.
Using this routing path information, the worst case end-to-
end delay of f
a
can be computed by adding up the worst
case per-hop delays as in Tindell et al.s end-to-end response
time analysis [10].
Let us explain this calculation for the example ow f
a
in
Figure 2(b). The ow f
a
s packet arrivals and transmissions
N2
N3
N4
N5
cluster4
cluster3
cluster4s SDs
cluster3s SDs
GTS
54
4
GTS
43
4
GTS
32
3
J
3
a
w
32
a
(1)
w
32
a
(2)
Delay
32
a
(1)
Delay
32
a
(2)
Figure 4. Calculation of worst case end-to-end delay
at each hop from N
5
to N
2
are depicted in Figure 4. At the
source node N
5
, packets of f
a
are generated periodically as
marked by the down-arrows in the N
5
s timeline. However,
each packet is transmitted using GTS
54
4
as marked by a
small shaded box in the N
5
s time line. If there are more than
one ows that pass through the link from N
5
to N
4
, packets
of multiple ows need to be properly scheduled by N
5
.
For the packet scheduling, we use a non-preemptive xed
priority scheduling assuming that packets from each ow has
a pre-xed priority. The pending packets are sorted by their
priorities and the GTS is used for transmitting the pending
packets in the priority order. Due to the GTS waiting and the
packet scheduling from multiple ows, the packet arrivals
at the second hop are not periodic anymore as shown in
N
4
s timeline. Therefore, more than one packets, generally
q packets, from the same ow can be pending at the time
of GTS. Such situation of q packets from the same ow is
well studied by Tindells analysis at each hop to calculate the
worst case end-to-end delay. Leveraging Tindells approach,
the worst case end-to-end delay, denoted by WCD
a
for ow
f
a
, can be computed by solving the following equations
at each hop of the end-to-end path for f
a
. The following
equations are for the case of the hop from N
i
to N
j
where
N
i
is the cluster head. The other case where N
j
is the cluster
head can be similarly computed and we skip it.
w
ij
a
(q) = (10)
_

q C
a
+ B
ij
a
+

bhp
ij
a
_
J
i
b
+w
ij
a
(q)
p
b
_
C
b
GTS
ij
i
_

BI
i
.
WCD
ij
a
= max
q=1,2,
J
i
a
+ w
ij
a
(q) (q 1) p
a
, (11)
where
hp
ij
a
is the set of higher priority ows than f
a
that pass
through the hop from N
i
to N
j
. This set can be easily
found thanks to the tree-based routing of all the ows.
C
a
is the transmission time of a packet of f
a
. This is
simply given as the packet size size
a
divided by the
transmission rate R, that is, C
a
= size
a
/R.
B
ij
a
is the blocking time due to the non-preemptive trans-
mission of a packet of a lower priority ow. It is given
as the largest transmission time out of all the packets
of lower priority ows that pass through the hop from
N
i
to N
j
.
J
i
a
is the worst case arriving jitter of f
a
due to scheduling
delays at the hops before N
i
. It is given as the worst
case delay of f
a
until reaching N
i
.
In Eq. (10), the intuitive meaning of w
ij
a
(q) is the length
of worse case time from the arrival of the rst packet of f
a
at N
i
until completely transmitting rst q packets of f
a
to
N
j
. Figure 4 shows examples of w
32
a
(1) and w
32
a
(2) at the
hop from N
3
to N
2
. In general, w
ij
a
(q) can be calculated
by adding (1) the total time for transmitting q packets, (2)
blocking time B
ij
a
due to the already started transmission
of a lower priority packet, and (3) largest possible delay
due to higher priority packets that arrive before the com-
plete transmission of q packets of f
a
, which is given as

bhp
ij
a
_
J
i
b
+w
ij
a
(q)
p
b
_
C
b
. In this term,
_
J
i
b
+w
ij
a
(q)
p
b
_
is the
largest possible number of packets of ow f
b
during the
time window w
ij
a
(q) assuming the worst case scenario that
the rst packet of f
b
is delayed the most, i.e., J
i
b
, and then
the succeeding packets arrive with the maximum rate, i.e.,
1/p
b
and thus the packet arrivals from f
b
are most packed
with the time window w
ij
a
(q). Note that w
ij
a
(q) appears in
both sides making Eq. (10) recursive. This recursive equation
can be solved iteratively starting with the initial assumption
of w
ij
a
(q) =
_
qCa+B
ij
a
GTS
ij
i
_
BI
i
until w
ij
a
(q) does not further
increase by additional packets of higher priority ows.
In Eq. (11), if we add the worst case delay of f
a
until
reaching N
i
, i.e., J
i
a
, to w
ij
a
(q), the result J
i
a
+ w
ij
a
(q)
becomes the worst case time between the rst packet gener-
ation time at the source node and the complete transmission
time of the rst q packets at the hop from N
i
to N
j
.
Figure 4 shows J
3
a
+w
32
a
(1) and J
3
a
+w
32
a
(2) for the example
hop from N
3
to N
2
. Thus, q-th packets delay denoted by
Delay
ij
a
(q) until reaching N
j
is J
i
a
+ w
ij
a
(q) (q 1) p
a
as shown in Figure 4. By picking the largest value of
J
i
a
+w
ij
a
(q)(q 1) p
a
out of all possible q = 1, 2, , the
worst case delay WCD
ij
a
until reaching N
j
can be found
as in Eq. (11). Fortunately, Tindell et al. proved that such
largest value exists out of q = 1, 2, , Q where Q is the
rst integer that satises w
ij
a
(Q) Q p
a
[10]. Therefore,
it is sufcient to check only a nite number of q.
By repeatedly using the above equations from the source
node to the destination node for each real-time ow f
a
, we
can compute the worst case end-to-end delay WCD
a
. If the
computed end-to-end delay WCD
a
is less than or equal to
the end-to-end deadline D
a
, we conclude that all the packets
of f
a
can be delivered to the nal destination before their
deadlines.
By repeating this check for all the real-time ows, we can
verify whether the given solution of cluster tree, i.e.,
ij
s,
power conguration, i.e., PW
i
s, and duty-cycle scheduling,
i.e., BI
k
, SD
k
, and GTS
j
k
, can deterministically guarantee
the end-to-end deadlines of all the real-time ows.
So far, we have formulated our problem as a formal op-
timization problem. However, it is a complex combinatorial
problem to nd the holistic solution of
ij
s, PW
i
s, BI
k
s,
SD
k
s, and GTS
j
k
s that maximizes the network lifetime, i.e.,
Eq. (2) while satisfying all above constraints, that is, (1)
valid tree construction constraints, (2) valid power congura-
tion constraints, (3) valid duty-cycle scheduling constraints,
and (4) end-to-end deadline guarantee constraints. Thus, it is
computationally intractable to exhaustively search the entire
solution space.
In order to manage such a high complexity of our holistic
optimization problem, we use a genetic algorithm. For the
genetic algorithm, we need (1) a chromosome representa-
tion and (2) a tness function. First, a chromosome string
represents one possible solution of our holistic optimization
problem. Thus, a chromosome string in our genetic algo-
rithm is a complete set of all the parameters of cluster-tree
construction, nodes power conguration, and duty-cycle
scheduling as shown in Figure 5. Second, the tness function
is used to evaluate the quality of each solution, i.e., tness
of each chromosome string. As the tness function, we use
our objective function, i.e., long-run average power function
in Eq. (1).

11

mm
PW
1
PW
m
. . . . . . . . . (BI
1
, GTS
1
1
, , GTS
1
m
)
Tree construction
(BI
m
, GTS
m
1
, , GTS
m
m
)
Power configuration Duty-cycle scheduling
Figure 5. Chromosome structure of the cluster-tree network
With the chromosome representation and the tness func-
tion, the genetic algorithm rst forms the initial population
of chromosome strings, which are seeded randomly in order
to cover broad points of the entire solution space. Then,
the genetic algorithm improves the chromosome strings
in the initial population from generation to generation by
repeating the following reproduction, mutation, repair, and
replacement steps:
Reproduction: This step selects two chromosome
strings as parents and combines them to create a new
chromosome which typically shares many characteris-
tics of its parents. In our genetic algorithm, the standard
weighted roulette wheel method [11] is used to select
two chromosome strings with better tness with higher
probability.
Mutation: This mutation step is taken probabilistically
and once taken it gives a random perturbation to the
reproduced chromosome string, which is necessary to
avoid getting stuck at a local optima. The random
perturbation is made by randomly choosing one value
out of
ij
s, PW
i
s, BI
k
s, SD
k
s, and GTS
j
k
and then
modifying it with a randomly generated value.
Repair: The reproduced and mutated chromosome
strings are mostly infeasible solutions for our problem.
Therefore, the repair step is necessary after the repro-
duction and mutation steps. For an efcient repair of
the infeasible chromosome string, we employ a greedy
repair method as follows:
valid tree construction: If the
ij
s in the chromo-
some string make cycles in the resulting graph, we
gradually remove edges from the longest one in
the cycles until the graph becomes acyclic. If the

ij
s result in a disconnected graph, we gradually
add new edges from the pair of two closest nodes
that do not make cycle until the graph becomes
connected.
valid power conguration: The transmission power
PW
i
of each node N
i
is always repaired such that
its signal can reach to its farthest neighbor, based
on the log-distance pass loss model in Eq. 7.
valid duty-cycle scheduling: With the BI
k
s, SD
k
s
in the chromosome string, if the duty-cycle
scheduling is not feasible, we try increasing BI
k
s
or decreasing SD
k
s so that the total sum of the
duty cycle, i.e.,

SD
k
/BI
k
, can be reduced,
which increases the chance of schedulability.
end-to-end deadline guarantee: If the worst case
end-to-end delay of any ow is greater than its end-
to-end deadline, we decrease BI
k
s along the ows
path while decreasing SD
k
and GTS
j
k
accordingly
such that the same ratio of SD
k
/BI
k
can be
maintained. This way, we can reduce the worst
case end-to-end delay of the ow while keeping
the same chance of duty-cycle schedulability.
Replacement: After the new chromosome string is
repaired as a feasible one, we replace a chromosome
string in the population with the new one so that the
population size can be kept the same. In order to
select the chromosome string to be replaced, we employ
an elitism strategy as in [11] because it ensures that
the best chromosome string in the current generation
always survives into the succeeding generation.
This repetitive evolution process of the population terminates
either at a sufcient number of generations, 2000 in our case,
or at the condition that a specied percentage, 70% in our
case, of the chromosome strings have the same best tness.
V. EXPERIMENTS
This section investigates how much improvement our
holistic optimization can make in terms of the long-
run average power of the most bottleneck node, i.e.,
max
1im
avgPW(N
i
) as dened in Equation (1). For this
investigation, we consider 10 ZigBee nodes, i.e., m = 10,
randomly placed in a 500m 500m rectangular area.
For each sensor node, we assume the TI CC2520 RF
transceiver [12], for which the transmission rate R is 250
kbps, the packet reception power PW
recv
is 55.5 mW, the
range of controllable transmission power PW
i
is (48.6 mW,
100.8 mW). Regarding the log-distance path loss model
in Equation (7), the CC2520 RF transceivers minimum
received RF strength RF
min
recv
is -85 dBm. For the path loss
exponent and the system loss constant C, we assume 3.5
and 0, respectively, as in [13].
On top of these sensor nodes, we randomly generate
periodic real-time ows. More specically, each ow f
a
s
packet size size
a
is randomly generated following uniform
distribution in the range of (1500 bits, 2500 bits). The
period p
a
and the end-to-end deadline D
a
are also randomly
generated following the uniform distribution in the ranges of
(4000 ms, 6000 ms) and (1000 ms, 3000 ms), respectively.
The ow f
a
s source node src
a
and destination node dst
a
are also randomly pick out of the above 10 nodes.
With these settings, we compare the following three
approaches:
Hans approach [1] that optimizes only the duty-cycle
scheduling assuming that the cluster-tree and nodes
powers are given. For the cluster-tree, we assume the
minimum spanning tree construction in [14] assuming
the maximum transmission power for each node.
Our approach called optimal clustering and max
power that optimizes both duty-cycle scheduling and
cluster tree construction while using the maximum
power for all the nodes.
Our approach called optimal clustering and optimal
power that optimizes all of duty-cycle scheduling,
cluster tree construction, and nodes power congura-
tion.
Figure 6 compares the three approaches in terms of the
long-run average power of the most bottleneck node as
varying the number of real-time ows, i.e., n from 1 to
6. Each point of Figure 6 is the average for 100 problems
with randomly generated 10 nodes and n ows as mentioned
above. As the number ows increases, that is, as the overall
system workload increases, the long-run average power
also increases for all the three approaches. However, the
slopes of the increase are signicantly different. For Hans
approach, the power consumption of the most bottleneck
node increases fast, since it does not distribute the ows
by leveraging the freedom of tree clustering. On the other
hand, our approach optimal clustering and max power
can optimally leverage the freedom of three clustering and
hence well distribute the ows. Thus, the load given to
the most bottleneck node can be limited even when the
number of ows gets larger. Due to this reason, the gap
0
5
10
15
20
25
30
1 2 3 4 5 6
l
o
n
g
-
r
u
n

a
v
e
r
a
g
e

p
o
w
e
r

(
m
W
)
# of flows
Hans approach
optimal clustering and max power
optimal clustering and optimal power
Figure 6. Long-run average power of the bottleneck node as a function
of the number of ows
0
5
10
15
20
25
30
500 - 1000
1000 - 1500
1500 - 2000
2000 - 2500
2500 - 3000
3000 - 3500
3500 - 4000
l
o
n
g
-
r
u
n

a
v
e
r
a
g
e

p
o
w
e
r

(
m
W
)
deadline (ms)
Hans approach
optimal clustering and max power
optimal clustering and optimal power
Figure 7. Long-run average power of the bottleneck node as a function
of the average deadline
between Hans approach and our approaches becomes larger
as the number of ows increases. In addition, our approach
optimal clustering and optimal power can further save the
power by using just adequate power rather than the max
power.
In order to investigate the effect of ow deadlines,
Figure 7 compares the three approaches by varying the
deadline generation range with the xed number of ows,
i.e., 4. The results are again the averages for 100 random
problems. When the deadlines are tight, in the case of Hans
approach, the long-run average power of the most bottleneck
node is pretty high. This can be explained as follows. In
Hans approach, there is no freedom of shortening routing
paths because of already xed cluster-tree. Thus, it should
meet the tight deadlines by optimizing only the duty-cycle
scheduling resulting in small BI values (see Eq. (10)). This
makes a large long-run average power especially in the most
bottleneck node. On the other hand, our approaches optimal
clustering and max power and optimal clustering and opti-
0
1
2
3
4
0 - 100 100 - 200 200 - 300 300 - 400
l
o
n
g
-
r
u
n

a
v
e
r
a
g
e

p
o
w
e
r

(
m
W
)
farthest distance (m)
optimal clustering and max power
optimal clustering and optimal power
Figure 8. Long-run average power of the bottleneck node as a function
of the farthest distance
mal power can enjoy the freedom of shortening the routing
path via optimal tree-clustering and hence our optimized
BI values can be larger even when the deadlines are tight.
Thus, our approaches result in a signicantly smaller long-
run average power. As the deadlines become loose, the duty-
cycle optimization alone in Hans approach can enjoy larger
room of the spare time until the loose deadline resulting on
larger BI values. Thus, Hans approach shows the decreasing
long-run average power slightly reducing the gap with our
approaches. However, even when the deadlines sufciently
long, we still can observe signicant gap between Hans
approach and ours.
The saving by our optimal power conguration largely
depends on the farthest distance between a parent and its
child among all parent-child pairs in the tree cluster network.
In other words, if the farthest distance is small, the parent
and child can communication with the power much smaller
than the max power. However, if the farthest distance is
large, they need to use the max power anyway to reach
each other, and hence the optimal power should set to the
same as the max power. In order to show this, Figure 8
compares our optimal clustering and max power and our
optimal clustering and optimal power for 500 random
problems with xed number of ows, i.e., 4. In each bin
of x-axis, i.e., (0-100), (100-200), (200-300), and (300-
400), we put the solutions with the farthest distance in the
corresponding range and average their tness, i.e., long-run
average power of the most bottleneck node. In the case of
optimal clustering and max power, the long-run average
power is almost constant regardless of the farthest distance.
This is because all the nodes use the maximum transmission
power without any power optimization. On the other hand,
in the case of optimal clustering and optimal power, the
long-run average power is much smaller when the farthest
distance is small. This clearly shows that the optimal power
conguration can make non-trivial power saving depending
on the given problem setting of nodes and ows, especially
when our solution can address the given problem with a
small farthest distance.
VI. CONCLUSION
In this paper, we propose a holistic optimization method
that builds the optimal IEEE 802.15.4/ZigBee network with
on-time delivery of all real-time data while maximizing
the network lifetime. The holistic optimization method
jointly addresses the three cross-related problems: (1) logical
cluster-tree construction, (2) power conguration for nodes,
and (3) duty-cycle scheduling.
We rst formulate the holistic optimization problem as a
formal optimization problem. Thanks to this formal problem
formulation, we can solve the holistic optimization problem
with a genetic algorithm. Our experimental study shows that
beyond the improvement by the optimal duty-cycle schedul-
ing which has been done by previous work, the optimal
tree clustering and optimal nodes power conguration can
make a further signicant improvement in the lifetime of
real-time IEEE 802.15.4/ZigBee network. This justies our
holistic optimization framework that optimizes all of the
duty-cycle scheduling, cluster tree construction, and nodes
power conguration all together.
Currently, we are applying the proposed holistic optimiza-
tion to build a ZigBee network for real-time monitoring of
a large scale building. In the future, we plan to extend the
optimization framework targeting coexistence of real-time
ows with deterministic and stochastic deadline guarantee
requirements.
REFERENCES
[1] J. Han, Global Optimization of ZigBee Parameters for End-
to-End Deadline Guarantee of Real-Time Data, IEEE Sen-
sors Journal, vol. 9, no. 5, pp. 512514, 2009.
[2] S. C. Ergen and P. Varaiya, Energy Efcient Routing with
Delay Guarantee for Sensor Networks, ACM Wireless Net-
works, vol. 13, no. 5, pp. 679690, 2007.
[3] M. Caccamo, L. Y. Zhang, L. Sha, and G. Buttazzo, An
Implicit Prioritized Access Protocol for Wireless Sensor Net-
works, in Proceedings of the 23rd IEEE Real-Time Systems
Symposium (RTSS), 2002.
[4] T. Watteyne, I. Aug` e-Blum, and S. Ub` eda, Dual-Mode
Real-Time MAC Protocol for Wireless Sensor Networks: A
Validation/Simulation Approach, in Proceedings of the 1st
International Conference on Integrated Internet Ad Hoc and
Sensor Networks (InterSense), 2006.
[5] O. Chipara, Z. He, G. Xing, Q. Chen, X. Wang et al.,
Real-Time Power-Aware Routing in Sensor Networks, in
Proceedings of the 14th IEEE International Workshop on
Quality of Service (IWQoS), 2006.
[6] T. He, J. A. Stankovic, C. Lu, and T. Abdelzaher, SPEED:
A Stateless Protocol for Real-Time Communication in Sensor
Network, in Proceedings of the 23rd International Confer-
ence on Distributed Computing Systems (ICDCS), 2003.
[7] E. Felemban, C.-G. Lee, E. Ekici et al., Probabilistic QoS
Guarantee in Reliability and Timeliness Domains in Wireless
Sensor Networks, in Proceedings of the 24th Annual Joint
Conference of the IEEE Computer and Communications So-
cieties (INFOCOM), 2005.
[8] ZigBee Alliance, ZigBee Specication,
http://www.zigbee.org, 2005.
[9] A. Koub aa, A. Cunha, and M. Alves, A Time Division
Beacon Scheduling Mechanism for IEEE 802.15.4/Zigbee
Cluster-Tree Wireless Sensor Netowrks, in Proceedings
of the 19th Euromicro Conference on Real-Time Systems
(ECRTS), 2007.
[10] K. Tindell and J. Clark, Holistic Schedulability Analysis
for Distributed Hard Real-Time Systems, Microprocessing
& Microprogramming (Euromicro J.), vol. 40, no. 2-3, pp.
117134, 1994.
[11] D. E. Goldberg, Genetic Algorithms in Search, Optimization,
and Machine Learning. Addison-Wesley, 1989.
[12] Texas Instruments, CC2520 Datasheet 2.4
GHz IEEE 802.15.4/ZigBee RF Transceiver,
http://focus.ti.com/docs/prod/folders/print/cc2520.html.
[13] V. S. Abhayawardhana et al., Comparison of Empirical
Propagation Path Loss Models for Fixed Wireless Access Sys-
tems, in Proceedings of the 61st IEEE Vehicular Technology
Conference (VTC), 2005.
[14] R. C. Prim, Shortest Connection Networks and Some Gen-
eralizations, Bell Sys. Tech. J., pp. 13891401, 1957.