Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 745 (2001)

(USC DC Other)

USC Computer Science Technical Reports, no. 745 (2001)

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Residual Energy Scans for Monitoring Wireless Sensor Networks
Yonggang “Jerry” Zhao, Ramesh Govindan
Computer Science Department/ISI
University of Southern California
Los Angeles, CA 90089
{zhaoy,govindan}@isi.edu
Deborah Estrin
Computer Science Department
University of California, Los Angeles
Los Angeles, CA 90095
estrin@isi.edu
Abstract
It is important to have continuously updated information about network resources and application activities after a
wireless sensor network is deployed in unpredictable environment. Such information can help notify users of resource
depletion or abnormal activities. However, the low user-to-node ratio and limited energy and bandwidth resources in
sensor networks make extracting states from each individual node infeasible. In this paper, we propose an approach to
construct abstracted scans of sensor network health by applying in-network aggregation of network states. Specifically,
we design a residual energy scan that approximately depicts the remaining energy distribution within a sensor network.
Simulations show that our approach has good scalability and energy-efficiency characteristics, compared to continuously
extracting the residual energy individually from each node.
1 Introduction
Wireless sensor networks have been attracting increasing research interest given the recent advances in miniaturization and
low-cost, low-power design. Unlike traditional computer networks such as the Internet, such a network will consist of a
large collection of small wireless, low-power, unattended sensors and/or actuators [ADL
98, EGHK99, KKP99, EGH00].
Sensor networks can enable “smart environments” which can monitor ambient conditions such as temperature, movement,
sound, light, location and others.
Wireless sensor network technology poses its unique design challenges. One important feature that distinguishes
sensor networks from traditional distributed systems is their need for energy efficiency. Many nodes in the emerging sensor
systems will be untethered, having only finite energy reserves from a battery. The scale of a sensor net’s deployment
will make recharging these energy reserves impossible. The requirement for energy-efficiency pervades all aspects of
the system design [PK00]. Another important feature that distinguishes wireless sensor networks from other distributed
systems is their unattended nature. In these networks, nodes are not necessarily deployed in a regular way. Because of
their compact form factor and their potential low cost, it might be possible for thousand of nodes to be autonomously
deployed in an unplanned fashion. The working environment for those sensor nodes might be unpredictable and could
affect the performance of the sensor network dramatically. The high node-to-human ratio also makes it infeasible to
maintain individual node constantly.
Given their unattended nature and their complexity, it is critical that the users be given continuously updated indications
of the sensor network health, i.e., explicit knowledge of the overall state of the sensor network after deployment. We call
such indications of network health scans.
1
Sensor network scans can provide early warning of system failure, aid in incremental deployment of sensors, or help
test sensor collaboration algorithms. For example, knowing the remaining energy resource distribution within a sensor
field, a user may be able to determine if any part of the network is about to suffer system failures in the near future, due
to depleted energy. Similarly, given the practical difficulties in precisely planning sensor field deployments, network scans
can help guide incremental deployment of sensors by indicating energy-depleted regions of the sensor field. By examining
the distribution of node density, communication quality and other resources in the sensor field, additional sensors can be
placed selectively on those regions short of resources. Finally, a sensor scan can be designed to depict the overall response
of the sensors to some known stimulus in a sensor field. Such information is a valuable tool for validating expected
functionality or fine-tuning detection algorithms.
However, such continuous monitoring wireless sensor networks leads to different challenges compared to existing di-
agnosis protocols for distributed systems [BH91, REG98, Bat95, TG98, BB99], or continuous monitoring in other domains
such as telecommunication networks, or power generation systems [Boy93]. The large number of sensor nodes in a sen-
sor field makes it infeasible, given energy and communication constraints, to collect detailed state information from each
individual sensor node and then process centrally.
In this paper, we propose an efficient monitoring infrastructure for sensor networks. Analogous to weather map or air
traffic radar images, our sensor network scans describe the geographical distribution of network resources or activity of
a sensor field. We design and evaluate a mechanism for collecting a residual energy scan (eScan). Such a scan depicts
an aggregated picture of the remaining energy levels for different regions in a sensor field, and may look like Figure 1.
Instead of the detailed information of residual energy at individual sensors, this scan provides an abstracted view of energy
resource distribution.
Our proposed approach to construct an eScan applies localized algorithms in sensor networks for energy-efficient in-
network aggregation of local representations of scans. Rather than collect all local scans centrally, this technique builds
a composite scan by combining local scans piecewise. At each step of aggregation, these partial scans are auto-scaled by
varying their resolutions. In this manner, the information content of the overall scan scales well with network size. We also
propose to apply incremental updates to scans. When the state of a node changes, it should not need to re-send its entire
scan. Rather, it should only need to send an update to a scan when its local state has changed dramatically. Furthermore, that
update should only traverse up the aggregation hierarchy if it radically impacts some aspect of the overall representation.
When local scans are aggregated, detailed information such as the residual energy at each individual node is lost. However,
the compactness of such an abstracted representation can reduce the communication and processing cost significantly. As
we show in this paper, the trade-off between reduced fidelity and increased lifetime is acceptable.
To evaluate the performance of our design, we perform a simulation-based comparison to centralized collection of
individual residual energy. We evaluate our design of residual energy scans by comparing the messaging costs of scan con-
struction and the relative errors introduced by aggregation and incremental update. We show that eScanning can achieve
significant saving on messaging cost over centralized collection, but only introduces acceptable and bounded error. Fur-
thermore, the savings increases with network size. These two results show that sensor network scanning have better
energy-efficiency and scalability characteristics than centralized collection of network state. To strengthen our result, all
the experiments are repeated for three different energy dissipation models and varied parameters.
The rest of the paper is organized as follows. In Section 2, we give a summary of related work and why they are
not applicable to sensor networks. Section 3 describes the design of residual energy scan collection. In Section 4, our
preliminary simulations results show that sensor network scanning is energy-efficient and scales well with network size,
compared to collecting all residual energy information centrally. Section 5 concludes with some future work directions.
2
2 Related Work
To our knowledge, there exists no ongoing or previous work that has attempted continuous monitoring of large-scale
distributed sensor networks. In this section, we review peripherally related areas: wireless sensor networks in general,
debugging and diagnosis protocols for parallel and distributed systems, monitoring other industrial systems and recent
work on coverage problem in sensor networks.
Wireless sensor networks can potentially support a variety of high profile applications [EGH00, Ten00]. Researchers
have been addressing various aspects of the design of sensor networks. For example, a design of wireless integrated network
sensors is described in [ADL
98]. Reference [HKB99] proposes a family of adaptive protocols (SPIN) for information
dissemination in energy-constrained wireless sensor networks. An energy-efficient paradigm (directed diffusion) for the
design of protocols for sensor networks is proposed in [IGE00]. Two important features of directed diffusion are data
aggregation and localized interaction. Combining data from several sensors, data aggregation can reduce data overlapping
and produce more accurate data. The control decisions are made based solely on the interactions with neighbors or nodes
with some vicinity. Localized interactions try to avoid the high cost to deliver data over long distance as much as possible
but still be scalable, robust and energy-efficient. Inspired by directed diffusion, we also apply these two principles in our
design of residual energy scan construction.
Debugging distributed systems is related to continuously monitoring sensor networks. Distributed diagnosis protocols
[HKR84, BH91, BB99] have been designed either for multiprocessor computers or for wired computer networks. Bates
[Bat95] describes a high-level debugging approach for such systems by using events and behavior models. It provides a
uniform view of heterogeneous systems and enables analysis to be performed in well-defined ways. Tarafdar and Garg
[TG98] summarize distributed debugging as observation and control. They propose using Predicate Control as an active
approach to debugging distributed systems, which involve a cycle of observation followed by controlled replaying of
computations, based on observation.
The Internet is growing every day but has been under-instrumented for a long time. Welles and Milliken [WM84]
proposed Loader Debugger Protocol, which is an application layer protocol for loading, dumping and debugging target
machines from hosts in a network environment. Written as early as in 1984, it reveals some key requirements for remote
debugging across computer networks. SNMP [CFSD96] is the de facto standard for inter-network management. Provid-
ing interfaces to a virtual information database, SNMP allows users to query or set values of managed network objects.
SCAN [REG98] provides a multicast-based continuous monitoring infrastructure with good scalability and robustness.
This enables the system to be robust to different kinds of failure e.g. failed nodes or network partition. NIMI [MMP97]
is a nation-wide infrastructure for detecting and debugging the performance problems within networks and their peers.
It can detect both short-term and long-term performance degradation. MINC [CDHD99] utilized the method for deter-
mining interior network performance from the edge measurement while sending multicast probes to measure end-to-end
performance outwards.
Industrial engineers have been designing monitoring mechanisms for power plants, gas companies, telephony systems
and others. For example, SCADA(Supervisory Control and Data Acquisition) technology [Boy93] was introduced more
than 50 years ago to enable a user to collect data and send commands to remote utilities in energy production, transport and
distribution systems. In order to provide real-time alert of crucial events, those systems invest heavily on sophisticated and
reliable devices for data acquisition and control. The deployment of those monitoring devices is also carefully engineered.
Reliability is always the first concern in these critical systems.
When instrumenting a sensor network, lots of ideas can be borrowed from those experiences. However, none of these
techniques is directly applicable to continuously monitoring sensor networks. First, the sheer number of sensor nodes
makes it infeasible, given energy and communication constraints in sensor networks, to centrally collect detailed state
from individual sensor node and then display abstracted view of data. Second, collaborative interaction between individual
3
nodes is, to some extent, inherent in wireless sensor network applications in order to achieve robustness and better accuracy.
Knowledge of the overall state of nodes over a region is more helpful than knowledge of the state of individual nodes. Third,
but most important, the monitoring activities within sensor network should be energy-efficient. Continuous monitoring
itself is inherently energy consuming, and more specifically, considering the high cost of communication in wireless sensor
networks, the messaging cost on monitoring data delivery should be carefully planned.
Most relevant to our work is the coverage problem in wireless sensor networks studied in [MKPS01]. Given the loca-
tions of sensor nodes, these techniques detect maximal breach path (using V oronoi diagrams and Delaunay triangulations
of the sensor field) and maximal support path along which there is poorest and best coverage of sensors, respectively. These
approaches and their distributed versions can detect one or more specific network vulnerabilities. Sensor network scanning
is complementary in that it can provide the network provider an approximate indication of when to invoke these other tools.
3 Residual Energy Scan
A residual energy scan (or eScan for short) depicts the remaining energy levels of sensor nodes. One possible representation
of an eScan of a sensor field is shown in Figure 1. Different regions of the sensor field are shaded differently, depending on
the average residual energy resources of sensors within that region. An eScan can help users to decide where new sensor
nodes need to be deployed to avoid energy depletion. It can also help verifying the behavior of energy-aware adaptive
routing protocols [XHE00]. An eScan is only one kind of sensor network scan. Not all scans are likely to be described in
the same way. Where applicable in the rest of this section, we also discuss design choices for other scans.
Figure 1: An Example of Residual Energy Scan
100%
0%
An eScan
3.1 System Model and Assumptions
Without loss of generality, the wireless sensor network we intend to monitor consists N sensor nodes distributed on a
2-D plane field. Each node is immobile. Each node can communicate with other nodes within certain range. Each node
knows its location on the plane. Obtaining reliable node location has been well studied in different contexts. In practice,
geographical location can be determined by Global Positioning System (GPS) with fair accuracy [Kap96]. Alternatively,
when a GPS device is not available or applicable, other localization systems can be used. For example, a few of the
sensor nodes called beacons know their coordinates in advance, either from GPS or predeployment. Relying on radio
signal strength or acoustic measurement, other nodes can approximate distances to beacons and decide their locations by
trilateration [BHE00, Gir00].
4
Sensor nodes are powered by batteries with normalized capacity of 100%. Each node can measure its local residual
energy by interface similar to APM (Advanced Power Management) or ACPI (Advanced Configuration and Power Inter-
faces). Each node executes one or more sensing tasks, which consumes energy during the whole lifetime of the sensor
network. Energy consumption in this network is dominated by inter-node communication. We emphasize the high energy
cost of communication compared to computation. For example, the energy consumed in transmitting a 1 kilobit packet
100m is approximately the same as performing 3M instructions on prototype wireless sensor nodes [PK00]. For this rea-
son, sensor nodes will prefer to perform significant local collaborative processing of data, rather than transmit data over
long distances.
We also assume that there are one or more special nodes (user gateways) at the network edge, from where human users
will collect residual energy scans. We intend to design network communication and aggregation mechanisms for delivering
energy scans to a user gateway with good scalability, robustness and energy efficiency characteristics.
3.2 Collecting Residual Energy Scans
The process of constructing a eScan of a sensor field can be briefly described as follows (Figure 2):
1. Determining Local Residual Energy: At each node, the residual energy level is measured periodically. The local
scan at a particular node comprises the node’s residual energy level and its location. A sensor node only need to
report its local eScan when there is a significant energy level drop compared to the last time it reported its eScan.
2. Disseminating eScans: Now that local eScans are available at each node, they must be disseminated across the
network to compute a composite eScan of the entire network. For this to happen, the user at the gateway expresses a
special INTEREST message for a network-wide eScan. This INTEREST message propagates throughout the network
by flooding. Upon receiving a INTEREST message, each node sets the sender as its parent node leading toward the
user gateway. Thus an aggregation tree is constructed whose root is at the user gateway. Each node then sends the
local eScans back towards the user. The aggregation tree is refreshed periodically to adapt network dynamics and
failures.
3. Aggregating eScans: Along the path to user gateway, nodes that receive two or more eScans may aggregate those
eScans according to several rules. If the eScans are topologically adjacent and have the same or similar energy
level, they can be aggregated into a tuple which contains a polygon that describes a collection of nodes, and the
range of residual energy levels at those nodes. The goal of aggregation is to reduce the messaging cost on collecting
eScans while losing little critical information content in the scans, and the detailed approach is discussed in the rest
of section.
There are several dimensions for us to explore in the design space, for example: What is the proper compact represen-
tations for scans? How do we aggregate scans to reduce messaging cost? Are there other ways to organize the aggregation
paths in terms of energy-efficiency and robustness? Which network characteristics are proper and which are not? All
these questions are important and interesting, but in this paper, we will focus on designs of representation and aggregation
schemes for eScans.
3.3 Abstracted Representation
Wireless sensor networks have limited energy resources, which makes centralized collection of individual node state infea-
sible. We use sensor network scans to represented abstracted view of a particular network characteristics. More precisely,
we can define a scan as a collection of (VALUE, COVERAGE) tuples.
5
Figure 2: Illustration of Residual Energy Scan Collection
(a) The user expresses interest in eScans.
"! #$&% "’( ) * + , - . / 0 1 2 * 3 4 5 676
(b) Each node selects its parent node to form an
aggregation tree
89;:"< =>?A@BC9D EFHGI"?AC:
(c) Two scans are aggregated into one along the
aggregation tree
VALUE is the quantitative representation of the network state that users are interested in. It may, in general, have a more
complex form than a single scalar value. In eScan, we use VALUE=(min, max), where min, max are the minimum
and maximal residual energy level of the nodes, respectively. For example in Figure 3(a), the eScan VALUE is
(35%,37%).
COVERAGE denotes the region in the sensor network that VALUE describes. In eScan, we use a polygon to describe the
COVERAGE of a scan. The nodes covered by this polygon share in common that their residual energy level falls in
the range of VALUE.min and VALUE.max. The vertices of COVERAGE polygon are the locations of those boundary
nodes. The polygon is not necessarily convex, but it is not self-overlapping (i.e., none of the edges crosses another).
In Figure 3(a) , the coverage polygon of the eScan is shown using a solid line.
Our choice of representation scheme for eScan determines the energy savings on messaging cost. By combining
locations of the nodes having similar energy level, the polygon representation is more compact. Intuitively, if all
J nodes
within a square region have similar values, instead of a list of
J locations for each node, our scheme represents the region
using information proportional to
K J , which is quite a significant reduction for larger
J . Of course, the price we pay
is the loss of detailed information about residual energy at each node and the locations for the interior nodes. However,
though this aggregated representation is lossy, it is still very helpful for a user as an indication of the distribution of network
6
Figure 3: Representation and Aggregation of eScans
L M N OPQ RS"T U V WX U Y WZ X[ N \;\ ] ^ _O ‘ a b c d ef b g b h ikj l m ‘ a b c d ef b g b h ikj n m ‘ a b c d ef b g b h ikj o m (a) A eScan Example
p q r sAt
uvw x y z{ x | z} p q r s~ uv"w x y z{x z} (b) Two eScans Before Aggregation
A
" (c) The eScan after Aggregation with T=10%
energy resource. As we have argued before, a user is not necessarily interested in the energy levels of individual sensor
nodes when monitoring a network of sensors in a large number. The approximate values of residual energy are enough for
the user to identify the near-depleted regions or discover energy consumption patterns. Another component of the price
for aggregation is the cost for local CPU processing, but we assume that processing costs are much lower than delivering
energy level data across the network.
Our VALUE and COVERAGE representation is a reasonable choice for another reason. Over long time scales, the energy
consumption pattern in a sensor network is expected to show spatial locality. We anticipate that all the nodes within a
certain neighborhood detect, and participate in the processing of, similar events, thereby expending similar energy on the
sensing task. If all nodes start out with comparable energy levels, spatial locality can results in good compressibility.
Where these assumptions are broken, of course, scan collection performance may degrade.
There exist alternative representation schemes for eScan and other scans. For example, VALUE in eScans can be in the
form of (avg, wgt) instead, where avg is the averaged residual energy and wgt is the number of nodes. When aggregating
multiple scans (Section 3.4), the aggregated VALUE can be determined. For those scans also capturing spatial locality, for
example, scans that show ambient sensing activities can be described in a manner similar to isothermal curves.
7
3.4 In-network Aggregation of eScans
Initially, each node constructs its local scan using its residual energy level plus its location, and reports this scan to its
parent node leading to the user gateway. Along the aggregation path, if a node receive two or more scans, it will try to
aggregate those scans into a composite one.
3.4.1 Aggregation Rules
One of the most important characteristics of the scan is that they can be aggregated. i.e. two or more scans can be combined
if their COVERAGES are adjacent and/or their VALUES are similar. For example, eScan A and eScan B can be aggregated
if
1. A.VALUE and B.VALUE are similar:
k ¡£¢⁄¥ ƒ§¤ ¡'«“‹fi›fl–†ƒ ƒ§‡ fi›fl·¢⁄¥ ›fl–'
•¶‚„†k ›fl·¢k¤ ¡£¢⁄¥ ƒ§¤ ›fl·¢⁄¥ ¤ ¡' ”…»
2. A.COVERAGE and B.COVERAGE are adjacent:
‰¿
`´ˆƒ˜k¯¤˘¤˙¨§¤¢¨˚¸
‰¿
`´ˆ˝˜⁄k¯¤˘¤˙¨§¤¢¨¤˚˛ ˇ— where T (tolerance) denotes the maximum relative error of residual energy value allowed by aggregation. R(resolution)
decides when two regions are adjacent. Function Shadow(P, R) extends a polygon P with a R-wide shadowed region.
For example, the shadow for A and B are marked as dashed lines in Figure 3(b). The resulted eScan C after aggregation
operation is
¯k§‡ fi›fl ˇ ›fl–†k¤ ›Afl·¢⁄k¤§‡ fi›fl–'
¯k§‡ ˇ †k ¡¢⁄¥ §¤ ¡'
¯k¯¤˘¤ƒ¨¤ ˇ† ‚ ˜ƒ ¯¤˘¤˝¨§¤¢⁄k¯¤˘¤˝¨§¤¢¨˚
where function Merge(P,Q,R) combines two polygons P,Q with resolution R. Obviously, › ˜¯¤˚
› ˜˚•
› ˜⁄˙˚¢
because the aggregation operation only removes vertices from COVERAGE but never adds new vertices. This
means messaging cost for the aggregated eScan is less than the sum of original eScans. An example of aggregation opera-
tion is shown in Figure 3(b,c), and the scan size is reduced by removing the location information for 5 nodes.
T and R are the parameters that control how deeply we aggregate scans. On the other hand, they also decide the
“fidelity” of outcome scans. In eScan, we usually assign a fixed value to R such as radio communication range or sensing
range, thus the quality of aggregated eScan solely depends on value tolerance T. The operation to test and aggregate two
eScans costs polynomial time.
The definition of aggregation operation depends on the underlying representation scheme for scans. If VALUE=(avg,wgt)
as defined in 3.3, the similarity test will be Æª†„ ÆªŁ ØŒº Æª† Æª ”…»
. The aggregated VALUE will be
˜ Æªæ &ª
Æªæ &ª
&ª
&ª
¢ƒ ˆ•‚ı
⁄ ˆ«‚`ı˚
. If the coverage is as simple as a list of nodes, we can merge the coverage lists of scans.
3.4.2 Incremental Update
Incremental updates are necessary for continuously monitoring a sensor network over long time scale. This is clearly
necessary from an energy perspective. When a node’s residual energy changes, it only need to resend an update to current
eScan instead of sending its entire eScan. Furthermore, the aggregation operation of eScan introduces error to aggregated
eScans. If the change of the update is within the aggregation tolerance T, it is not necessary to relay the update along the
aggregation hierarchy to the user.
8
In order to support incremental update, each branch node in the distribution tree (Figure 2) maintains a finite cache
of eScans. Upon receiving one or more eScan updates, they are first compared with old eScans in cache. If an update is
covered by an old eScan and their value are similar according to tolerance T, this update will be dropped because it does
not change the scan significantly. If the update has a value that falls out of the value range of such old scan, the cached
scan will be invalidated from cache by this recent update. The process of incremental update of eScans is described in
Algorithm 1 in Appendix.
3.5 Discussion
In Section 3.3 and 3.4, we have described the core of residual energy scan construction: abstracted representation of resid-
ual energy distribution and its corresponding aggregation operations. By applying in-network aggregation of abstracted
residual energy distribution, we can achieve good scalability and energy-efficiency characteristics in eScan collection.
There are some additional design issues that we are currently investigating.
3.5.1 Complementary Tools
Abstracted indications of network characteristics may fail to reveal interesting and useful information in some scenarios,
compared to complete collection of node states. Some critical information may be lost because of abstraction. There exist
solutions complementary to sensor network scans. As we state in Section 2, centralized and distributed diagnosis protocols
for sensor networks are complementary to sensor network scans. In some cases, sensor network scans can be used as
input to those diagnosis protocols to identify particular network problems. Complementary to sensor network scans is drill
down. When a human observer detects an anomaly in a scan, they may need to obtain detailed information from a specific
region of sensor field. The mechanisms for drill down are similar to those for query distribution and response aggregation
described in [IGE00].
3.5.2 Aggregation Tolerance Adaptation
When aggregating two or more scans, the value of aggregation tolerance decides the size and quality of resulted composite
scan. When the user initiates the collection of residual energy scan, he/she can put the tolerance value to indicate how
deeply each node should aggregate scans. However, to balance the savings in aggregation and the loss of accuracy in scans,
each node should adaptively adjust its aggregation operation locally. For example, if a node keeps receiving a lot of scan
updates, it can increase the aggregation tolerance value to reduce the size of resulted scan. If the node only receives a few
escan updates or those updates are very similar to each other, it can reduce the the aggregation tolerance value to generate
more detailed scans of residual energy.
In general, given a resource budget that can be spent on scanning, the aggregation tolerance adaptation should provide
residual energy scans as detailed as possible. We have not well studied designing adjustment policies and its impact on
constructing eScans, but in principle aggregation tolerance adaptation is a important and necessary feature for monitoring
a real wireless sensor networks.
3.5.3 Aggregation Path Maintenance
In previous description of sensor network scans, we have not addressed the robustness issue in residual energy scans. For
example, a node failure will partition the aggregation tree thus all the offspring nodes are not able to send eScan updates to
the user.
In Section 3.2, we propose that the user periodically refresh INTEREST messages to adapt node failure and network
dynamics. However, flooding over the whole network is cost and should only be issued with low frequency.
9
Complementary to refreshment of aggregation paths, we are considering local recovery of aggregation tree. One
possible simple scheme we are currently investigating is as follows. When a node N detects that its parent fails, it sends
a special MARK message to all of its children and such MARK messages are further propagated to N’s k-th grandchildren
in the sub tree using reliable communication primitives. Those offspring nodes will have temporary knowledge that N is
their (grand)parent. Node N then broadcasts a REPAIR-REQUEST message to its neighbors and only those nodes that are
not N’s offsprings will reply with a REPAIR-RELY message. N choose a non-offspring neighbor as its parent node. If there
is no response, N will keep trying the repair process but for only for p times. Local recovery of the aggregation tree will be
overridden upon reception of INTEREST messages.
We have not well studied the performance of this protocol. Intuitively, it will avoid loops, and perform well in cases
of sparse node failures. If there are massive failures in a small region, this protocol may take time to converge to the new
aggregation path. However, the reconstruction of aggregation tree can compensate this drawback.
4 Experimental Results
In order to evaluate our design of residual energy scans, we compare its performance to centralized collection of individual
residual energy information from each node. We use a stand-alone C++ package to simulate the eScan construction
process for large scale sensor networks. In this section, we present our results and discuss their implications and possible
applications.
4.1 Metrics
The key performance criterion for residual energy scanning is the energy consumed in communicating the scans to the user
node (Section 3.1). We define the following metrics to evaluate the performance of residual energy scanning:
Messaging Cost: Sensor network applications consume energy from time to time. After each sensing activity event, the
energy dissipation may be significant enough to invoke an eScan update. We define the messaging cost
§
for continuous
monitoring as
ˇł…ø
º ¯ º Jßœ
˜›ıŁfl´Ł ¶ flı˚ where
is the number of sensing events happening per node,
¯ º is the sum of messaging cost on collecting eScan updates
during this time frame, and
J is the network size. We define
§
as the cost of centralized collection of residual energy
information without any in-network processing. Compared to the cost of centralized collection,
is expected to reveal
the potential savings from aggregation an incremental update. We quantify such savings as cost ratio
¨ ˇ : a large
¨ indicates significant savings.
Distortion: Aggregation with certain tolerance introduces error into eScans when compared with the actual node
residual energy. We quantify the “fidelity” of an eScan snapshot by the root mean square error between perceived residual
energy values in an eScan and the actual values:
ˇ ø º º “‹ º J where
º is the estimated residual energy of node
fl in the eScan and
º is the real value.
J is the size of the network.
only reflects the absolute error due to aggregation and incremental update. We further define distortion as the fraction of
the absolute error to average residual energy in the sensor network:
ˇ fl ø º † º ' 10
There are other important metrics for the performance of network scans. For example, due to propagation delay, an
eScan update arrived at user gateway is actually a delayed view of the sensor network. The latency distortion is another
metric of accuracy. Such a metric is also impacted by network failures and dynamics. We have not studied this metric yet
but leave it for future work.
4.2 Energy Dissipation Model
There is one missing piece when we define those metrics: Obviously,
and
may be sensitive to the energy consumption
patterns. How do we model the energy dissipation in a sensor network? To our knowledge, there is no realistic model or
empirical data on energy dissipation of large-scale sensor networks. We propose two energy dissipation models that
emphasize different kinds of sensing application activity.
We propose a UNIFORM DISSIPATION model, to capture uniformly distributed sensing activity. More precisely, during
a sensing event, each node
fl in the network has a probability of initiating a local sensing activity, and every node within
a circle of
centered at
fl consumes fixed amount of energy
. This latter feature of our model is inspired by collaborative
sensing algorithms, for example, the beam-forming algorithm described in [YHR
98]. In this algorithm, triggered by
physical activities on the sensor field, the nodes within some vicinity send their data to a particular head node, where all
data are centrally processed to detect particular patterns. However, the geographical correlation expressed in this model
depends on the specific value of the parameters. For instance, if
is very small compared to the network diameter, the
model may not be able to show significant spatial correlation.
In the UNIFORM DISSIPATION model, the residual energy at all nodes decreases at approximately the same rate. How-
ever, in a realistic environment, different regions in a sensor field may have different energy dissipation rates. To model
this, we propose a HOTSPOT DISSIPATION model. In this model, there are
hotpots uniformly distributed in random on
the sensor field but their locations are fixed during the simulation. Each node
fl has a probability of ˇ ˜`˚
to initiate
a local sensing activity, and every node within a circle of
centered at
fl consumes fixed amount of energy
; where
is a density function and
ˇ ›Afl
ø Œ fl “ Œ ' is the distance from
fl to nearest hotspot. We use two density functions:
˜;&˚ ˇ &Æ is the density function for exponential distribution, where the hotspot effect drops quickly with increasing
distance; and the Pareto density
˜„˚ ˇ ˜;¥†˚
Æ , where the impact of the hotspot falls off more gradually than an
exponential distribution with the same value of
.
4.3 Settings
Following the assumptions described in Section 3.1, our experiments were conducted on a regular square sensor grid
consisting of
J ˇ œ nodes. Each node can communicate with its 4 nearest neighbors. The user gateway is at the
upper left corner of the grid. To form an aggregation tree, each node uniform randomly chooses its left or up neighbor as its
parent node. We also assume a perfect MAC layer which implies that the wireless communication channel only consumes
energy when sending and receiving data packets; There is no loss or overhead due to contention or environment changes.
Each node begins with residual energy of 100%. The eScan algorithm in the experiments uses 16 bits to represent
the residual energy and 32 bits for every node location. An eScan is represented as a collection of a (min, max) value
segment plus a coverage polygon. The estimated residual energy in a eScan for a node covered by a polygon is
ØŒº Ø Æ in the corresponding value segment. The messaging cost for transporting an eScan only includes the data size for the eScan
but ignores other overhead such as packet headers. Each node aggregates eScans with tolerance
» . Incremental update is
supported with a cache of size 30. All the three energy dissipation models share the same parameters
ˇ and
ˇ .
No auto-scaled aggregation adjustment is simulated. Each node measures its residual energy periodically , an eScan update
is triggered if the energy drops more than 0.1%.
11
Each run of our experiment corresponds to one choice of random aggregation tree and parameters for one of the energy
dissipation models. To compute continuous monitoring cost
, we also run energy dissipation simulation for
ˇ events before starting eScan process. We continue to simulate energy dissipation for another
ˇ
events in the same
time of eScanning. We then stop energy dissipation simulation but continue eScan process until all updates arrive at user
gateway. We then compute the cost
§
and relative distortion
for the final snapshot. For each run we also compute the
cost
"!
for continuously collecting residual energy level and compute
¨ ˇ# .
For each set of experiments , the number of runs were adjusted to obtain acceptable 95% confidence intervals.
4.4 Results
Figure 4: Messaging Cost Ratio
¨ and Distortion
for eScans
2
4
6
8
10
12
14
16
18
20
100 200 300 400 500 600
Cost Ratio
Network Size (N)
T= 1%
T= 2%
T= 5%
T=10%
T=25%
T=30%
(a) Cost Ratio for Uniform Model
(p=0.1)
2
4
6
8
10
12
14
16
18
100 200 300 400 500 600
Cost Ratio
Network Size (N)
T= 1%
T= 2%
T= 5%
T=10%
T=25%
T=30%
(b) Cost Ratio for Hotspot Model (Ex-
ponential, a=0.5)
2
4
6
8
10
12
14
16
18
100 200 300 400 500 600
Cost Ratio
Network Size (N)
T= 1%
T= 2%
T= 5%
T=10%
T=25%
T=30%
(c) Cost Ratio for Hotspot Model
(Pareto, a=0.5)
0
2
4
6
8
10
12
14
100 200 300 400 500 600
Distortion (%)
$ Network Size (N)
T= 1%
T= 2%
T= 5%
T=10%
T=25%
T=30%
(d) Distortion for Uniform Model
(p=0.1)
0
2
4
6
8
10
12
14
100 200 300 400 500 600
Distortion (%)
$ Network Size (N)
T= 1%
T= 2%
T= 5%
T=10%
T=25%
T=30%
(e) Distortion for Hotspot Model (Ex-
ponential, a=0.5)
0
1
2
3
4
5
6
7
8
9
10
11
100 200 300 400 500 600
Distortion (%)
$ Network Size (N)
T= 1%
T= 2%
T= 5%
T=10%
T=25%
T=30%
(f) Distortion for Hotspot Model
(Pareto, a=0.5)
Figures 4(a)(b)(c) plot
¨ of different aggregation tolerance
» , as a function of the network size
J for different energy
dissipation models and Figures 4(d)(e)(f) plot the corresponding distortion for each set of experiments.
The energy-efficiency of residual energy scanning can be observed when we evaluate cost ratio together with distortion
introduced by aggregation and incremental update. For example, for a network of 400 nodes using the uniform dissipation
model ( ˇ% &
), applying aggregation with a tolerance of 10% can save messaging costs by a factor of 12.5, but only
introduces around 5% distortion. Beyond a certain tolerance level, however, the gains are almost the same. For example,
in Figure 4(a), the curves for T=25% and T=30% are not distinguishable. At these levels, almost the entire network is
aggregated into one polygon. Results for the two hotspot dissipation models also shows the same trend.
Given a fixed aggregation tolerance value, the cost ratio increases with the network size, which indicates that residual
energy scan construction with aggregation and incremental update has better scalability that centralized collection with no
12
aggregation. This trend is common to all three energy dissipation models. The larger the network size, the more significant
the cost difference between delivering aggregated data and raw data across the network. However, for smaller aggregation
tolerances, these results seem to be less dependent on network size. This confirms one of our design principles that given
the high cost for communication, it is preferable to process data locally to aggregate data rather than disseminating raw
data over long distance.
The cost ratio
¨ can be plotted as a function of aggregation tolerance
» in Figure 5(a). From this graph, the cost ratio
increases sharply for larger aggregation tolerance. However the curve converges gradually when
» increases. Figure 5(b)
shows distortion v.s
» , where distortion is roughly a linear function of aggregation of tolerance with slope of around
.
Combined with (a), it can help the user to choose proper aggregation tolerance to save significant message cost but with
limited distortion.
Figure 5: Cost Ratio and Distortion v.s. Aggregation Tolerance
(N=400, Uniform: p=0.1, Exponential: a=0.5, Pareto: a=0.5)
2
4
6
8
10
12
14
16
5 10 15 20 25 30
Cost Ratio
Aggregation Tolerance %
Uniform N=400
Uniform N=400
Exp N=400
Exp N=400
Pareto N=400
Pareto N=400
(a) Cost Ratio v.s. Aggregation Tolerance
0
2
4
6
8
10
12
14
16
5 10 15 20 25 30
Distortion (%)
Aggregation Tolerance %
Uniform N=400
Exp N=400
Pareto N=400
(b) Distortion v.s. Aggregation Tolerance
In Figure 5(a) the two curves of Hotspot-Pareto and Hotspot-exponential dissipation models roughly show the same
trend because they share the exact same number of hotspots and locations. However the curve for Pareto model outperform
the case of exponential distribution. Given the same shape factor of
ˇ’ ( , the Pareto model of energy dissipation
tends to have less localized impact, leading to a smaller deviation of residual energy across nodes, which increases the
aggregatability of the scan.
Besides these functions, we have also studied the performance sensitivity to different simulation parameters such as
the number of hotspots in energy dissipation model, probability function shape factor
. Our experiments show that small
perturbations in those parameters does not change the experimental results significantly.
5 Concluding Comments
Continuously monitoring resource distribution and network activity will be a integral component for future sensor net-
works. To our knowledge, there is no other ongoing or previous work on continuous monitoring large-scale distributed
sensor networks. Our design of residual energy scans provides an overall abstracted view of residual energy in an energy-
efficient manner. Instead of collecting the raw residual energy data from individual nodes, we apply in-network aggregation
to composite residual energy scans. Our design is an instance of trading off local processing cost against the savings in
communicating raw data over long distance. Simulation results show that our approach has good scalability and energy-
efficiency characteristics, compared to extracting the residual energy individually from each node. To some extent, residual
energy scanning itself is a unique application on sensor networks. Our research on this topic will also enrich understanding
13
of sensor network design in general.
We will continue to refine and evaluate our design, especially those issues stated in Section 3.5. Residual energy scan
is only one kind of abstracted indication of sensor network state. Different network resource or performance metrics may
require different techniques to achieve good energy-efficiency, scalability and robustness characteristics. We intend to
investigate other sensor network scans and their impacts in the design of monitoring mechanism for sensor networks.
References
[ADL
98] G. Asada, T. Dong, F. Lin, G. Pottie, W. Kaiser, and H. Marcy. Wireless integrated network sensors: Low
power systems on a chip, 1998.
[Bat95] P. Bates. Debugging Heterogeneous Distributed Systems Using Event-Based Models of Behavior. ACM
Transactions on Computer Systems, 13(1), 1995.
[BB99] Douglas M. Blough and Hongying W. Brown. The Broadcast Comparison Model for On-Line Fault Diagnosis
in Multicomputer Systems. IEEE Transactions on Computers, 48(5):470–493, 1999.
[BH91] A. Bagchi and S. L. Hakimi. An Optimal Algorithm for Distributed System Level Diagnosis. In Proc. 21st Int.
Symp. on Fault-Tolerant Computing (FTCS-21), pages 214–21, Montreal, Que., Canada, 1991. IEEE Computer
Society Press.
[BHE00] N. Bulusu, J. Heidemann, and D. Estrin. GPS-less Low Cost Outdoor Localization for Very Small Devices,
2000.
[Boy93] S. Boyer. SCADA: Supervisory Control and Data Acquisition, 1993.
[CDHD99] R. Caceres, N. G. Duffield, J. Horowitz, and D.Towsley. MINC: Multicast-Based Inference of Network-
Internal Loss Characteristics. Proc. IEEE Infocom, 1999.
[CFSD96] J. Case, M. Fedor, M. L. Schoffstall, and C. Davin. Simple Network Management Protocol. Internet Request
for Comments 11157, 1996.
[EGH00] Deborah Estrin, Ramesh Govindan, and John Heidemann. Embedding the Internet. Communications of the
ACM, 43(5):39–41, May 2000. (special issue guest editors).
[EGHK99] D. Estrin, R. Govindan, J. Heidemann, and S. Kumar. Next Century Challenges: Scalable Coordination in
Sensor Networks. Proceedings of the 4th Annual ACM/IEEE International Conference on Mobile Computing
and Networking, 1999.
[Gir00] L. Girod. Development and Characterization of an Acoustic Rangefinder, 2000.
[HKB99] Wendi Rabiner Heinzelman, Joanna Kulik, and Hari Balakrishnan. Adaptive Protocols for Information Dis-
semination in Wireless Sensor Networks. In Mobile Computing and Networking, pages 174–185, 1999.
[HKR84] Seyed H. Hosseini, Jon G. Kuhl, and Sudhakar M. Reddy. A Diagnosis Algorithm for Distributed Computing
Systems with Dynamic Failure and Repair. IEEE Transactions on Computers, 33(3):223–233, 1984.
14
[IGE00] Chalermek Intanagonwiwat, Ramesh Govindan, and Deborah Estrin. Directed Diffusion: A Scalable and
Robust Communication Paradigm for Sensor Networks. In Proceedings of the 6th Annual International Con-
ference on Mobile Computing and Networking (MOBICOM-00), pages 56–67, N. Y ., August 6–11 2000. ACM
Press.
[Kap96] E. Kaplan. Understanding GPS: Principles and Applications, 1996.
[KKP99] J. M. Kahn, R. H. Katz, and K. S. J. Pister. Next Century Challenges: Mobile Networking for ”Smart Dust”.
Proceedings of the 4th Annual ACM/IEEE International Conference on Mobile Computing and Networking,
1999.
[MKPS01] Seapahn Meguerdichian, Farinaz Koushanfar, Miodrag Potkonjak, and Mani Srivastava. Coverage Problems
in Wireless Ad-hoc Sensor Networks. IEEE Infocom, 2001.
[MMP97] J. Mahdavi, M. Mathis, and V . Paxson. Creating a national measurement infrastructure (nimi).
http://www.psc.edu/networking/nimi, May 1997.
[PK00] G. J. Pottie and W. J. Kaiser. Embedding the Internet: wireless integrated network sensors. Communications
of the ACM, 43(5):51–51, May 2000.
[REG98] A. Reddy, D. Estrin, and R. Govindan. Large Scale Fault Isolation. IEEE Journal on Selected Areas of
Communications, 1998.
[Ten00] David Tennenhouse. Embedding the Internet: proactive computing. Communications of the ACM, 43(5):43–
43, May 2000.
[TG98] A. Tarafdar and V .K. Garg. Predicate Control for Active Debugging of Distributed Programs. Symposium on
Distributed and Parallel Debugging, 1998.
[WM84] Christopher Welles and Walter Milliken. Loader Debugger Protocol. Internet Request for Comments 909,
1984.
[XHE00] Ya Xu, John Heidemann, and Deborah Estrin. Adaptive energy-conserving routing for multihop ad hoc net-
works, 2000.
[YHR
98] K. Yao, R. Hudson, C. Reed, D. Chen, and F. Lorenzelli. Blind Beamforming On A Randomly Distributed
Sensor Array System, 1998.
Appendix
15
Algorithm 1 Incremental Update of eScan
/* Cache of Aggregated Scans*/
FIFO_QUEUE Cache=
— ;
While (TRUE){
Case ENERGY_LEVEL_CHANGED:
U=buildScan(ResidualEnergy, Location);
Send U to Localhost;
Case CACHE_TIMED_OUT:
/*Time Out the Oldest eScan in Cache */
Cache.pop();
Case ESCAN_RECEIVED:
eScans {U1,U2,...}=recvScans();
For each S in Cache s.t S covers Ui {
If
Ø Æ *)
Ø Ææ +,A
Ø Æ- Ø‡Œº*)
Ø‡Œº
+,A
ØŒº . /0 ) ØŒº ) Ø Æ-21
” » Discard Ui; /* Drop insignificant scan update */
Else
Cache.delete({S}); /* Drop out of date cached scan */
}
Cache.append({U1,U2,...});
For any pair S1 and S2 in Cache has similar VALUE and adjacent COVERAGE{
S=Aggregate(S1,S2);
Cache.del({S1,S2});
Cache.append({S});
}
For each U in Cache that covers any Ui
Send out eScan U
}
16

Abstract (if available)

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 774 (2002)

PDF

USC Computer Science Technical Reports, no. 692 (1999)

PDF

USC Computer Science Technical Reports, no. 750 (2001)

PDF

USC Computer Science Technical Reports, no. 732 (2000)

PDF

USC Computer Science Technical Reports, no. 669 (1998)

PDF

USC Computer Science Technical Reports, no. 704 (1999)

PDF

USC Computer Science Technical Reports, no. 603 (1995)

PDF

USC Computer Science Technical Reports, no. 724 (2000)

PDF

USC Computer Science Technical Reports, no. 771 (2002)

PDF

USC Computer Science Technical Reports, no. 677 (1998)

PDF

USC Computer Science Technical Reports, no. 746 (2001)

PDF

USC Computer Science Technical Reports, no. 841 (2005)

PDF

USC Computer Science Technical Reports, no. 873 (2005)

PDF

USC Computer Science Technical Reports, no. 631 (1996)

PDF

USC Computer Science Technical Reports, no. 706 (1999)

PDF

USC Computer Science Technical Reports, no. 888 (2007)

PDF

USC Computer Science Technical Reports, no. 731 (2000)

PDF

USC Computer Science Technical Reports, no. 723 (2000)

PDF

USC Computer Science Technical Reports, no. 697 (1999)

PDF

USC Computer Science Technical Reports, no. 957 (2015)

Description

Yonggang "Jerry" Zhao, Ramesh Govindan, Deborah Estrin. "Residual engery scans for monitoring wireless sensor networks." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 745 (2001).

Asset Metadata

Creator Estrin, Deborah (author), Govindan, Ramesh (author), Zhao, Yonggang "Jerry" (author)

Core Title USC Computer Science Technical Reports, no. 745 (2001)

Alternative Title Residual engery scans for monitoring wireless sensor networks (title)

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)

Tag OAI-PMH Harvest

Format 16 pages (extent), technical reports (aat)

Language English

Unique identifier UC16270332

Identifier 01-745 Residual Energy Scans for Monitoring Wireless Sensor Networks (filename)

Legacy Identifier usc-cstr-01-745

Format 16 pages (extent),technical reports (aat)

Rights Department of Computer Science (University of Southern California) and the author(s).

Internet Media Type application/pdf

Source 20180426-rozan-cstechreports-shoaf (batch), Computer Science Technical Report Archive (collection), University of Southern California. Department of Computer Science. Technical Reports (series)

Access Conditions The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

Description Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)