Measurement and Analysis on the Packet Delivery
Performance in A Large Scale Sensor Network
Wei Dong
†
, Yunhao Liu
‡§
, Yuan He
‡
, Tong Zhu
§
†
CS College, Zhejiang University;
‡
TNLIST, Tsinghua University;
§
CSE Dept. HKUST
dongw@zju.edu.cn, {yunhao, he}@greenorbs.com, zhutong@cse.ust.hk
Abstract—Understanding the packet delivery performance of
a wireless sensor network (WSN) is critical for improving system
performance and exploring further development and applications
of WSN techniques. In spite of many empirical measurements in
the literature, we still lack in-depth understanding on how and
to what extent different factors contribute to the overall packet
losses with respect to a complete stack of protocols at large scales.
Specifically, very little is known about (1) When, where, and
under what kind of circumstances packet losses occur. (2) Why
packets are lost. As a step towards addressing those issues, we
deploy a large-scale WSN and design a measurement system for
retrieving important system metrics. We propose MAP, a step-
by-step methodology to identify the losses, extract system events,
and perform spatial-temporal correlation analysis by employing
a carefully examined casual graph. MAP enables us to get a
closer look at the root causes of packet losses in a low-power ad-
hoc network. This study validates some earlier conjectures on
WSNs and reveals some new findings. The quantitative results
also shed lights for future large-scale WSN deployments.
I. INTRODUCTION
As an emerging technology that bridges cyber systems and
the physical world, wireless sensor networks (WSNs) are
envisioned to support numerous unprecedented applications.
We have witnessed many research studies, deployments of real
systems, and substantive practical applications in recent years.
We are still facing severe challenges in designing scalable,
long-lived, and high-performance WSN systems. Some of the
difficulties come from the fact that the current understanding
of WSN is still limited. Therefore, it is necessary to conduct
empirical measurements in real-world WSN systems, so that
we can better understand the behaviors of large-scale WSNs
and facilitate the design of such systems.
In the past years, many WSN protocols have been reported
and shown to be effective in testbed or small-scale networks.
On the other hand, it is not uncommon to see that many real
deployments often adopt a set of tailored protocols to fulfil the
application’s requirements. We believe that it is important to
understand the performance of some well-principled protocols
in combination at a large scale. We are interested in the
question: whether they are reliable enough to facilitate the
development of future WSNs.
Many deployments have reported the overall packet delivery
performance [1], [2]. Also, many empirical measurement stud-
ies show how some specific factors impact the packet delivery
performance via controlled experiments [3], [4]. However, we
usually do not know how and to what extent different factors
contribute to the overall packet losses with a complete stack
of protocols at a large scale. Specifically, very little is known
about (1) When, where, and under what kind of circumstances
packet losses occur. (2) Why packets are lost. Answers to the
above questions are critical for improving system performance
and exploring further development and applications of WSN
techniques.
Understanding the packet delivery performance in an oper-
ating WSN is challenging due to the following facts. First,
data packets might be lost during multi-hop transmissions
and thus data collection is incomplete by nature. It is very
difficult to acquire the complete information of the internal
status. In-depth understanding requires detailed measurement
of networking and system metrics, but basically this is far
from affordable for resource and energy constrained sensor
nodes. Second, operational efforts to disclose the root causes
behind packet losses are insufficient, and few efforts have
been validated to be effective at large scales. Third, fine-
grained measurements usually demand mechanisms which
incur unnegligible operational or capital expenses.
As a step towards addressing those challenges, we deploy
GreenOrbs, a large-scale and long-term WSN system in the
wild. The network we measure is in continuous operation
since Dec. 2010 with nearly 400 nodes. For the sensor node
hardware, we use the commonly used TelosB nodes. For the
software, we use TinyOS and its radio stack, including the LPL
MAC, the CTP collection protocol, and the Drip dissemination
protocol.
Based on GreenOrbs, we propose MAP, a practical method-
ology for Measuring and Analyzing the Performance of a
large operating WSN. MAP incorporates a well-designed
measurement system for retrieving networking and system
metrics. MAP includes three steps for analyzing the packet
losses. First, it uses robust algorithms to identify the losses as
well as important system events. Second, it carefully tracks
the interactions inside the WSN system by means of a causal
graph. Third, it examines the temporal-spatial correlations
among system events having casual relationships.
Using this methodology, we are able to conduct a deep
examination of packet losses. For example, for a recent de-
ployment of our system, Figure 1 shows the spatial distribution
of packet losses along with the geometric network topology.
The red nodes are those with packet delivery ratio (PDR)
less than 90%, and the length of the radius indicates the
number of lost packets. We note that while previous reports
usually show us the delivery performance as in Figure 1,
they usually lack detailed analysis on classifying the losses
into smaller categories that can be useful for further analysis.
978-1-4673-5946-7/13/$31.00 ©2013 IEEE
2013 Proceedings IEEE INFOCOM
2679