Content uploaded by Oscar Novo
Author content
All content in this area was uploaded by Oscar Novo on Dec 04, 2018
Content may be subject to copyright.
Gateway Selection in Capillary Networks
Nicklas Beijar, Oscar Novo, Jaime Jim´
enez and Jan Melen
Ericsson Research
Jorvas, Finland
Email: firstname.lastname@ericsson.com
Abstract—The world is adopting machine-type communica-
tion, wherein sensors and actuators blend seamlessly with the
environment around us, bringing a new ubiquitous computing
and communication era – a shift that is, to some extent,
illustrated by the explosive growth of the Internet of Things
(IoT). Capillary Networks play an important role in the growth
of IoT, enabling wireless sensor networks to connect and use the
capabilities of cellular networks through Capillary Gateways. In
that sense, Capillary Gateways facilitate the seamless integration
of wireless sensor networks with cellular networks. Therefore,
an optimal selection of the Capillary Gateways by the wireless
sensor network is crucial for balancing the load between the
gateways and optimizing the end-to-end path through both
networks. This paper describes a set of possible gateway selection
architectures and presents an algorithm for determining the
gateway selection based on policies and a set of constraints.
Then, the paper describes our implementation of two selected
architectures, discussing the solutions and challenges encountered
during implementation. Finally, the paper evaluates the traffic
and load generated by gateway selection.
Keywords—M2M, IoT, Internet of Things, Capillary Network,
self-organizing networks.
I. INTRODUCTION
With the growing presence of short-range radio-access
technologies and the proliferation of multiple heterogeneous
devices connected to Internet, the society is moving towards
a new vision, where a wide range of objects become part
of the Internet. The decreasing connectivity and technology
costs, as well as the increasing network penetration, are key
factors enabling that vision called the Internet of Things [1].
Internet of Things (IoT) encompasses any device that utilizes
embedded technology to communicate and interact with other
devices via Internet. This includes everything from automo-
tives, environmental sensors, and industrial machinery to home
appliances and wearable devices. The application possibilities
are endless and will profoundly change existing industries
such as health, logistics, agriculture and manufacturing as well
as create new unforeseen ones. Although IoT is emerging
as a technology, the current communication paradigm will
need to go beyond the traditional communication scenarios
– which have primarily been built to provide connectivity for
human interaction – and evolve into a scenario of billions of
ubiquitous interconnected devices.
According to market predictions [2], the majority of the
IoT devices are expected to use short-range radio technologies
such as Bluetooth Low Energy [3], IEEE 802.15.4 [4] or IEEE
802.11ah [5]. In this case, cellular networks can play a valuable
role in connecting the short-range radio networks, leveraging
on its ubiquity and advanced connectivity for backhaul, and
integrating security and network management into the scene.
Fig. 1: Capillary networks
The short-range network would act as the fine-grained last-
hop network connecting the tiniest devices. Thereby, they are
referred to as Capillary Network [6], [7], comparing them with
the tiny blood vessels connecting cells to the arteries. Capillary
Networks go beyond sensor networks, with not only sensor
data being transferred from devices to the cloud, but also cloud
applications being able to perform actuation on the devices.
Capillary networks consist of a set of devices connected
using short-range radio access technologies to a more powerful
device called Capillary Gateway. The gateway connects the
devices to the cellular backhaul network, transporting data to
an IoT Cloud service, which aggregates the incoming data
and manages devices and gateways. Figure 1 illustrates a
communication model between different capillary networks
(WIFI, Bluetooth) and two capillary gateways using the same
cellular network as backhaul (LTE). The sensor data is sent to
an IoT Cloud service that offers data aggregation and storage.
The IoT Cloud service also enables operators to offer services
to enterprise customers such as connectivity management of
the devices.
Typically, Capillary Network deployments do not involve
any network planning. The networks should be able to auto-
matically configure themselves in order to be easy to deploy.
This typically implies over-provisioning the network with
redundant gateways, which may have different properties in
terms of connectivity, capacity and – in the case of battery
powered gateways – also energy. Given the assumption that
there are multiple capillary gateways within reach offering
communication services with different properties, the capillary
network should be able to direct the sensors to the best capil-
lary gateway in order to achieve a specific goal. Examples of
such goals can be minimizing latency, maximizing availability
or providing load balancing. The mechanism that leads sensors
to the best capillary gateway is called Gateway Selection.
This mechanism reduces manual configuration of network
connectivity, provides redundancy of gateways and optimizes
the communication path across network technology borders.
This paper presents and evaluates various solutions for
gateway selection. Specifically, Section II presents three dif-
ferent architectures based on the required information, while
Section III further details the mechanism for calculating the
selection. Section IV applies the solution to existing proto-
cols and describes our prototype implementations. Different
approaches taken in the implementation of the gateway selec-
tion mechanism are evaluated in section V, and Section VI
concludes the paper.
II. GATE WAY SELEC TI ON ARCHITECTURE
The gateway selection process depends on various types
of information that must be collected from the capillary
network. These gateway selection parameters are the input
for the gateway selection algorithm. Several architectures are
possible, depending on where in the network the selection
process is executed. The choice of location affects the need
to transport information and thereby the performance. This
section discusses the gateway selection parameters along with
the location of the gateway selection process.
A. Gateway Selection Parameters
Gateway selection relies on three types of information:
reachability information, constraints and policy. According to
these three parameters, the gateway selection algorithm selects
the best candidate for the capillary network.
1) Reachability information: Reachability information de-
scribes the possible connectivity between the IoT devices in a
capillary network and the capillary gateways around them. The
possibility of a link is determined by the existence of a radio
signal with sufficient strength in the short-range network. The
reachability information is therefore, at it simplest, described
as a set of possible connectivity links. However, the informa-
tion can be augmented with link quality data such as signal
strength, uptime or packet loss for each link. The reachability
information can be detected by either the gateway or by the IoT
devices. IoT devices typically detect connectivity by listening
to the beacon signals that gateways transmit.
2) Constraints: The constraint information describes the
properties of the network and the nodes that are included in the
selection process. Gateway properties can include the load of
the gateway (described as the processor load or by the number
of connected devices), the remaining battery level, or the cost
of use (backhaul costs, transit fees). Constraints describing the
cellular link quality and bandwidth allow selecting a gateway
with good uplink, so that the path through both network types
can be optimized. Most constraints can be specified on a per
gateway basis.
3) Policy: The policy determines the goal of the gateway
selection. Fundamentally, a policy is a set of priorities that
determines how the various constraints affect the best choice
of gateways. Policies are static and are defined by the network
Fig. 2: Control Point Architecture
management. The policy model is described in more detail in
III-A.
B. Location of the Gateway Selection Process
The gateway selection process can be controlled at various
locations in the network. The control point can be centralized,
distributed between the gateways, or distributed between the
devices. The location of the node(s) that controls the selection
process affects the need to transport the gateway selection
parameters to different places in the network. Below, the paper
discusses the alternatives in more detail:
1) Centralized: In the centralized alternative, there is a
centralized gateway selection server managing the selection as
shown in Figure 2 A. The reachability information is sent from
the devices or from management proxies in the gateways to
the management server. The constraint information is collected
from the gateways and sent to the management server. The
policy information is updated from the management interface
to the management server. The centralized server runs the
gateway selection algorithm creating a target distribution.
Changes in the target distribution are sent via the management
proxies to the concerned devices.
Since all the information is centrally available, there are no
convergence problems in the selection process. However, the
management server acts as a single point of failure and the
capacity of the management server limits the scalability of the
network. This is the major drawback in this design, although
it could be solved by redundant servers.
2) Distributed: In the distributed alternative, the selection
of the best candidate is done by the gateways themselves as
shown in Figure 2 B. In this approach, the gateways share the
reachability, constraint and policy information among them and
every gateway runs the same gateway selection algorithm to
choose the right candidate.
The main challenge of this alternative is the communication
needed between the gateways to synchronize the information.
To carry out this communication, either the cellular links or
the capillary network itself must be used. Both alternatives
have limited capacity. Moreover, because of the time required
by replication, the information at the gateways might be
inconsistent at some points in time. That inconsistency might
cause devices to get inconsistent commands from different
gateways during short periods of time.
In the case of large domains, the communication should
be limited to a minimum set of gateways to reduce delays.
Thus, gateways replicate information only to other nearby
gateways. Alternatively, one gateway could be established
as the decision-making node. This approach would be very
similar to the centralized approach above. This reduces the
time for replicating the information but, instead, the system
would become less robust due to having a single point of
failure.
3) Autonomous devices: In the autonomous devices alterna-
tive, the IoT devices make their own decision to which gateway
they connect. Figure 2 C shows this model. The reachability
information is then produced and utilized locally by the
devices. The constraint and policy information is requested by
every single IoT device from the gateways or from particular
servers in the network. In the latter case, the IoT device must
indicate to a constraint server the list of potential gateways,
and the constraint server replies with the constraints of the
aforementioned gateways.
One of the challenges with this approach is the complexity
required at the devices to run the gateway selection algorithm.
Typical IoT devices have limited processing power and mem-
ory, and executing the algorithm consumes the device’s battery.
Thus, this approach would be limited to IoT devices with
enough capacity or, in some cases, used in a hybrid approach
where the limited devices use a centralized approach instead.
Furthermore, it is difficult to implement load distribution in
this approach, since there is no coordination between the IoT
devices. Oscillation easily appears. For example, a gateway
with low load may cause a race of devices connecting to it,
suddenly increasing the load above the average.
Fig. 3: Policy Tree
III. GATEWAY SELE CT IO N PROCESS
This section describes how the gateway selection process
works and how the best candidate is chosen according to the
gateway selection parameters. First, we describe how to model
the gateway selection parameters within the gateway selection
algorithm. Afterwards, we present the gateway selection algo-
rithm itself and provide an implementation example.
A. Algorithm Input
The gateway selection algorithm takes the policy, the
constraints and the reachability information as input in order
to produce a selection of gateway for each device as output.
1) Policies: To enable expressive policies, the policies are
described using a decision tree. A node in the tree can have
ordered branches toward other nodes as depicted in Figure 3.
Each branch carries a condition determining when the branch
is traversed. Processing starts from the root and traverses the
first of the ordered branches for which the condition is fulfilled.
The condition can specify that a branch is to be traversed for
a given (1) device, (2) gateway, (3) interface, or (4) node
type. Once no other branches can be traversed, the policy
description of that node is applied. The policy description is a
list of constraint names and the corresponding weights of those
constraints. Additionally, a constant value can be given with a
special constraint named ”priority”. Using this simple model,
a versatile range of policies can be described. The policy tree
can be described using JavaScript Object Notation (JSON) or
eXtensible Markup Language (XML). The former is used in
our implementation.
2) Reachability: Reachability information is collected from
devices and/or gateways. Based on the reachability informa-
tion, a full or partial reachability graph is built. The reach-
ability graph Rhas a link (g, d)between gateway gand
device dif either ghas reported reachability to d, or dhas
reported reachability to g. A timer is attached to each link in
the reachability graph. Reachability links that have not been
refreshed will be deleted from the graph after a time out.
3) Constraints: Constraints are key-value pairs describing
the properties of a gateway. The key is the constraint name c
and the value is a numeric value vc(g)of the constraint for a
specific gateway g.
B. Gateway Selection Algorithm
The algorithm operates by iterating through all devices to
find a target gateway for each device. In order to maintain
stability, the devices are iterated in the order of joining the
network. Thus, a recently joined device does not cause changes
in the allocation of the existing devices, but rather a suitable
gateway is determined for the newly joined device instead.
To find the gateway for a device d, a preference value
p(g)is calculated for each gateway gthat is in the set of
reachable gateways G={g:∃(g, d)∈R}. For the specific
combination of gateways and devices, a node in the policy
tree is obtained as described in the previous subsection. The
policy node defines a set of constraints Cwith a weight wcfor
each constraint c∈C. Additionally, a constant value kmay
be defined (otherwise k= 0). The constant value kis defined
in Section III-A as ”priority”. The preference of a gateway g
is calculated by weighting the constraint value vc(g)with the
corresponding weight wc:
p(g) = k+X
c∈C
wcvc(g)(1)
The gateway with the highest preference is selected for the
device d.
gsel(d) = argmax
g∈G
p(g)(2)
Additionally, for the purpose of load distribution, a particu-
lar constraint cconn(g)is dynamically defined for each gateway
g. This constraint reflects the number of devices allocated to
a gateway during a particular point in the algorithm. At the
start of the algorithm, the value is reset to cconn(g) = 0,
∀g∈G. Each time a node is allocated to a gateway gthe
constraint is increased: cconn(g) = cconn (g)+1. The constraint
is used in the policy and the preference calculations like other
constraints.
C. Example
Let us use an example to illustrate the use of policies
to obtain different selection results. The example policy de-
scription in Figure 4 contains a generic policy (lines 11-12),
an exception policy for a specific device (lines 2-7), and an
exception policy for a given device type (lines 8-10). The
conditions for selecting the exception policies, i.e. the branches
in the policy tree, are specified on lines 2 and 8, respectively.
The generic policy gives the same weight for considering
both battery and load, but with different sign. Thus, the
preference of a gateway increases when the battery level of
the gateway increases, while the preference decreases when the
Fig. 4: Policy in JSON Format
load level increases. A gateway Awith constraints {”load”: 1,
”battery”: 5}gets a preference of −2∗1+2∗5 = 8, while
a gateway Bwith constraints {”load”: 3, ”battery”: 4}gets a
lower preference of −2∗3 + 2 ∗4 = 2. Thus, gateway Ais
selected.
The type-specific policy defines a branch for devices of
the type ”alarm”. The node at the branch instructs that alarm
devices should be connected to the gateway with the highest
reliability constraint.
The device-specific policy is defined by a branch for a
given device identity. Within the branch, a particular gateway
is defined (using a branch in the policy tree) to have a priority
of 100 while all other gateways have the priority of 1; thus,
the device is allocated to the indicated gateway unless it is
unavailable. The special constraint ”priority” is used to specify
constant values.
IV. IMPLEMENTATION
During our implementation phase, both the distributed
and the centralized architecture were implemented. The au-
tonomous architecture was deemed impractical for constrained
devices due to its requirements on processing power and
memory, the usage of which were already close to the limits.
This section describes the implementation from a technical
viewpoint while the next section will explain the reason behind
each choice made.
A. Implementation of Distributed Architecture
In the distributed architecture, gateways share their con-
straints and reachability information. Each gateway imple-
ments the gateway selection algorithm in a replicated way
based on the information. The information must be synchro-
nized between gateways to ensure that all gateways obtain the
same result.
Since gateways need to share information, they need to
have connectivity to each other. This could be achieved using
the cellular uplink, the capillary network itself, or through a
separate network. Connecting all gateways with each other in
a mesh would not be scalable, therefore the network was split
into gateway selection domains. Within each domain, gateways
share information and devices can only move between the
gateways in the domain. 1In our prototype implementation,
we connected gateways with Ethernet for simplicity.
For distributing information between the gateways, we
extended the Open Shortest Path First (OSPF) [8] routing pro-
tocol with custom link state advertisements. When a constraint
value of a gateway changes, this gateway advertises the new
constraint value to all other gateways over OSPF. To reduce
the backhaul traffic caused by downloading policies from
the management server, the gateway with the highest OSPF
identifier is selected as the master. The master periodically
downloads the policy and distributes any changes between
the gateways. The policy is provided using a HTTP REST
[9] interface to the gateways. Consequently, the management
server only has a minimal role in supporting gateway selection
by providing the policy. 2
In this implementation, TMote Sky [10] devices were
used as sensors. They run 6LoWPAN [11] over an IEEE
802.15.4 interface. We based our device implementation on
the Contiki operating system, with the RPL (IPv6 Routing
Protocol for Low-Power and Lossy Networks) [12] routing
protocol between the gateway and the device. RPL implements
the messaging for allowing the gateway and the device to
detect each other. At the application layer devices imple-
ment two protocols: Constrained Application Protocol (CoAP)
[13] for data communication, and Lightweight M2M Protocol
(LWM2M) [14] for device management part. LWM2M objects
at the device represent the reachable gateways, the target
gateway selection and the currently connected gateway.
The IEEE 802.15.4 interface of the gateway is implemented
by using a USB-connected TMote Sky device, which operates
as the root for RPL. RPL uses a Destination-Oriented Directed
Acyclic Graph Identifier (DODAGID) to identify the RPL
root. The root device provides a HTTP REST interface [9]
toward the software in the gateway. This REST interface allows
the gateway software to query the root about devices in the
network that the root can reach. Reachability information is
thus generated by the gateway. Reachability in this sense is
defined as the existence of a RPL route between the device
and the gateway.
When a device joins the network, it first selects a random
root device, i.e. a random gateway. The gateway detects the
existence of this device by periodically querying the root. Each
gateway advertises the connected devices over a custom OSPF
link state advertisement to other gateways. Consequently, every
gateway learns about all devices in the network.
Each time there is a change in reachability, policy or
constraint information obtained via OSPF, the gateway selec-
tion algorithm is restarted. As OSPF quickly floods changes
between gateways, all gateways use the same information and
starts the algorithm roughly at the same time. It is important
that all gateways processes the devices in the same order so
that all gateways obtain the same result. Implementing the join
order, as described in Section III-B, is difficult because of
1In the case gateway selection is needed between the domains, a hierarchical
solution combining distributed and centralized architectures can be used.
2The management server may additionally be responsible for bootstrapping,
registration, access control, configuration, and other management functions for
both devices and gateways. These are out of the scope of this paper.
Fig. 5: Gateway Selection Sequence Diagram for the centralized architecture
the need to synchronize clocks. Therefore, devices are ordered
according to their MAC address instead.
To make the device switch a gateway, the gateway to
which the device currently is connected sends a command to
the device. The command is implemented as a LWM2M like
command transported over CoAP. The command contains the
DODAGID of the target gateway. Additionally, the command
can include a list of alternative gateways to be used in case
the target gateway becomes unavailable.
B. Implementation of Centralized Architecture
The selection of the gateway in the centralized architecture
is controlled by a central node: the gateway selection server.
Implementation-wise, the gateway selection server is either
integrated as a part of the management server or implemented
as a separate component connected to the management server.
In our implementation, the gateway selection server is part of
the management server. Additionally, the management server
is connected to a management interface, which provides a
dashboard for managing the IoT network. The management
interface graphically presents the gateways devices connected
to it and allows editing the policy.
The main function of the gateway selection server is to
collect the reachability, constraints and policy information as
input and generate the target allocation of devices to gateways.
Our server exposes a HTTP REST interface for the gateways
to post updated constraint information. Constraints can also
be edited from the management interface, which is useful for
testing and demos. The policy is locally stored on the gateway
selection server and updated by the management interface
through a set of HTTP REST commands. Both policies and
constraints are described in JSON format.
In the centralized implementation, two types of devices
are supported: IEEE 802.15.4 based devices with RPL routing
and IEEE 802.11b/g based devices. Depending on the device
type, the reachability information can be obtained either from
gateways or from the devices.
The first case is utilized on TMote Sky devices operating
over IEEE 802.15.4 radio. The devices run RPL [12] to create
routes toward the RPL root, and the root provides a REST
interface for the gateway to obtain the list of the neighbours
reachable from the root. The gateway posts changes in the
reachability information to the gateway selection server.
The second case is used in Libelium Wasmote devices [15]
with IEEE 802.11b/g [5] wireless capabilities. Libelium Wasp-
mote devices provide a device centric view of the reachability.
The Waspmote performs a scan over all channels to find the
SSIDs of the surrounding gateways. This scan is initiated by
a command from the server. The reachability information is
posted via the gateway to the gateway selection server.
Our gateways support several interfaces per gateway. Thus,
a gateway can have both IEEE 802.15.4 and IEEE 802.11b/g
interfaces, and even several interfaces of the same type. For
IEEE 802.15.4, the interface is identified by the Destination-
Oriented Directed Acyclic Graph Identifier (DODAGID), and
in IEEE 802.11b/g the interface is identified by a Service
Set Identifier (SSID). Consequently, reachability and gateway
selection commands contain these interface identifiers.
After all the input information is gathered, the gateway
selection server finds out the best gateway interface for every
device and sends the interface identifier information to the
affected devices. Messages are forwarded via the gateway in
which the device currently is connected to. In our implemen-
tation, different protocols are needed to send the command to
different device types. The TMote Sky supports CoAP, and the
implementation works like in the distributed implementation.
Since the Waspmote do not support IPv6 and CoAP, we
implemented a HTTP REST interface to the device instead.
In a final implementation, without the above restrictions, the
Lightweight Machine to Machine (LWM2M) [14] protocol
would be used instead.
The management of the devices from the gateway selection
server could be implemented with transparent gateways, allow-
ing management commands to be sent to devices directly from
the management server to the device. However, in a practical
implementation, a few issues suggest involving the gateways
as a proxy on the management path. Firstly, the reachability
information within the IEEE 802.15.4 wireless network needs
to be collected from the gateway perspective. That implies the
gateway needs to act as a LWM2M client as well. Secondly,
in the case of NAT [16] between the Management Server and
the devices, a proxy in the gateway simplifies the tunnelling
of inbound commands over an outbound connection. Thirdly,
Libelium Wasmote devices [15] do not support CoAP [13]. The
LWM2M protocol is transported on top of CoAP. That means
that a conversion between CoAP and HTML was needed and
the gateways are the best candidates due to their processing
power. All these reasons made us decide to implement a
management proxy in each gateway acting as a back-to-back
server and client.
Figure 5 shows the interaction of the different elements
involved in sharing the reachability information between the
Libelium Waspmote devices and the Management Server com-
ponent. First of all, the Libelium Waspmote device automati-
cally connects to a network. Once the device has connected, it
will send a registration message (1) to a gateway in which its
address is predefined in the sensor’s EEPROM. That informa-
tion will contain the device’s MAC address and the SSID of
the network that the sensor is connected to. After the sensor
is registered, the gateway will ask (3) the sensor to scan all
the available access points (AP) around it. Then, the Libelium
Waspmote will perform an active probe scan of access points
in all the 13 channels and returns (4) to the gateway the MAC
address, signal strength, SSID name, and security mode of the
found access points. The scan command output [17] format is:
Channel RSSI Security WPA Conf WPS MAC Address SSID
The gateways will forward that information to the man-
agement server. The management server will read the list of
reachable gateways and send a command (6,7) to the device.
That command will include in the payload a list of target
gateways in priority order to allow for backups in case of
failure of the target gateway. The sensor will try to connect
to the first option or, subsequently, to the consecutive options
if that fails. Once the device changes gateway, the registration
message (9,10) is triggered again.
V. EVAL UATIO N
In this section, we evaluate the three architecture options
and provide some insights that help to clarify our preference
for the centralized solution.
Figure 6 shows the message paths for different types of
operations in the three architectures. The figure focuses on
messages over the two wireless interfaces: the mobile backhaul
and the capillary network. The figure assumes that Device1
and Device2are reachable from Gateway2while Gateway1
cannot reach any device.
Policies are distributed infrequently as they are rather static.
In the centralized architecture the new policy is not transported
over any wireless interface. In the distributed architecture, the
updated policy must be sent to each gateway separately. In the
autonomous architecture, the new policy must be sent to all
gateways, which forward it to all reachable devices.
In the centralized architecture, it is sufficient to transport
the updated constraints of a gateway to the management server.
In the distributed architecture, the constraints must be sent
to all gateways over the mobile network. The transmission
between a pair of gateways requires one message in the uplink
direction and one in the downlink direction. In the autonomous
architecture, the gateway sends its updated constraint to all
reachable devices.
Reachability updates are frequent if devices are mobile.
Updated reachability information generated by a device must
traverse both the capillary and the mobile networks to reach
the management server in the centralized architecture. In the
distributed architecture, the message must be flooded between
gateways, each transmission passing over the mobile network
twice. In the autonomous architecture, the reachability infor-
mation is only used locally by the device.
Each time a parameter changes, the target gateway of one
or more devices may change. In the centralized architecture,
a command to switch gateway is sent over the two wireless
interfaces to the affected devices. In the distributed archi-
tecture, each gateway locally performs the gateway selection
Fig. 6: Messages used in the evaluation
algorithm. Messages are needed to be sent only to the affected
devices over the capillary network interface. In the autonomous
architecture, the device performs the calculation locally and
switches gateway without any messaging.
Distributed Centralized Autonomous
Cell Cap Cell Cap Cell Cap
Policy N0 0 0 N M
Constraints 2N(N−1) 0 1 0 0 D
Reachability 2N(N−1) 1 1 1 0 0
GW Selection 0 1 1 1 0 0
TABLE I: Number of messages per operation
Table I shows an analysis of the traffic required for each of
the above operations. To make the comparison more generic,
we count the number of HTTP REST requests instead of
the number of bytes or messages. Thus, we do not consider
message size, connection setup, acknowledgement, fragmen-
tation and packet loss. We separately consider messages in
the cellular uplink (abbreviated as Cell) and the capillary
network (abbreviated Cap). In the distributed solution, we
assume that the gateways communicate with each other via
the cellular network, so that each request traverses the cellular
network twice. We denote the number of gateways with N
and the number of devices with M. The average number
of devices within reach of a gateway is denoted D, which
typically is D∝M/N. For flooding in OSPF, we assume
full connectivity, thus a message from one gateway is sent to
N−1other gateways. Sending a message to all devices under
a gateway is assumed to be implemented with unicast using
HTTP REST, rather than broadcast.
Based on Table I, we can analyse the total traffic generated
for different scenarios. The traffic is dependent on the frequen-
cies of updates in the gateway selection parameters. We denote
the frequency of policy updates with fpol, the frequency of
constraint changes per gateway with fconst, and the frequency
of reachability updates per device with freach. The frequency
of gateway selection commands, fsel, is proportional to the
total number of updates fpol +Mfreach +N fconst as each
update may, with a given probability, cause a device to be
allocated to another gateway. In our analysis, a selection
command is sent once per 10 updates, but the choice of value
turned out to have only a minor effect on traffic. Ddenotes
the average number of devices within reach of a gateway.
We assume every devices can be reach from three different
gateways. The default values of the parameters are given in
Table II.
Parameter Default value
fpol 1 update / 24 h
fconst 1 update / 1 h
freach 1 update / 10 min
fsel (fpol +Mfreach +N fconst )/10
D3M/N
TABLE II: Default values of parameters in the analysis
Figure 7 shows the total number of requests in the network
for a network with 10000 devices and 1000 gateways. Figure
8 shows the same scenario for a network, where the same
10000 devices are distributed between 100 gateways. In both
figures, the frequency of reachability updates freach varies
between one update per 1 minute and one update per 24 hours.
This correspond to different degree of device mobility. We see
that the distributed architecture generates several magnitudes
higher traffic than the other architectures. The centralized
architecture generates least traffic, except when mobility is
high, when the autonomous architecture performs better. In
Figure 9 the frequency of constraint updates freach varies
between one update per 1 minute and one update per 24 hours.
This mainly affects the traffic in the autonomous architecture.
We noticed that varying the other parameters only has a minor
effect on the traffic. The last table, Table III, evaluates the
load on the management server. In all evaluated cases, the
centralized architecture showed the highest load on the server.
1"
10"
100"
1,000"
10,000"
100,000"
1,000,000"
10,000,000"
100,000,000"
1,000,000,000"
Distributed"
Centralized"
Autonomous"
Requests"per"minute"
1"min"
10"min"
1"h"
24"h"
Fig. 7: Requests for various reachability update frequencies, 100 gateways
1"
10"
100"
1,000"
10,000"
100,000"
1,000,000"
10,000,000"
100,000,000"
1,000,000,000"
10,000,000,000"
100,000,000,000"
Distributed"
Centralized"
Autonomous"
Requests"per"minute"
1"min"
10"min"
1"h"
24"h"
Fig. 8: Requests for various reachability update frequencies, 1000 gateways
1"
10"
100"
1,000"
10,000"
100,000"
1,000,000"
10,000,000"
100,000,000"
1,000,000,000"
10,000,000,000"
Distributed"
Centralized"
Autonomous"
Requests"per"minute"
1"min"
10"min"
1"h"
24"h"
Fig. 9: Requests for various constraint update frequencies, 1000 gateways
VI. CONCLUSION
In this paper, we addressed the problem of gateway selec-
tion in a capillary network. Without any mechanism, a device
typically selects the gateways randomly, without considering
the uplink capacity or the properties of the gateways. Gateway
selection allows devices to make informed selection in order to
globally optimize the network. The paper presents an overview
of different gateway selection solutions for capillary networks,
some of which we implemented. The solutions mainly differ in
the location where the selection decision is made. According
to our analysis, a distributed architecture creates a high amount
of traffic in the cellular network. The centralized architecture
has low traffic both in the capillary and the cellular network but
the load on the management server is high. The autonomous
devices make the network simple with low traffic and low load
on the server, but the device itself has a high complexity as it
Load on Server
Distributed Nfpol
Centralized Nfconst +Mfreach +M fsel
Autonomous Nfpol
TABLE III: Load on the Management Server
must perform the gateway selection. This complexity is often
too demanding for the current sensors. Our implementation
utilizes flexible policies, which takes into account multiple
factors to select the optimal gateway. With automatic gateway
selection, load can be distributed between gateways, while con-
sidering gateway properties, such as current load. This allows
a better utilization of the network resources and optimization
of the end-to-end path through both networks.
ACKNOWLEDGMENT
The authors would like to thank Tero Kauppinen, Miika
Komu and Mert Ocak for their contribution in the implemen-
tation of the prototypes and Petri Jokela for his feedback.
REFERENCES
[1] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of
things (iot): A vision, architectural elements, and future directions,”
Future Gener. Comput. Syst., vol. 29, no. 7, pp. 1645–1660, Sep. 2013.
[Online]. Available: http://dx.doi.org/10.1016/j.future.2013.01.010
[2] O. Mazhelis, H. Warma, S. Leminen, P. Ahokangas, P. Pussinen,
M. Rajahonka, R. Siuruainen, H. Okkonen, A. Shveykovskiy, and
J. Myllykoski, “IoT SOTA report 2013 - Internet-of-Things Market,
Value Networks, and Business Models: State of the Art Report,” Jan.
2013. [Online]. Available: http://www.internetofthings.fi/
[3] “Bluetooth standards.” [Online]. Available: https://www.bluetooth.org/
en-us/specification/adopted- specifications
[4] “IEEE 802.15 standard.” [Online]. Available: http://www.ieee802.org/15
[5] “IEEE 802.11 standard.” [Online]. Available: http://www.ieee802.org/
11/
[6] S. Singh and K.-L. Huang, “A robust m2m gateway for effective
integration of capillary and 3gpp networks,” ser. Advanced Networks
and Telecommunication Systems (ANTS), S. Singh and K.-L. Huang,
Eds. 2011 IEEE 5th International Conference, 2011, pp. 1–3.
[7] V. Misic, J. Misic, X. Lin, and D. Nerandzic, “Capillary machine-to-
machine communications: The road ahead,” in Ad-hoc, Mobile, and
Wireless Networks, ser. Lecture Notes in Computer Science, X.-Y. Li,
S. Papavassiliou, and S. Ruehrup, Eds. Springer Berlin Heidelberg,
2012, vol. 7363, pp. 413–423.
[8] RFC 5340; OSPF for IPv6, IETF.
[9] L. Richardson and S. Ruby, RESTful Web Services. O’Reilly Media,
2008.
[10] “TMote Sky Datasheet.” [Online]. Available: http://www.eecs.harvard.
edu/∼konrad/projects/shimmer/references/tmote-sky-datasheet.pdf
[11] RFC 6282; Compression Format for IPv6 Datagrams over IEEE
802.15.4-Based Networks, IETF.
[12] RFC 6550; RPL: IPv6 Routing Protocol for Low-Power and Lossy
Networks, IETF.
[13] RFC 7252; The Constrained Application Protocol (CoAP), IETF.
[14] Lightweight Machine to Machine, Technical Specification, Open Mobile
Alliance (OMA), 2014.
[15] “Libelium webpage.” [Online]. Available: http://www.libelium.com/
[16] RFC 2663; IP Network Address Translator (NAT) Terminology and
Considerations, IETF.
[17] “Microchip, Wifly Command Reference.” [Online]. Available: http:
//ww1.microchip.com/downloads/en/DeviceDoc/50002230A.pdf
