CFMS: A Cluster-based Convergecast Framework for Dense Multi-Sink Wireless Sensor Networks

International Journal of Wireless & Mobile Networks (IJWMN), Vol.15, No.1, February 2023
DOI:10.5121/ijwmn.2023.15101 1
CFMS: A CLUSTER-BASED CONVERGECAST
FRAMEWORK FOR DENSE MULTI-SINK WIRELESS
SENSOR NETWORKS
Gokou Hervé Fabrice Diédié1
, Koigny Fabrice Kouassi2
and Tchimou N’Takpé2
1
Laboratory of Mathematics and Computer Science, Université Peleforo Gon Coulibaly,
Korhogo, Ivory Coast
2
Laboratory of Mathematics and Computer Science, Université Nangui Abrogoua,
Abidjan, Ivory Coast
ABSTRACT
Convergecast is one of the most challenging tasks in Wireless Sensor Networks (WSNs). Indeed, this data
collection process must be conducted while copying with packet collisions, nodes’ congestion or data
redundancy. These issues always result in energy waste which is detrimental to network efficiency and
lifetime. This paper is aimed to address these problems in large-scale multi-sink WSNs. Inspired by our
previous work MSCP, we designed a lightweight protocol stack that seamlessly combines clustering, path-
vector routing, sinks’ duty cycling, data aggregation and transmission scheduling in order to minimise
message overhead and packet losses. Simulation results show that this solution can mitigate delay while
significantly increasing packet delivery and network lifetime.
KEYWORDS
Wireless Sensor Networks, Multi-sink, Clustering, Convergecast, Aggregation tree, Scheduling
1. INTRODUCTION
Wireless Sensor Networks (WSNs) are composed of small resource-constrained monitoring
devices also called sensor nodes and a sink (i.e. gateway). The latter is in charge of connecting
the resulting infrastructure to the outside world. Nowadays, WSNs are pervasive and used for
various applications in domains like ecology, health, security, transportation, research or industry
[1] - [3].
In such networks, sensor nodes detect physical events and transmit the derived data to the
gateway through a multi-hop communication pattern [4]. This process (commonly referred to as
convergecast), implies routing a large flow of data to a single point [5][6]. Interference and
congestion are inherent to this kind of traffic and lead to packet losses [7]. They always provoke
delays and energy waste, hence reduction of network efficiency and lifetime [8].
A common solution especially in dense WSNs, involve increasing the numbers of sinks,
clustering the sensor nodes, reducing redundancy, controlling collision or congestion via data
aggregation and transmission scheduling [9][10] [11]. Unfortunately, most of such contributions
found in the literature are not scalable or make use of assumptions, metrics and strategies that are
actually too restrictive for any real-world implementation.

2
In this paper, we propose to combine clustering, path-vector routing and scheduling schemes in
order to dynamically assign a proper sink to each sensor node, duty cycle nodes activity,
construct a collision-free data aggregation tree, while balancing network load. The resulting
framework helps collect periodical data, mitigate delay, minimize packet losses and prolong
network lifetime.
The main contributions of this work are summarized as follows:
- a hop limitation clustering technique that mitigates message overhead and network flooding;
- a route discovery processes that simultaneously consider links’ asymmetry, interference, and
nodes’ congestion ;
- a fully-distributed loop-free path-vector routing strategy that minimizes delays between sensor
nodes and sinks;
- a duty-cycling strategy that helps balance loads for both sensor nodes and sinks;
- a tree construction and scheduling scheme for collision-free data aggregation.
The rest of the paper is organized as follows: Section 2 surveys the related contributions; then,
the proposed solution is detailed in Section 3; the performance evaluation process, the results,
and discussions are presented in Sections 4 and 5 followed by conclusion in Section 6.
2. RELATED WORK
In WSNs, data aggregation and traffic scheduling-aware convergecast schemes are generally
classified into tree-based solutions and cluster-based ones [12] - [15]. These works as well can be
categorized according to the routing, the aggregation or the traffic scheduling techniques they
use. However, regardless the strategy used only cluster-based are actually scalable [16]. This
paper is aimed to focus on such solutions.
LEACH by Heinzelman et al. [17] is one of the early solutions proposed in this category. It has
inspired many contributions for the past two decades [18] [19]. Regrettably, most of them are
dedicated to only one sink [20]. In this IoT era, many of these single-sink oriented solutions are
still proposed despite their low applicability [21] - [23].
Yu and Li [24] proposed the first hierarchical solution for the minimum-time collision-free
aggregation scheduling problem in a multi-sink environment. A Voronoï-based scheduling
strategy is suggested to construct a dominating set for all the sensor nodes in forests (i.e. clusters)
of trees around each sink. Schedules are then made as well as for dominatee and dominator
nodes. Unfortunately, Voronoï partition requires nodes to know their exact positions.
Goyal et al. [25] In this paper, an improved data aggregation technique for cluster-based UWSN
(Under Water Sensor Network) is proposed where an efficient sleep-wake up algorithm is used
for aggregating the sensed data and TDMA based transmission schedule is used to avoid intra and
inter cluster collisions to reduce the data redundancy and ensure energy efficient collision-free
transmission.
Khan et al. [26] suggested a typical multi-layer cluster-based strategy for UWSNs. Nodes are
grouped in equal size layers according to their depth. CHs are elected based on their residual
energy and distance to sinks. After cluster formation CHs assign a TDMA schedule to other
members to help them send data which is aggregated by CHs. This scheme is based on one-hop
clusters which could foster interference between CHs.

3
Rajput and Kumaravelu [27] used a Fuzzy-c-means-based strategy to balance the size of clusters
while optimizing the number of sinks and their locations in the deployment area. Regrettably, this
solution is limited to applications with a deterministic sink deployment and is not scalable; since
the latter deployment and the clustering process are centralized in the distant base station.
Singh and Nagaraju [28] proposed a Sink-As-CH strategy [29]-[31] to help reduce number of
intermediary hops. Sensor nodes are assigned to clusters according to their Euclidean distance to
the sink. Packets are then routed by constructing a Wiener minimum spanning tree based on an
ABC (Artificial Bee Colony) optimization scheme. Opportunistic coding is finally used at each
node to select the best flows passing through intermediary nodes and to aggregate data of
intersecting flows. Regrettably, packet transmission is implicitly scheduled via a queuing process.
Daas et al. [32] designed a distributed multi-hop cluster-based routing protocol where paths are
selected using both hop count and link state via Signal-to-Noise-Ratio. This protocol also
leverages to help balance the load of both sensor nodes and sinks. However, this solution requires
fixed size clusters thus can be hardly applied to randomly deployed networks, hence to monitor
phenomena in harsh environments.
Verma et al. [33] proposed a bio-inspired optimisation scheme heterogeneous WSNs. A genetic
algorithm-based clustering strategy is used to select CHs via a metric combining residual energy,
distance to the sink, and node density. To help mitigate fast energy depletion in sinks’ one-hop
neighbourhood, they suggest deploying sinks outside the region of interest in order to directly
connect CHs to sinks. Unfortunately, this strategy is too restrictive for real-life applications
which requiring random deployment.
Recently, another bio-inspired solution was proposed by Chada and Gugulothu [34]. They
suggested an optimisation-based scheme using the multi-objective black widow metaheuristics.
The latter mimics the cannibalism behaviour of spider black widows. Candidate CHs referred to
as spiders are chosen leveraging their residual energy and distance to the sinks via a pheromone-
based strategy. Data can be aggregated by CHs before transmission. However, no explicit
scheduling scheme is proposed to cope with possible collision.
3. PROPOSED SOLUTION
This section is aimed to present our solution named CFMS (Convergecast Framework for Multi-
Sink networks). We discuss our motivations, objectives and assumptions, before detailing the
latter solution.
3.1. Motivations and Objectives
As shown in previous section, few state-of-the-art cluster-based multi-sink convergecast
protocols have been proposed so far. In our previous work, named MSCP [35] we proposed a
proactive loop-free scheme to increase network efficiency and lifetime, besides scalability, by
considering interferences and link asymmetry, while duty cycling activity of both sensor nodes
and sinks. Unfortunately, MSCP struggles to cope with data redundancy, collisions and node
congestion. Furthermore, by allowing each node to individually send its data, MSCP is actually
more suitable for event-driven applications. This work is aimed to address these issues while
providing a good trade-off between network efficiency and lifetime to periodic (time-driven)
application.

4
3.2. Assumptions
We assume that:
- nodes are equipped with an omni-directional radio;
- each node has a unique identifier (ID);
- nodes are uniformly and randomly deployed in the area of interest;
- nodes’ connection is modelled as an UDG (Unit Disk Graph);
- each node can assess distances through the received signal strength or a specific localization
protocol;
- the number of sensor nodes is higher than the number of sinks;
- each node has a finite size buffer using a FIFO policy;
- nodes are static.
Figure.1 Process stack of CFMS.
3.3. Description
As shown in Figure 1, CFMS is composed of three processes, namely the clustering process, the
path-finding process, and the data aggregation-scheduling process. The first two processes are
respectively inspired from MSCP presented in our previous work [35].
We detail in the following sub-sections these processes.
3.3.1. Clustering Phase
This is a two-step cyclic phase starting with neighbour discovery followed if necessary by the
creation of a new cluster. Each node u must give a unique integer number to each of its
neighbours. This number is randomly chosen in interval 1; ( )
N u
 
  ; where ( )
N u denotes the set
of these neighbours.
We use a scheme similar to the one proposed for the clustering process inside the CONSTRUCT
protocol [31]. Note that each cluster is identified by its Cluster Head ID (CHID); hence, each
node knows the cluster it belongs to. Besides, we use a Sink-As-CH strategy [29], i.e. each sink
creates a cluster in its k-hop neighbourhood; k is given as a parameter such as k =
max_hop_count.

5
S
Z
V
Y
R
W
U
T
X
I
O
Q
L
D
N
P
K
M
J
B
A
G
F
C
H
E
Leaf-member
Cluster-Head
(CH)
CH-ring-neighbour
Sink
Sink-ring-neighbour
CH Z's neighbouring
cluster gateway
Sink A's 2-hop range
gateway
Simple member
CH Z's cluster
gateway 2
CH Z's cluster
gateway 1
Figure.2 Nodes’ statuses after the clustering phase
The duration of CHs’ service time is fixed as a parameter. Therefore, after its mandate a CH must
abandon its status and launch a new election process in its k-hop neighbourhood. Moreover, when
two CHs move next to each other, the one with less cluster mates will eventually lose its role and
become a new cluster mate; ties are randomly broken Figure 2 depicts the different statuses
applicable to each node in a section of the network after the cluster formation phase. Equation (1)
helps calculate the CH’s fitness score; where , , , , ( )
n Ei Er CC S t respectively denote the number
of its neighbours, its residual energy, its initial energy, its clustering coefficient and its stability
index (see [31] for details).
 
1
1 ( )
score
n Er
CC S t
n Ei
   
 

   
      
   
   
(1)
3.3.2. Intra-Cluster Convergecast Sub-Tree Construction
After the creation of a new cluster, the CH must construct the intra-cluster data forwarding sub-
tree and interconnect it to its inter-cluster counterpart. To do so, CH u broadcasts in its k-hop
neighbourhood a TREE-REQ message containing its ID then triggers a timer (TREE-timer) and
waits for any response during _ _ ( )
wait
max hop count t u
 seconds. This duration is calculated using
Equation (2); where rtt(.) denotes the round-trip-time (in seconds) experienced with node v
during neighbour discovery. N(u) is node u’s neighbourhood.
 
( ) max rtt( ); ( )
wait
t u v v N u
  (2)
When receiving a TREE-REQ message from a neighbour u , a node v must increment the value
of the hop-count field to 1 then estimates the transmission delay ,
u v
 using Equation (3); where
,
u v
C , ,
u v
L and u
S respectively denote the capacity (i.e. maximum transmission rate) of link ( , )
u v ,
the received packet length on this link and the queuing delay (sojourn time) of this message
inside neighbour u .
Note that u
S is provided by the latter neighbour via a dedicated field in the message.
,
,
,
u v
u v u
u v
L
S
C
   (3)

6
, , 2 ,
log (1 )
u v u v u v
C W SINR
   (4)
,
u v
C is estimated using Equation (4); where ,
u v
W and ,
u v
SINR respectively denote the bandwidth of
link ( , )
u v and the Signal Interference plus Noise Ratio experienced on this link.
Node v then increments the total-delay ield with the estimated delayf ,
u v
 . Node v must also
insert its ID into the first-forwarder field if the latter node is a neighbour of the TREE-REQ
message’s sender.
A node must respond to a TREE-REQ message by broadcasting a TREE-ACK message if this
node is a neighbouring gateway (i.e. is affiliated to another CH but has at least one neighbour
inside the sender’s cluster), is a leaf-mate (i.e. a cluster member with only one neighbour) or is a
CH-ring-neighbour (i.e. has exactly two neighbours of which one is the latter CH) see Figure 2
for illustration.
TREE-ACK messages contain the sender’s ID, its CH’s ID, the value of the first-forwarded field
provided by the received TREE-REQ message, a list of IDs of its neighbouring sinks, the delay to
each of these sinks, and the estimated total delay of the TREE-REQ message.
Figure.3 Structure of convergecast table
Note that TREE-REQ messages are forwarded by a node if the sender is a cluster mate (CHID
=ID), this message is received for the first time, and its hop_count field’s value is below
max_hop_count +1. Moreover, to mitigate the protocol overhead, TREE-REQ messages are
never sent back to a forwarder.
When forwarding a TREE-REQ message to its neighbours, a node sends the list of pairs (unique
number, neighbour’s ID). The recipient neighbour then concatenates the given unique number to
the content of this message’s cast_vector field.
TREE-ACK messages are forwarded following the same rules applied to TREE-REQ except that
the CH’s ID is not compared to the forwarder ID.
After receiving a TREE-ACK message sent by a neighbouring cluster gateway, the CH node u
must update its convergecast table (as shown in Figure 3) by calculating the delay to each
destination (i.e. sink) via Equation (5); where ,
u x
 , ,
u s
 and ,
s x
 respectively denote delays from
CHu to sink x , from u to the sender s and from the sender s to sink x .The first forwarder
provided by the TREE-REQ message is the next-hop. For the same destination only the lowest
delay so far experienced is kept in this table.
, , ,
u x u s s x
  
  (5)

7
TREE-ACK messages sent by other eligible nodes help CH to determine total delay on the
senders-to-CH paths.
After TREE-timer expiration, the CH must reply to each received TREE-ACK message, via the
first forwarder node, by sending a CAST message of which field type is respectively set to 1 if
the receiver (i.e. sender of the TREE-ACK message) is a neighbouring cluster gateway and 0
otherwise.
Each CAST message also contains the list of sinks discovered by the CH and the delays to reach
them. Before sending this message, CH copies into the cast_vector field, content of the
corresponding field of the received TREE-ACK message.
When receiving a CAST message with type field equals to 1, a node u must update its
convergecast table by calculating its transmission delay ,
u x
 to each sink x as expressed by
Equation (5); where ,
u x
 , ,
s x
 and ,
s u
 respectively denote delays from node u to sink x , from the
sender s to sink x and from the sender s to node u .
The latter must analyse content of the cast_vector field, then extract the unique number of which
position equals value of the hop_count field. Neighbour that was assigned this unique number is
chosen as the next hop.
, , ,
u x s x s u
  
  (6)
However, when a leaf-member or a CH-ring-neighbour (see Figure 2 for illustration) receives a
CAST message, it must respond by sending a CAST-ACK. Except that a CH-ring-neighbour
must send a CAST-ACK only after receiving a CAST message forwarded by its non-CH
neighbour.
Note that CAST and CAST-ACK messages are forwarded following the same rules applied to
TREE-REQ and TREE-ACK ones.
A typical CAST-ACK message contains the list of sinks and the delays to reach each of them.
The cast_vector field embeds the forwarders’ unique number. All these information were
extracted from the CAST message previously sent by the CH.
After receiving a CAST-ACK message a node u must use Equation (5) to update its
convergecast table by calculating the delay ,
u x
 to reach each sink x via the CH s ; then must
extract from the cast_vector field, the unique number of which position equals value of the
hop_count field; so as to choose as next hop, the neighbour that was assigned the latter unique
number.
3.3.3. Inter-Cluster Convergecast Sub-Tree Construction
Note that only sinks are allowed to periodically (i.e. at a beginning of a new duty-cycle)
broadcast a HELLO message to clusters located at most max_cluster_hop_count hops.
After sending such a HELLO message a sink u must trigger a timer (HELLO-timer) and wait for
any response from next-hop gateways during _ _ _ ( )
wait
max cluster hop count t u
 seconds (see
Equation (2)). HELLO messages are forwarded like TREE-ACK messages.
Once a node receives a new HELLO message, if the forwarder’s CHID is different from its own
CHID, the latter node must increment the cluster_hop_count field’s value to 1 and forward this

8
message only if the value of this field is lower than max_cluster_hop_count. This node must also
insert its ID into the first-forwarder field if it is a neighbour of the message’s sender (i.e. the
sink). Only a range gateway (i.e. last cluster hop gateway), a leaf-mate (i.e. member with only
one neighbour), sink-ring-neighbour (i.e. a node that has exactly two neighbours of which one is
the sender sink), must respond by broadcasting a HELLO-ACK. (See Figure 1 for illustration).
Note that a HELLO-ACK message is forwarded only by nodes that have previously received a
HELLO message from the same sink.
After receiving a new HELLO-ACK from a neighbouru , a node v estimates the transmission
delay ,
u v
 using Equation (3); then increments the total-delay field with the latter estimated
delay ,
u v
 and finally forwards the message.
After HELLO-timer expiration, the sink must reply to each received HELLO-ACK message, via
the first forwarder node, by sending a CAST message where the field type is set to 0. This
message also contains the total delay value extracted from the HELLO-ACK. Such CAST
messages are forwarded using the same rules applied to the HELLO ones.
When a node receives a CAST message, it must reply by sending a CAST-ACK. Except that
sink-ring-neighbours must send a CAST-ACK message only after receiving a CAST message
forwarded by its non-sink neighbour. CAST-ACK messages must also contain the total delay
value extracted from the CAST ones. Besides, the cast_vector field of each CAST-ACK message
contains the values provided by its corresponding CAST message.
Therefore, when receiving a CAST-ACK message sent by node s , a node u can update its
convergecast table by calculating its transmission delay ,
u x
 to the sink x using Equation (5) where
,
u x
 , ,
s x
 and ,
s u
 respectively denote delays from node u to sink x , from the sender s to
sink x and from the sender s to node u . Then the latter must analyzes content of the cast_vector
field, in order to extract the unique number of which position equals value of the hop_count field.
The neighbour that was assigned the latter unique number is chosen as the next hop.
To help balance the load, every sink must shift to sleep mode after wake
t seconds, i.e. if its energy
consumption ratio  reaches a threshold  defined as a parameter.  is calculated using
Equation (7) where i
E and f
E respectively denote sink’s energy at the beginning and at the end
of its current duty-cycle. Before being inactive, sinks must alert their cluster mates by
broadcasting a SLEEP message. The latter will be forwarded like HELLO messages.
i f
i
E E
E


 (7)
This idle state will last sleep
t seconds. The latter duration is determined using Equation (8) where
 denotes the number of times a sink has been active; β > 0 is the Weibull distribution shape
parameter; R is uniformly and randomly chosen in the interval  
0;1 .
1
1 1
ln( )
sleep
t
R


  (8)
It is also noteworthy to mention that a CH must also trigger a timer (BUILD-timer) to help
reconstruct the intra-cluster part of the convergecast tree after build
t seconds. This duration is

9
determined using Equation (9); where i
t denotes the mean time between wake-ups of the ith
neighbouring sink while  is given as a parameter.
1
min( , ,..., )
, 2
i i n
build
t t t
t 


  (9)
Note that i
t is calculated then included in HELLO messages sent by sink i. Sink’s range
gateways also spread this information along with the list of active sinks when sending TREE-
ACK messages.
3.3.4. Data Sending Process
After each cluster member has found a path to the nearest sink, the data sending process must be
scheduled in order to favour aggregation and mitigate collisions or nodes’ congestion. To do this,
the data sending coordinator (i.e. the CH or the sink) broadcasts a SCHED message containing
the ID of a gateway randomly chosen then triggers a SCHED-timer.
When receiving this message, only periphery-members whose paths to their nearest sink pass
through the latter gateway must reply by sending a SCHED-ACK.
Note that inside a cluster the term “periphery-members” refers to nodes (except gateways) that
are located _ _
max hop count hops from the coordinator.
The sender of a SCHED-ACK message must specify the total delay to reach the designated
gateway in addition to the lowest capacity (i.e. forwarded node’s buffer length) encountered on
the path towards this gateway. This information is recorded during the path-finding process. After
expiration of the SCHED-timer, the data sending coordinator constructs a schedule leveraging all
the received SCHED-ACK messages. To do this, it must infer from ˆ
x
f (i.e. the buffer remaining
space) of gateway x the duration data
t of the current process.
Equation (10) helps to estimate data
f the portion of data that each periphery-member can send to
this gateway. x
N denotes the set of periphery-members eager to send data to gateway x .
ˆ
x
data
x
f
f
N
 (10)
Equation (11) helps estimate ( )
data
t u the time slot assigned to periphery-member u to send its
data. ,
u x
 denotes the delay between periphery-member u and the chosen gateway x .
,
( )
data data u x
t u f 
  (11)
After calculating the schedule (i.e. the set of durations ( )
data
t u ), the coordinator broadcasts an
AGGR message containing the chosen gateway’s ID and the amount of data data
f to be sent by
each periphery-member. The coordinator randomly chooses the periphery-member which should
start the process. The coordinator then triggers a timer AGGR-timer for ˆdata
t seconds. This
duration is estimated using Equation (12).

10
ˆ ( )
x
data data
u N
t t u

  (12)
When receiving such a message, only a periphery-member that has previously sent a SCHED-
ACK containing the chosen gateway’s ID, must start the data sending process. If so, such a
periphery-member will send its data through a DATA message to the next neighbour according to
its convergecast table.
Upon receiving such a message, a node must aggregate its own data (if destined to the same
gateway), before sending the result.
Note that, in this snowball effect, gateways act like ultimate aggregators of the data collected by
other cluster members.
After expiration of AGGR-timer, coordinator ends the current transmission session by randomly
choosing among the remaining gateways a new ultimate aggregator before broadcasting a new
SCHED message and so on.
Note that all these messages are forwarded based on the same rules applied to those used for the
path-finding process.
Figure 4 a) depicts SCHED-ACK messages sent by some periphery-members whose shortest path
to sink A goes through gateway T randomly chosen by the CH Z. Figure 3 b) shows the
convergecast and aggregation tree constructed by the CH based on all the SCHED-ACK
messages received. Gateway T is the ultimate aggregator of data sent by periphery-members U,
R, Y and X. Node U is randomly chosen to initiate the data sending process according to
durations in the schedule calculated by CH. Figure 5 and 6 depicts the processes used by CFMS.
Figure.4 Example of Convergecast tree construction: a) Periphery-members responses. b) The resulting
Direct Acyclic Graph (DAG).

11
Update
Convergecast
Table
New
TREE-ACK
Message
Start
End
Broadcast
TREE-REQ
Message
CH
Check for
Messages
Trigger
TREE-timer
New
TREE-REQ
Message
No
Apply TREE-REQ
messages
forward and
discard rules
Broadcast
TREE-ACK
message
Yes
TREE-timer
or HELLO-timer
triggered
New cluster or
BUILD-timer
expired
Enough
Energy
Check
Battery
No
No
New
HELLO
Message
Discover
Neighbourhood
and Recluster
Sink
Apply TREE-ACK
messages forward
and discard rules
Yes
No
Apply HELLO
messages forward
and discard rules
No
Broadcast
HELLO
Message
No
New
HELLO-ACK
Message
New
CAST
Message
Yes
Yes
No
No
Yes
Trigger
CAST-timer
Yes
TREE-timer
or HELLO-timer
expired
Yes
Yes
Yes
No
No
New
SLEEP
Message
No
No
Enter into the
data sending process
No
WAKE-timer
expired
Broadcast
SLEEP
Message
No Trigger
SLEEP-timer
Yes
SLEEP-timer
triggered
SLEEP-timer
expired
Trigger
WAKE-timer
Yes
No
Yes
Yes
Yes No
TREE-ACK
eligble
Yes
CAST-timer
expired
Apply HELLO-ACK
messages forward
and discard rules
Yes
HELLO-ACK
eligble
No
Broadcast
HELLO-ACK
message
Yes
Trigger
HELLO-timer
Apply SLEEP
messages
forward and
discard rules
CAST-timer
triggered
Yes
Yes
Trigger
BUILD-timer
Apply CAST
messages forward
and discard rules
CAST-ACK
eligble
Broadcast
CAST-ACK
message
New
CAST-ACK
Message
Yes
Apply CAST-ACK
messages forward
and discard rules
No
Yes
Yes
No
No
Broadcast
CAST
Message
No
Figure.5 Flowchart of the path-finding process of CFMS

12
New
SCHED-ACK
Message
Start
End
Check for
Messages
New
SCHED
Message
No
Apply SCHED
messages
forward and
discard rules
Send
SCHED-ACK
message
Yes
SCHED-timer
triggered
Yes
New
AGGR
Message
Apply SCHED-ACK
messages forward
and discard rules
No
Apply AGGR
messages forward
and discard rules
New
DATA
Message
Yes
Yes
SCHED-timer
expired
Yes
No
No
Yes
Perform the
latter Task
SCHED
eligble
Yes
Scheduled
Task
Yes
AGGR
eligble
No
Apply DATA
messages
discard rules
Data
Send
new DATA
message
No
Yes
No
Yes
Broadcast
AGGR
Yes
No
Yes
Broadcast
SCHED
Message
Choose a new
gateway
Sink
or CH
Enter into the
path-finding
process
Trigger
SCHED-timer
No
No
Trigger
AGGR-timer
Get
process
duration
AGGR-timer
triggered
AGGR-timer
expired
No
Aggregate
Forward
DATA
message
Yes
Check
AGGR-timer
No
Gateway to
schedule
No
Yes
No
Chosen
gateway
Buffer
full
Yes
Check
Data
No
Yes
No
Yes
Figure.6 Flowchart of the path-finding process of CFMS

13
4. EXPERIMENTAL SET-UP
In this section, we detail the extensive simulation campaign we carried out. Experiments were
made with OMNeT++ simulator version 6.0 [36]. We used the energy consumption model
proposed by Heinzelman et al. [17]; the other parameters are summarized in Tables 1 - 4. The
results were compared to those we obtained with our previous work named MSCP (Multi-Sink
Convergecast Protocol) and the solution proposed by Singh and Nagaraju [28] that we will refer
to as S&N.
Table 3 presents parameters we used to randomly and uniformly vary link quality; namely, PRR
(Packet Reception Ratio), SNR (Signal to Noise Ratio), SINR (Signal Interference plus Noise
Ratio) and LQI (Link Quality Indicator) [37]. Tables 2 and 4 are inspired by the specifications of
the IEEE 802.15.4 standard [38]. We evaluated the ability of three protocols to efficiently transfer
data to the sinks through three metrics, namely, the average end-to-end delay, the packet delivery
ratio, and the network lifetime [32]. To do this, we randomly and uniformly deployed sensors and
sinks varying their population as described in Table 1, respectively using, a 100 and a 2 steps
scale so that the sink-to-sensor ratio is 0.02. We specifically investigated how the latter
populations influenced these metrics. To vary link quality, we used the uniform distribution to
randomly change parameters described in Table 3. Each 2.5s, 30% to 50% of nodes were
randomly chosen to send data. The latter was queued in a buffer in order to wait for a
transmission schedule when we applied CFMS. We used in CFMS the same data aggregation
scheme proposed for S&N. The reason for that is twofold; first, providing a specific data
aggregation scheme is beyond the scope of this paper and secondly, we wanted to evaluate CFMS
and S&N on the same basis.
This experiment was replicated 50 times for each variation of the number of nodes. Results were
averaged with a 95% confidence interval. The experiment started after all the nodes were
deployed and was ended according to two network lifetime definitions respectively when a sensor
and a sink depleted is energy.
Table.1 Simulation general parameters
Parameter Value
deployment zone
number of sensors
number of sinks
sensors’ transmission ranges
sinks’ transmission ranges
max_hop_count
max_cluster_hop_count
sensors’ initial energy
sinks’ initial energy
self-discharge per second
Eelec (see [17])
efs (see [17])
eamp (see [17])
d0 (see [17])
length of a data packet
β Weibull distribution shape
Wu,v bandwith of a link (u,v)
 Threshold of energy consumption ratio
Nodes’ buffer size
1000 m X 1000 m
100 - 1000
2 - 20
127 m
250 m
2
2
0.2 J
20 J
0.1 J
50 nJ/bit
10 pJ/bit/m2
0.0013 pJ/bit/m4
87 m
2000 bits
3
2.4 GHz
0.01
100 data packets

14
Table.2 Transition state delay
Rx (s) Tx (s) Sleep (s) Idle
Rx - 1 194 -
Tx 1 - 194 -
Sleep 5 5 - -
Idle - - - -
Table.3 Link quality parameters
PRR SNR (dBm) SINR(dBm) LQI
Excellent 1 ]40; 60] ]30; 40] ]106; 255]
Good ]0.75; 1[ ]25; 40] ]15; 30] ]102; 106]
Medium ]0.35; 0.75] ]15; 25] ]5; 15] ]80; 102]
Poor [0; 0.35] [0; 15] [0; 5] [0; 80]
Table.4 Transition state energy consumption
Rx (mW) Tx (mW) Sleep (mW) Idle
Rx - 62 62 -
Tx 62 - 62 -
Sleep 1.4 1.4 - 1.4
Idle - - 1.4 -
5. RESULTS AND DISCUSSIONS
In the following, we analyse and explain results of the experiments described in the previous
section.
5.1. Average End-To-End Delay
Figure 7 shows the effect of number of nodes on packets delivery delay. These delays appear to
increase according to network size when using CFMS and MSCP. In contrast, when MSCP is
applied, delays increase only for networks with less than 500 sensors then decrease.
Figure.7 Number of sensor nodes vs. End-to-end delay

15
Figure.8 Number of sensor nodes vs. PDR
This is mainly due to the delay caused by the data aggregation process. However, it is noteworthy
that when using CFMS delays slightly increase for sparse networks (with number of sensors >
500); but actually remain constant for dense ones. In other words, CFMS helps to better mitigate
additional aggregation delay than S&N. Indeed, CFMS use the delay-based route selection
scheme inspired from MSCP during its path-finding phase. Decisions implicitly consider both
links state (via the SINR metric) and level of congestion of the intermediary nodes. Therefore,
sensor nodes always choose the nearest sink considering lowest transmission and queuing delays.
In contrast, S&N leverages only the Euclidean distance to the sinks.
5.2. Packet Delivery Ratio
Figure 8 suggests that the three evaluated protocols yield ratios higher than 90%. However, the
values obtained with MSCP are around 98% while CFMS reaches values above 99% irrespective
of the number of sensor nodes. This is also due to the path selection metric (SINR + congestion
level) used by both CFMS and MSCP. Except that the latter protocol helps mitigate packet losses
due to both collisions and nodes’ congestion by scheduling data transmissions. Hence, the ratios
close to 100%. Whereas the data delivery process of S&N is aimed to prevent congestions.
Figure.9 Number of sensor nodes vs. Network lifetime (Until a node dies)

16
Figure.10 Number of sensor nodes vs. Network lifetime (Until a sink dies)
5.3. Network Lifetime
Figures 9 and 10 show that irrespective of the protocol used, network lifetime increases with its
size; since high node degrees help provide more alternative routes. However, MSCP and CFMS
yield the best durations. This is mainly due to their route selection metric that considers both link
state and nodes’ congestion; as discussed above, this scheme help mitigate packet losses hence
energy consuming retransmissions. Additionally, these protocols favour CHs’ rotation and sinks’
sleep/wake scheduling. Furthermore, MSCP and CFMS significantly mitigates energy waste, by
reducing the range and the number of signaling messages. However, the data aggregation and
transmission scheduling process of CFMS helps mitigate congestions and collisions while further
reducing the number of packets.
On the contrary, S&N yields the worst results since this protocol only leverages Euclidean
distance and link node’s congestion level to route packets to the sinks. No policy is explicitly
applied in this protocol to mitigate energy wastes due to collisions or links’ state variation.
6. CONCLUSIONS
In this paper we presented CFMS, a framework for convergecast in large-scale multi-sink WSNs.
We combined clustering technique to path vector-based routing and duty cycling schemes for
both sensor nodes and sinks. This solution was inspired from our previous work MSCP (Multi-
Sink Convergecast Protocol). During route selection phase, transmission and queuing delays are
still estimated by considering nodes’ level of congestion and link interferences via a SINR
(Signal-to-Interference-plus-Noise-Ratio)-based metric. However, we replaced nodes ‘ability to
individually transmit data by a collaborative scheme. This implied the design of two processes
inside clusters as well as in sinks’ one-hop neighbourhood. The first one is dedicated to data
aggregation tree construction and the second one to transmissions scheduling. The duration of
each transmission session is calculated according to gateways’ congestion level in order to
minimise collisions. The resulting fully-distributed and proactive strategy helped balance the
load between sinks, mitigate message overhead and packet losses. Simulation results showed that
CFMS enhances packet delivery ratio and network lifetime; though slightly increasing latency
obtained with MSCP. CFMS outperforms S&N a related protocol recently proposed in the
literature.
As a future work, we will focus on scenarios with mobile sinks and sensor nodes. We also plan
to provide this solution with security, a specific data fusion module and an explicit congestion
control scheme that considers packet priority as required by real time applications.

17
REFERENCES
[1] Kayhan Erciyes, (2019) Case study: Environment monitoring by a wireless sensor network,
Computer Communications and Networks,.Springer International Publishing.
[2] Dionisis Kandris, Christos Nakas, Dimitrios Vomvas, and Grigorios Koulouras, (2020)
“Applications of wireless sensor networks: An up-to-date survey”, Applied System Innovation, Vol.
3, No.1, pp.14.
[3] Kamal Gulati, Raja Sarath Kumar Boddu, Dhiraj Kapila, Sunil L. Bangare, Neeraj Chandnani, and
G. Saravanan, ( 2022) “A review paper on wireless sensor network techniques in internet of things
(IoT)”, Materials Today: Proceedings, Vol. 51, pp. 161–165.
[4] Asside Djedouboum, Ado Abba Ari, Abdelhak Gueroui, Alidou Mohamadou, and Zibouda
Aliouat. (2018) “Big data collection in large-scale wireless sensor networks”, Sensors, Vol 18, No
12:4474.
[5] Deepak Sharma, Amritesh Ojha, and Amol P. Bhondekar. “Heterogeneity consideration in wireless
sensor networks routing algorithms: a review”, The Journal of Supercomputing, Vol.75, No.5, pp.
2341–2394.
[6] Filipe Araujo, André Gomes, and Rui P. Rocha, (2020) “Towards optimal convergecastin wireless
ad hoc networks”, Ad Hoc Networks, Vol.107, No 102214.
[7] Jobish John, Gaurav S. Kasbekar, and Maryam Shojaei Baghini, (2021) “Maximum lifetime
convergecast tree in wireless sensor networks. Ad Hoc Networks, Vol.120, No.102564.
[8] Arnab Roy, Joseph Lalnunfela Pachuau, and Anish Kumar Saha, (2021) “An overview of queuing
delay and various delay based algorithms in networks”, Computing, Vol.103, No.10, pp.2361–2399.
[9] Aysegul Tuysuz Erman, Thijs Mutter, Lodewijk van Hoesel, and Paul Havinga, (2009) A cross-
Layered communication protocol for load balancing in large scale multi-sink wireless sensor
Networks, International Symposium on Autonomous Decentralized Systems, IEEE.
[10] Tarunpreet Kaur and Dilip Kumar, (2019) “Computational intelligence-based energy efficient
routing protocols with QoS assurance for wireless sensor networks: a survey”, International Journal
of Wireless and Mobile Computing, Vol.16, No.2 172.
[11] Yuan Chai and Xiao-Jun Zeng,“Delay-and interference-aware routing for wireless mesh
network”, (2020) IEEE Systems Journal, Vol. 14, No. 3, pp. 4119–4130.
[12] Yunquan Gao, Xiaoyong Li, Jirui Li, and Yali Gao, (2019) “Distributed and efficient minimum-
data aggregation scheduling for multichannel wireless sensor networks”, IEEE Internet of Things
Journal, Vol. 6, No.5, pp. 8482–8495.
[13] Louie Chan, Karina Gomez Chavez, Heiko Rudolph, and Akram Hourani, (2020) “Hierarchical
routing protocols for wireless sensor network: a compressive survey”, Wireless Networks, Vol. 6,
No.5, pp.3291–3314.
[14] Shashi Bhushan, Manoj Kumar, Pramod Kumar, Thompson Stephan, Achyut Shankar, and Peide
Liu. (2021) “FAJIT: a fuzzy-based data aggregation technique for energy efficiency in wireless
Sensor Network”, Complex & Intelligent Systems, Vol. 7, No.2, pp.997–1007.
[15] Shamim Yousefi, Hadis Karimipour, and Farnaz Derakhshan, (2021) “Data aggregation
mechanisms on the internet of things: A systematic literature review”, Internet of Things, Vol.
15, No.100427.
[16] Mandeep Kaur and Amit Munjal, (2020) “Data aggregation algorithms for wireless sensor network:
A review”, Ad Hoc Networks, Vol.10, No.102083.
[17] Wendi Beth Rabiner Heinzelman, Anathan.P. Chandrakasan, and Hari Balakrishnan (2002) “An
application-specific protocol architecture for wireless microsensor networks”, IEEE Transactions on
Wireless Communications, Vol. 1, No. 4, pp. 660–670.
[18] Ikram Daanoune, Baghdad Abdennaceur, and Abdelhakim Ballouk. (2021) “A comprehensive
survey on LEACH-based clustering routing protocols in wireless sensor networks”. Ad Hoc
Networks, Vol. 114 No.102409.
[19] Fanpyn Liu, (2021) “Majority decision aggregation with binarized data in wireless sensor
networks”, Symmetry, Vol. 13, No.9:1671.
[20] Amin Shahraki, Amir Taherkordi, Oystein Haugen, and Frank Eliassen, (2021) “A survey and
future directions on clustering: From WSNs to IoT and modern networking paradigms”, IEEE
Transactions on Network and Service Management, Vol.18, No.2, pp. 2242–2274.

18
[21] G. Saranraj, K. Selvamani, and P. Malathi, (2021), “A novel data aggregation using multi objective
based male lion optimization algorithm (DA-MOMLOA) in wireless sensor network”, Journal of
Ambient Intelligence and Humanized Computing, Vol.13, No.12, pp. 5645–5653.
[22] Sedigheh Sadat Sharifi and Hamid Barati, (2021) “A method for routing and data aggregating in
cluster-based wireless sensor networks”, International Journal of Communication Systems, Vol. 34,
No. 7.
[23] Vani S. Badiger and T.S Ganashree, (2022) “Data aggregation scheme for IOT based wireless
sensor network through optimal clustering method”, Measurement: Sensors, Vol.24, No.100538.
[24] Bo Yu and Jianzhong Li, (2011) Minimum-time aggregation scheduling in multisink sensor
networks, 2011 8th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad
Hoc Communications and Networks, IEEE.
[25] Nitin Goyal, Mayank Dave, and Anil Kumar Verma, (2017) Improved data aggregation for cluster
based underwater wireless sensor networks. Proceedings of the National Academy of Sciences,
India Section A: Physical Sciences.
[26] Wahab Khan, Hua Wang, Muhammad Shahid Anwar, Muhammad Ayaz,Sadique Ahmad, and Inam
Ullah, (2019) “A multi-layer cluster based energy efficient routing scheme for UWSNs”, IEEE
Access, Vol.7, pp. 77398–77410.
[27] Anagha Rajput and Vinoth Babu Kumaravelu , (2020) “Fuzzy-based clustering scheme with sink
selection algorithm for monitoring applications of wireless sensor networks”, Arabian Journal for
Science and Engineering, Vol. 45, No.8, pp. 6601–6623.
[28] Amit Singh and A. Nagaraju, (2020) “Low latency and energy efficient routing aware network
coding-based data transmission in multi-hop and multi-sink WSN”, Ad Hoc Networks, Vol.107,
No.102182.
[29] Aarti Jain and B.V.R. Reddy. (2014) Sink as cluster head: An energy efficient clustering method for
wireless sensor networks, International Conference on Data Mining and Intelligent Computing
(ICDMIC), IEEE.
[30] Gokou Hervé Fabrice Diédié, Boko Aka, and Michel Babri, (2018), “Energy-efficient ant colony-
based k-hop clustering and transmission range assignment protocol for connectivity construction in
dense wireless sensor networks”, Journal of Computer Science, Vol.14, No. 3, pp. 376–395.
[31] Gokou Hervé Fabrice Diédié, Boko Aka, and Michel Babri, (2019) “Self-stabilising hybrid
,connectivity control protocol for WSNs”, IET Wireless Sensor Systems, Vol. 9, No. 1, pp. 6–24.
[32] Mohamed Skander Daas, Salim Chikhi, and El-Bay Bourennane, (2021) “A dynamic multi-sink
routing protocol for static and mobile self-organizing wireless networks: A routing protocol for
internet of things”. Ad Hoc Networks, Vol.117, No.102495.
[33] Sandeep Verma, Neetu Sood, and Ajay Kumar Sharma, (2019) “Genetic algorithm based optimized
cluster head selection for single and multiple data sinks in heterogeneous wireless sensor network”,
Applied Soft Computing, Vol. 85, No.105788.
[34] Sampath Reddy Chada and Narsimha Gugulothu, (2022) “Secure and energy aware cluster based
routing using trust centric-multiobjective black widow optimization for large scale WSN”,
International Journal of Electrical and Computer Engineering Systems, Vol. 13, No.7, pp. 525–532.
[35] Gokou Hervé Fabrice Diédié, Koigny Fabrice Kouassi, and Tchimou N’Takpé, (2022) Multi-sink
convergecast protocol for large scale wireless sensor networks. In Signal, Image Processing and
Embedded Systems Trends, Academy and Industry Research Collaboration Center (AIRCC).
[36] Varga, A.: 2022. Available at ‘https://omnetpp.org’, accessed on August 2022.
[37] Carlo Alberto Boano, Marco Antonio Zuniga, Thiemo Voigt, Andreas Willig, and Kay Romer,
(2010) The triangle metric: Fast link quality estimation for mobile wireless sensor networks, 19th
International Conference on Computer Communications and Networks, IEEE.
[38] Bruno Bougard, Francky Catthoor, Denis Clarke Daly, Anantha Chandrakasan, and Wim
Dehaene, (2005) Energy efficiency of the IEEE 802.15.4 standard in dense wireless microsensor
networks: Modeling and improvement perspectives, Design, Automation and Test in Europe, IEEE.

19
AUTHORS
Gokou Hervé Fabrice Diédié received his Ph.D. in Computer Science at Université
Nangui Abrogoua (UNA) in Abidjan, Ivory Coast in September 2018. He received his
M.S. in Computer Science from the same university in 2012. His research interests
include distributed algorithms, self-organization in ad hoc networks and intelligent
transportation systems. He is currently a Professor and Researcher in Computer Science
and a member of Laboratory of Mathematics and Computer Science at Université
Peleforo Gon Coulibaly of Korhogo, Ivory Coast.
Koigny Fabrice Kouassi received his M.S. in Computer Science at Université Nangui
Abrogoua (UNA) in Abidjan, Ivory Coast in November 2021. He received his BS in
Computer Science from the same university in 2018. His research interests include
distributed algorithms and hoc networks. He is a member of and Researcher in
Computer Science and a member of Laboratory of Mathematics and Computer Science
at UNA.
Tchimou N'Takpé received his Ph.D. in Computer Science at Nancy University,
France, in January 2009. He received his MS in Engineering from Ecole Nationale
d'Electricité et de Mécanique at Nancy, France, in 2005. His research interests include
distributed and parallel algorithms and Networking. He is currently a Professor and
Researcher in Computer Science and a member of Laboratory of Mathematics and
Computer Science at Université Niangui Abrogoua (UNA) in Abidjan, Ivory Coast.

CFMS: A Cluster-based Convergecast Framework for Dense Multi-Sink Wireless Sensor Networks

More Related Content

Similar to CFMS: A Cluster-based Convergecast Framework for Dense Multi-Sink Wireless Sensor Networks (20)

Recently uploaded (20)

CFMS: A Cluster-based Convergecast Framework for Dense Multi-Sink Wireless Sensor Networks