Professional Documents
Culture Documents
Abstract—Recently, mobile devices have been introduced in various networks (LLNs). Protocols, such as Collection Tree Protocol
wireless sensor network (WSN) applications in order to solve complex (CTP) [11] and Routing Protocol for Low-Power and Lossy
tasks or to increase the data collection efficiency. However, the Networks (RPL) [30], allow a system developer to simply choose
current generation of low-power WSN protocols is mainly designed to
support data collection and address appli-cation-specific challenges a network layer protocol that suits the target application well rather
without any particular considerations for mobility. In this paper, we than designing a completely new routing protocol.
introduce MoMoRo, a mobility support layer that can be easily However, in these new CPS and IoT applications, mobil-ity
applied to existing data collec-tion protocols, thereby enabling and actuation becomes an integral part of the low-power
mobility support in the network. MoMoRo robustly collects
wireless network and cause radically different network traffic
neighborhood information and uses a fuzzy estimator to make link
quality estimations. This fuzzy estimator continuously reconfigures its patterns than those found in traditional data collection appli-
thresholds for determining the fuzzy sets, allowing MoMoRo to easily cations. Hence, protocols specifically designed for stationary
adapt to changing channel environments. Furthermore, MoMoRo networks are ill equipped to tackle the low-power data collec-
includes an active destination search scheme that allows disconnected tion challenges found in hybrid networks with both stationary
mobile nodes with sparse traffic to quickly reconnect if there are and mobile nodes. Furthermore, they cannot provide adequate
packets in the network destined to this mobile node. We evaluate
MoMoRo both indoor and outdoor and show that a continuously performance in these applications; therefore, a new approach is
moving device in a MoMoRo-enabled RPL (i.e., IPv6 Routing Protocol needed to address the mobility issue in these networks.
for Low-Power and Lossy Networks) network can achieve a high Previous efforts to support mobility in LLNs show that a
packet reception ratio of up to 96% and stay connected in areas where multiple-tier network architecture that separates the role of data
RPL alone cannot and with less than half the packet overhead needed forwarding from the mobile nodes is effective in achieving high
by the well-known Ad hoc On-Demand Distance Vector routing
protocol.
reliability, but such an architecture requires a large set of nodes
to form the backbone network and require careful consideration
in the deployment phase [7], [16].
Index Terms—Fuzzy estimation algorithm, Internet Engineer-
ing Task Force Routing Protocol for Low-Power and Lossy Net- While this two-tier architecture captures the essence of a hybrid
works (RPL), node mobility support. network, we seek a solution with less rigid division between data
forwarding and mobile nodes, thereby reduc-ing the constraints
I. INTRODUCTION imposed on the underlying data collection backbone. Furthermore,
we target low-power wireless network applications in which 1)
W ITH the emerging cyber physical systems (CPS) and Internet of
Things (IoT), low-power wireless networks are now taking their next step
human-scale mobility exists, 2) only low amount of traffic is
transmitted in the network, and 3) data collection is the main traffic
toward widespread usage in during the majority of the time, but point-to-point traffic is used to
various everyday life applications. Examples of such applica-tions deliver urgent alerts or configuration/actuation messages to
include wireless physiological data sensing for clinical care, individual nodes. Given the maturity of the available low-power
military-related sensing applications, and robotic network systems wireless network proto-cols and that data collection is still a
for various automation and sensing tasks [4], [5], [20], cornerstone CPS and IoT application, we draw from previous
[17]. A decade worth of research by the wireless sensor net- research efforts into data collection protocols and leverage the
work (WSN) community provides the foundation for forming routing capabilities these protocols provide. Specifically, we
such wireless networks with proven network layer protocols introduce a mobility support layer, called MoMoRo, to ensure that
that addresses the unique challenges in low-power and lossy high packet delivery rates can be achieved even with mobile nodes
in the network. At the same time, MoMoRo does not introduce
Manuscript received August 10, 2012; revised September 5, 2013 and significant performance overhead when all the nodes are
November 5, 2013; accepted December 20, 2013. This work was supported by stationary. Our previous work shows that the reason behind this
the Korea Evaluation Institute of Industrial Technology and the Korean
performance degradation with current generation collection
Ministry of Science, ICT and Future Planning under Project 10041725 (De-
velopment of application service and software to support sensor terminals for protocols is the rapid change in the wireless environment caused
providing personalized service based on smart devices). by nodes moving into areas with different wireless characteristics.
J. Ko is with Electronics and Telecommunications Research Institute,
Daejeon 305-350, Korea (e-mail: jeonggil.ko@etri.re.kr).
In other words, protocols for stationary networks are designed to
M. Chang is with the Department of Computer Science, Johns Hopkins compensate for temporal changes in the wireless environment,
University, Baltimore, MD 21218 USA. whereas the spatial changes caused by mobility happen too rapidly
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
for these protocols to adapt in time [16].
Digital Object Identifier 10.1109/JSYST.2014.2299592
Hence, in order to effectively support device mobility in LLNs, A. Mobility in LLNs
the mobility support layer must be able to provide robust link Numerous efforts exist to provide a routing protocol with
quality estimations for the changing neighborhood. In addi-tion, high packet reception ratio (PRR) for WSNs. One of the most
since bidirectional communication is an important feature in CPS successful efforts is the CTP [11]. As the default data collection
and IoT applications, the mobility support layer must also be able protocol in TinyOS 2.x, CTP provides high PRR routing paths
to assist the underlying routing protocol to quickly reestablish its with minimal end-to-end costs by provisioning routes based on
bidirectional routes. MoMoRo accomplishes this by actively the four-bit link estimator [10]. While CTP performs well for
transmitting probes once it detects that the node has been stationary node data collection, it lacks the support for routing
disconnected from its current parent in the data collection tree. to individual nodes in the network and was not designed to
Link quality estimators, such as the expected number of provide any form of mobility support. More recently, Hydro
transmissions (ETX), received signal strength (RSS), and symbol and RPL were introduced to provide IPv6 routing to LLN
error rate (SER) variance, collected from the probes are analyzed systems. These protocols provide support for data collection
using a fuzzy logic estimator to determine good quality links in the traffic as well as point-to-point traffic [9], [30]. However, like
new neighborhood. Using a fuzzy estimator allows us to CTP, Hydro and RPL were not designed to support mobile
compensate for the uncertainties in the wireless channel conditions nodes. Further mobility-related work in sensor networks either
while effectively classifying the wireless links’ quality. only support sink node mobility [21], require local or global
Furthermore, to allow the fuzzy estimator to easily adjust to time slots [3], or require the use of global location infor-mation
varying environment conditions, we use an adaptive scheme to [1].
continuously reconfigure the thresholds used to determine the WSN systems such as MEDiSN [16], CodeBlue [22], and the
fuzzy sets. Links that are estimated to have good quality by the Washington University physiological data collection system [7]
fuzzy estimator are suggested to the underlying routing protocol so have proposed some customized solutions to support mobile
that it can select a route that optimizes its respective end-to-end devices in a WSN. MEDiSN and the Washington University
metric. Additionally, MoMoRo includes an active destination physiological data collection system form a two-tier network by
searching scheme that initiates adaptively scoped floods to search simplifying the routing capabilities of the mobile devices and
for “missing” destinations. This scheme is activated when a mobile placing a designated backbone routing tree network of stationary
node has not yet realized that it has been disconnected from the devices for data forwarding. While such an approach simplifies the
previous route; thus, a route reconfiguration has not been initiated issue of maintaining continuous connectivity to mobile devices,
at the new neighborhood, and the node is missing from its previous one drawback is that, in order to increase the connectivity range,
neighbors. the backbone needs to be extended, even when some locations of
Our contribution in this paper is the mobility support layer the coverage area may not have any traffic to forward (e.g., an area
MoMoRo, which we evaluate using a series of experiments with only infrequent node visits needs the same backbone
with stationary and mobile Tmote Sky devices [23] installed in infrastructure as a well-visited area). In addition, to ensure high-
an indoor testbed and validate the performance in an outdoor quality links, the deployment of the backbone network requires
environment as well. Our results show that MoMoRo supports careful provisioning [7], [16]. The CodeBlue system incorporates
the underlying collection protocols well when human-scale mo- a fully ad hoc routing protocol and can theoretically solve this issue
bility is introduced, with high packet reception performance and [6]. However, the performance of CodeBlue quickly degrades with
low additional packet overhead. We also present performance increasing hop counts, large number of devices, and user mobility
comparisons with the Ad hoc On-Demand Distance Vector [22]. Our approach is an effort to combine these two approaches
(AODV) routing protocol, which is specifically designed for and tries to maintain a routing tree, which includes both stationary
mobile networks. Our results indicate that MoMoRo achieves and mobile devices, while effectively supporting connectivity for
up to 96% packet delivery performance with less than half of collection and point-to-point traffic with mobile nodes on top of
the overhead introduced by an AODV implementation designed this base routing architecture.
specifically for LLNs.
The structure of this paper is as follows. Section II introduces
related work in three different related research domains. Next, in
Section III, we present MoMoRo as a mobility support layer for B. MANETs
low-power wireless networks and evaluate its performance in The well-known AODV routing algorithm is a reactive rout-ing
Section IV. Finally, we conclude this paper with a summary and a protocol designed for mobile ad hoc networks (MANETs)
list of future research direction in Section V. [24]. While AODV adapts to networks with frequent link
changes, the high overhead and delay from the route discovery and
maintenance makes it less suitable for resource-limited networks
II. RELATED WORK
such as LLNs. Nevertheless, implementations such as TinyAODV
Research on supporting mobile devices in wireless net-works and Not So Tiny (NST)-AODV represent efforts to bring AODV
has been widely studied in various wireless technique domains. to LLN applications [13], [27]. We will use NST-AODV, an
We use this section to present an overview of the existing improvement to TinyAODV, in our evaluations to show how an
literature in various domains and position our work AODV implementation for LLNs performs with a typical data
accordingly. collection application and human mobility.
Dynamic source routing (DSR) is another on-demand routing
protocol for MANETs that uses source routing for packet
forwarding [15]. By using source routing for all the packets, it
introduces a relatively high overhead. In addition, because a node
must have knowledge of the entire route before attaching the
source routing header, and the fact that DSR does not have a local
route repair mechanism for broken links, the overhead of
maintaining the routes can become very expensive. Finally, in
DSR, path maintenance (e.g., new route requests) occurs when an
already transmitted packet encounters a broken link; thus, this
recovery phase can introduce a large latency. Fig. 1. Diagram of how MoMoRo interacts with the network stack.
Furthermore, most MANET routing protocols aim to estab-
lish end-to-end paths for data transmissions [2]. In other words, this information through a fuzzy estimator to identify links that are
when a single link on the path breaks (e.g., due to mobility or currently good and usable to hint the routing layer in making its
link fluctuations), the entire end-to-end path is reconstructed, route selections. The remainder of this section provides detailed
increasing routing overhead and latency. Similar overheads are descriptions of the MoMoRo mobility support layer.
also encountered when using proactive routing schemes such as
Destination-Sequenced Distance Vector routing [25]. On the A. Mobility Support Layer in the Networking Stack
other hand, our scheme’s basic routing paths are based on the
paths selected by a tree-based routing protocol, and we focus As Fig. 1 shows, the operation of MoMoRo is implemented
on maintaining these paths using the assistance from our in the networking stack as an additional layer between the link
mobility support layer to perform local routing path repairs layer and the routing layer. We note that only mobility/routing
with minimal overhead. configuration-related information/packets are exchanged be-
Finally, we observe that mobile LLNs carry different traffic tween the three layers. All other packets, such as typical data
types than typical MANETs. Typically, in MANET research, packets, do not pass through the MoMoRo layer and thus do
the main focus of the routing protocols is to establish any-to- not add any additional processing overhead. The main benefit
any routing paths; whereas, for mobile LLNs, collection is the of implementing MoMoRo as an independent layer rather than
main traffic pattern, and point-to-point traffic support is an a software module within the routing layer is that it potentially
application-specific requirement and deals with a significantly provides the flexibility to easily apply and reuse the MoMoRo
smaller number of packets [30]. It is therefore reasonable for code with different routing protocols rather than reimplement-
mobile LLNs to optimize their routes for the collection traffic ing the procedures when the routing layer changes.
and use the already established routes to support additional
traffic patterns.
B. Realizing a Link Disconnection
C. Mobile Device Handoff Typically, a routing layer collects information on the number of
transmission attempts for each unicast packet to perform link
Many wireless research domains propose device handoff quality estimations (see Fig. 1). Once a node realizes that a packet
schemes to provide seamless connectivity on mobile devices transmission to node n has failed (i.e., the number of
(see [19], [28], and references in [26]). However, we argue that retransmissions reaches the maximum), the routing layer notifies
for LLNs, such techniques are too aggressive. Given that the MoMoRo that the link with node n may be potentially lost. At this
primary consideration in designing an LLN system is to point, MoMoRo makes one last attempt to reach n by transmitting
optimize for the limited resources available on the devices (e.g., a unicast packet. This attempt allows the node to suppress any
bandwidth, energy, etc.), wasting packets to perform detailed additional overhead due to route reconstruction if the reason
analysis of whether a link would break or not or to select the behind the failed transmission is due to instantaneous link quality
best possible link at any point in time [8], [14] is a luxury. fluctuations. If an acknowledgment is received, no further action
Rather, the characteristics of LLNs suggest that we utilize a is made. However, if the acknowledgment is not received in
scheme to quickly and accurately recover from link losses and
timetinit_probe, MoMoRo first determines if n is the default route.
find new routes. In other words, a reactive scheme is more
If this is true, an operation to reconstruct the default route takes
suitable for LLNs than a proactive scheme in which we try to
place (see Section III-C); otherwise, MoMoRo will start the search
predict future link disconnections.
for the packet’s final destination since the route to this missing
destination node is broken (see Section III-D).
III. MoMoRo: THE MOBILITY SUPPORT LAYER
MoMoRo supports node mobility by performing two major
operations. First, it tries to react quickly to packet losses to either
C. Reconstructing Default Routes
reconstruct the default route (e.g., route toward the root of the
routing tree) or notify a nonroot destination that its route is Once a nodeA determines that the default route is lost and,
currently unavailable. Second, when reconstructing the default therefore, the data collection tree needs to be repaired, Mo-
route, it quickly gathers neighborhood information and passes MoRo quickly broadcasts countbroadcast messages to its (new)
wireless neighborhood. Nodes that hear this broadcast will send a
unicast reply. Using the received replies, node A computes the
ETX of the link, the average RSS on the link, and the link’s SER
variations. We explain the three metrics in detail below.
1) ETX: ETX deals with the link quality in an experimental way
and represents the expected number of transmission attempts
needed to make a single successful packet trans-mission
[30]. In wireless networking, ETX is considered as one of
Fig. 2. Diagram of our fuzzy membership function for determining the fuzzy
the most accurate link quality estimation metrics 2
sets ofETX, RSS, and σSER . The values of the Ax variables are application
for various routing protocols [30], [11]. In MoMoRo, we and channel environment dependent or can be adaptively set.
compute the ETX using the countrx replies that we
receive from each link as follows: given that it provides a structured environment for multiple input
count variables to result in a single output. In addition, since it is difficult
ETX = broadcast
to extract the meaning of a metric using static thresholds in a
countrx dynamic channel environment, by using a fuzzy-logic-based
where countrx is the actual number of probe packets that estimator, we can take the advantage that the input metric does not
are received. need to be exact but rather can cope with some unexpected
2) Average RSS: The average RSS of the received packets, variations using overlapping regions in defining the fuzzy sets.
i.e., RSS, can provide us with a rough estimate of the Specifically, each metric value is passed through a fuzzy
expected connection quality. We compute RSS only membership function, where the fuzzy set is determined for each
using countrx packets. metric (e.g., input metric value is x% “good,” y% “medium,” and
3) SER Variance: Previous work shows that the SER vari- z% “bad”). By summing up the result of the fuzzy membership
ance increases noticeably as the distance of the link function for each metric, we can predict the expected quality of a
increases to a point where potential link disconnections link in various dimensions. The output of the fuzzy estimator will
may occur [18]. On both the popular TI CC2420 and indicate if the link that is being diagnosed is expected to be either
CC2520 radios, the link quality indicator (LQI) values of “good,” “ok,” or “bad.”
probe packets can provide us with a rough estimate of the The main challenge in defining an effective fuzzy estimator is
SER encountered as a packet travels a wireless link. Since properly defining the fuzzy sets and the membership functions. For
all IEEE 802.15.4 radios are required to report an LQI this, we first start by defining a simplified and general diagram of
value, this metric is widely available on WSN platforms. a fuzzy estimator, as shown in Fig. 2. In other words, defining the
We compute the variance of the SER (LQI) values using variables A0, A1, A2, and A3 for each input metric (e.g., ETX,
the following equation: RSS, and σSER2) becomes a critical issue since configuring
inefficient set boundaries will naturally lead to improper link
σSER2 = SER2 − SER2.
quality detection. Using the base diagram in Fig. 2, the rest of this
We point out that the metrics introduced above have their own section discusses about how we configure the fuzzy sets for each
benefits and drawbacks. Specifically, while the ETX metric is an input metric that we target.
accurate measure of link quality, it takes multiple packets to 1) ETX: Among the input metrics for our fuzzy estimator,
achieve reasonable accuracy. Furthermore, the RSS metric can be the fuzzy sets for the ETX metric can be easily config-
collected quickly but can show high variance with changes in the ured to meet application-specific targets. For example, an
channel conditions, and lastly, σSER2 presents an accurate application may consider links with a 90% PRR as good
evaluation of a link condition with minimal latency but typically and therefore set the threshold for ETX’s fuzzy
shows the link quality as a binary metric (e.g., good or bad) [18]. membership function as 1.11 Tx. Furthermore, if 50% is
To compensate for these benefits and drawbacks, we need a way to the threshold for determining bad links, this can be
intelligently combine the metrics so that we can exploit the benefits translated as 2.00 ETX and used as the threshold for
of each metric. Furthermore, for each metric, we should come up ETX’s fuzzy membership function. These thresholds can
with a single answer to the following question: “Is this neighbor be set to tolerate for some unexpected variations (e.g.,
good enough to connect to?” The simplest and possibly the most 5%), thereby determining A0 = 1.05, A1 = 1.18, A2 =
accurate answer to this question could be obtained by normalizing 1.82, and A3 = 2.22.
each metric, identifying the relationships between various metrics 2) σSER2: For the fuzzy membership function of σSER2, our
and then using an optimization algorithm to end up with a measurements and previous work suggest that 0 < σSER2
probability of potential link (dis)connection. However, in a <∼ 3 indicates a good link (e.g., link with a short distance
dynamic channel environment, it is difficult to get an accurate and a high expected PRR), ∼ 3 < σSER2 <∼ 50 indicates
relationship func-tion over different link metrics, and the resulting a link that works well with some fluctuations, and σSER2
probability would require another threshold for the algorithm to >∼ 50 indicates a link with bad quality [18]. Again, since
output a firm decision (i.e., yes or no). we are defining fuzzy thresholds, we can further define
safety bounds (i.e., ranges of A1−A0 and A3−A2) for the
As an alternative, we consider the use of a fuzzy-logic-based σSER2 metric to account for realistic variations and
link quality estimator. Using fuzzy logic is beneficial uncertainties. As a result, for σSER2, one
bar for selecting a good link and try to assure that links that fall in
the “good” category have better quality links than before.
Upon evaluating the potential links in the fuzzy estimator for
RSS, if the number of links that are classified as good is too small
(e.g., < countgood_limit), MoMoRo will try to decrease the values
of A0 and A1 by 5 dBm and recursively run the fuzzy estimator
with the same links once more. This processes repeats until more
than countgood_limit are classified as good or until A1 reaches A2.
This decrease in thresholds is to assure that MoMoRo can supply
the routing layer with enough links to compare from and to make
sure that the increase in A0 and A1 values from the previous case
does not filter out all of the potential links by setting a bar that is
too high. In practice, and in our evaluations, countgood_limit is set
to 1.
Given that RSS is a metric that is heavily dependent on the
channel conditions, this adaptiveness of the fuzzy estimator
allows MoMoRo to be easily deployed in various environments
in which the wireless channel dynamics are difficult to model
or even unknown. We will show in our evaluations that using
this adaptiveness helps reduce the amount of overhead needed
to identify good links.
Fig. 7. PRR of the collection and downward traffic from/to the mobile node for
MoMoRo, RPL, and AODV.
Fig. 8. Amount of overhead introduced during the period when one node was
mobile in the network.
to react to such unexpected disconnections. On the other hand, Fig. 9. (Top) Collection and (bottom) downward PRR for a mobile device over
both MoMoRo and AODV start taking more aggressive actions, time using 30-s buckets. The first and last two buckets show the PRR before
and after the mobile phase, whereas the five buckets in the middle show the
leading to a larger overhead but also a high PRR, with the PRR of MoMoRo, RPL, and AODV during the mobility phase.
overhead from the AODV network almost twice as large as the
MoMoRo network. point PRR (bottom) for a total of 4.5 min (i.e., 1 min before the
We note that our experiments were conducted at various times mobility occurred, 2.5 min of the mobility period, and 1 min after
during the day. Specifically, for each protocol, we run five the mobility ceased) split into 30-s bins. We notice that, while
experiments, evenly spaced, during 10 A.M. to 1 A.M. As a result, MoMoRo and AODV show a continuous high PRR over the entire
despite the fact that our experiments were done using a single period, RPL’s PRR performance drops as mobility is introduced.
testbed, the temporal variations of the channel environment over In fact, during the fifth 30-s period, the mobile node was
the day introduce a new channel environment for the networks to completely disconnected from the network (resulting in a PRR of
operate in. The small variations we see with MoMoRo suggest that 0%). Even during this period, we see in Fig. 10, where we plot the
it can adapt well to various channel conditions when compared amount of overhead packets in the entire network during the same
with AODV or RPL. time frame, that RPL did not increase its amount of overhead
In order to examine the packet loss distribution over time, we packets in order to discover new routes. On the other hand, when
select a single trace from each configuration for comparison taken losses occur, both MoMoRo and AODV start generating overhead
at nighttime when the wireless channel was stable. Fig. 9 presents packets to reconfigure the routes, and as a result, both achieve high
the mobile node’s collection PRR (top) and point-to- PRR. In this case, AODV
Fig. 10. Number of overhead packets in the network over time using 30-s
intervals. The first and last two cases show the overhead packets generated
before and after the mobile phase, whereas the five cases in the middle show
the overhead packets generated during the mobile phase.