Magnetic diffusion: Scalability, reliability, and QoS of data dissemination mechanisms for wireless sensor networks

(1)

Magnetic diﬀusion: Scalability, reliability, and QoS

of data dissemination mechanisms for wireless sensor networks

Hsing-Jung Huang, Ting-Hao Chang, Shu-Yu Hu, Polly Huang

*

Electrical Engineering Department, National Taiwan University, No. 1, Sec. 4, Roosevelt Road, Taipei, Taiwan Available online 3 March 2006

Abstract

Envisioning a new generation of sensor network applications in healthcare and workplace safety, we seek mechanisms that provide timely and reliable transmissions of mission-critical data. Inspired by the physics in magnetism, we propose a simple diffusion-based data dissemination mechanism, referred to as the magnetic diffusion (MD). In that, the data sink, functioning like the magnet, propagates the magnetic charge to set up the magnetic field. Under the influence of the magnetic field, the sensor data, functioning like the metallic nails, are attracted towards the sink. We compare MD to the state-of-the-art mechanisms and find that MD: (1) performs the best in timely delivery of data, (2) achieves high data reliability in the presence of network dynamics, and yet (3) works as energy efficiently as the state of the art. These suggest that MD is an effective data dissemination solution to the mission-critical applications.

1. Introduction

The technological advances have enlightened a future of intelligent and pervasive computing and communication. In that, miniature and robust sensor nodes would be able to generate, pre-process, and communicate metrics about the environment. For instance, auto-sensing of the room temperature allows tuning of the air conditioning to the level just as necessary. Auto-detection of pulse anomalies allows prevention of irreversible damages caused by diseas-es that are preceded by arrhythmia.

The emergence of the sensors and sensor networks can impact a wide variety of applications from national securi-ty to consumer electronics. Envisioning a new generation of sensor network applications in healthcare and work-place safety, we seek mechanisms that provide reliable and timely transmissions of mission-critical data. Even though wall power will be more accessible in consumer electronics applications, the sensor nodes are likely to run

on limited battery power for the ease of deployment. Thus, energy eﬃciency remains an important design challenge.

Much of the related work in sensor network data dis-semination either emphasizes the energy-efficient design or stresses the reliable data transfer. We have not yet been able to identify any mechanism comparable to the pro-posed mechanism, magnetic diffusion (MD), that aims at achieving timely delivery, reliability, and energy efficiency. MD is a simple data dissemination mechanism that promises all the above properties. The inspiration comes from the magnetic interactions in the nature. Consider the data sink as a magnet and the data as nails. The data will be attracted towards the sink according to the magnetic field just as the metallic nails being attract-ed towards the magnet. The magnetic field is establishattract-ed by setting up the proper magnetic charges on the sensor nodes within the range of data sink. The strength of the charge is determined by the hop distance to the sink and the level of resource available at the sink. The data will be propagated based on the magnetic field from low to high magnetically charged nodes. This way of

dissemi-nating data results in optimal delay multi-path

forwarding.

*

Corresponding author.

E-mail addresses: [email protected], [email protected],

[email protected],[email protected](P. Huang).

(2)

We are able to observe from the simulation results that MD does (1) perform the best in timely delivery of data, (2) achieve high data reliability in the presence of network dynamics, and yet (3) work as energy eﬃ-ciently as the state-of-the-art mechanisms. We thus con-clude that MD is a promising data dissemination solution to the mission-critical applications.

2. Related work

There is a significant volume of related work in sensor network data dissemination. MD is particularly related to research in energy-efficient routing and reliable data delivery. To put our work in context, we compare and contrast the existing work to our mechanism. In general, we are not able to identify any mechanisms comparable to MD that aims at achieving timely delivery, reliability, and energy efficiency. This is primarily due to the limited attention in sensor network design for mission-critical applications.

We choose to compare MD to the mechanisms that we are able to ﬁnd actual implementations for the frequently used

Motes/TinyOS platform [28,29]. Directed diﬀusion [1,2]

and ﬂooding are the mechanisms considered in our evalua-tion. To facilitate discussions in the later sections, we give an overview of the directed diﬀusion mechanism, following the summary of existing work.

2.1. Energy-eﬃcient data dissemination

There are three major classes of energy-eﬃcient data dissemination mechanisms, the cluster-based, random-walk, and location-aware mechanisms. The cluster-based

approaches[4–6] evenly distribute energy load among the

sensors in the network. The idea is to form clusters for local-ized data dissemination and send only the aggregated data across the clusters to conserve energy. Random-walk

approaches[7,8]randomly select the next hop targets to

dis-seminate data. These mechanisms are energy eﬃcient in that they avoid the sending of control messages and the

mainte-nance of routing states. Location-aware schemes [9,10]

exploit the nodes’ geographical location information to min-imize the cost of disseminating data. The location informa-tion might not be easy to obtain for certain sensor network applications.

More closely related to our work are the mechanisms dealing with efficient data dissemination to mobile sinks [11,12]. MD is similar in that our mechanism also aims at energy-efficient data dissemination for dynamic networks. MD is yet different for that the mechanism design takes into account mobile non-sink nodes and considers the on–off type of network dynamics.

2.2. Reliable data dissemination

There are two major approaches towards better data delivery reliability. The passive approach retransmits data

when losses are detected, whereas the active approach aims, instead, at avoiding losses.

More speciﬁcally in the passive approach category, some

proposed mechanisms[13,14]recover data end-to-end, and

the others[15,16]retransmit hop-by-hop. In the end-to-end

mechanisms, the sink nodes track the status of data deliv-ery and send request messages to other nodes for recovdeliv-ery. The hop-by-hop mechanisms implement the data recovery between two neighboring nodes. When a data packet is lost, the intermediate sending node retransmits.

In the active approach category, several schemes[17–19]

attempt to improve reliability by avoiding congestions. By avoiding congestions, these schemes indirectly lower the loss rate and improve the overall reliability. In addition,

some schemes [20–22] avoid losses by selecting the data

paths of lower loss rates.

Our work is closely related to[23], a multi-path

mecha-nism. This work has suggested that disseminating data over multiple paths improves the reliability. The number of paths selected in the mechanism depends on the priority of the data and the network situation. Our MD is similar to this mechanism in that we also disseminate data over multiple paths for reliability. The primary diﬀerence is in our choices of the multiple paths. Opting for energy eﬃ-ciency and timeliness of data delivery, we scope our choice to all the available shortest paths.

2.3. Directed diﬀusion

Directed diﬀusion[1,2]is a data-centric data

dissemina-tion protocol for wireless sensor networks. The mechanism achieves energy eﬃciency by means of selecting empirically good paths and in-network aggregation. There are two dis-semination modes in directed diﬀusion: one-phase pull and two-phase pull.

2.3.1. One-phase pull

In one-phase pull (OPP), the sink periodically broad-casts an interest message to each of its neighbors. The inter-est message speciﬁes the data the sink is interinter-ested. The neighboring nodes continue to broadcast the interest mes-sage to their own neighbors if the interest mesmes-sage is not a duplicate. This action will be repeated until all nodes have received the interest message. Then, the source with the matching data selects the path of shortest latency according to the interest arrival time, and disseminates data along this path to the sink. The data path is periodi-cally refreshed by interest messages.

2.3.2. Two-phase pull

Two-phase pull (TPP) operates in a similar way with a slight complication. The sink also periodically broadcasts an interest message to its neighbors, and the neighbors operate similarly to those in OPP. The main diﬀerence is at the broadcasting of exploratory data from the data source to all nodes in the sensor network. After the sink receives this exploratory data packet, it reinforces a good

(3)

path back upstream by sending a reinforcement message. This reinforced path is determined according to metrics such as delay, loss rate, and available bandwidth. The node that receives the reinforcement message acts in the same way and forwards the reinforcement message back upstream until the source receives this reinforcement mes-sage. After that, the source sends data along the reinforced path.

To adapt to network dynamics, the reinforced path is periodically refreshed by periodic exploratory data. To conserve energy due to the ﬂooding of data packets, the interval of periodic exploratory data is set to be signiﬁcant-ly longer than that of the periodic interest messages.

Please note following important diﬀerences between OPP and TPP that will be referred later while comparing the performance of OPP, TPP, and MD:

• To avoid collisions, TPP implements its own random wait mechanism to prevent synchronized sending of exploratory data. Each node selects a random waiting time when receiving an exploratory data. There is only one data transmission path in OPP. It does not need to schedule a random wait before propagating data. • To adapt to failures, OPP refreshes the data path by

periodic interest message, whereas TPP refreshes the reinforced path by periodic exploratory data.

• The frequencies of sending exploratory data and interest messages are diﬀerent. The frequency of periodic explor-atory data is twice as low as that of periodic interest messages.

3. Magnetic diﬀusion

Magnetic diffusion is simple and yet powerful. Consider the data sink as a magnet and the data as metallic nails. The data will be attracted towards the sink according to the magnetic field just as the nails are attracted towards the magnet. The magnetic field is established by setting up the proper magnetic charges on the sensor nodes within the magnetic influence of the data sink. The strength of the charge is determined by both the hop distance to the sink and the level of resource available on the sink. The data will be propagated based on the magnetic field from

low-charge to high-low-charge nodes (Fig. 1). This way of

dissemi-nating data promises the following properties.

Multiple paths. By the simple principle of data traveling towards the center of attraction, the sensor nodes forward data coming strictly from the nodes with lower charges (Fig. 2). Forwarding data this way, the paths selected by MD are optimal. Furthermore, there can be multiple next hops to forward the data. Thus, MD sends data in an opti-mal multi-path fashion.

Load balancing. When there are multiple magnets in the environment, the multiple magnetic fields will find their balance. The nails will thus be attracted by the magnets nearby. The same principle applies to MD. When there exist multiple sinks, their magnetic influences,

decrement-ing MD charges from the sinks, will strike a balance (Fig. 3) and attract the data nearby. One advantage of such property is that whenever a new sink joins or an old sink leaves this system, a new balance will be re-established automatically.

Resource awareness. In the real world, some data sinks might be more resource-limited than the others. For exam-ple, there could be sink nodes supplied by unlimited wall power, while others run on batteries. By assigning sink charges based on the amount of resource at hand, we can achieve the goal of balancing the use of heterogeneous resource and potentially prolonging the network lifetime without human intervention.

Fig. 1. Data moving towards the magnet (i.e., the sink).

×s3 ×s3 ×s3 ×s3 ×s2 ×s2 ×s1

Fig. 2. Multiple paths: multiple nodes with equal charge, s1 < s2 < s3.

(4)

In a healthcare example, a patient with chronic diseases may wear biometric sensors for long-term monitoring of the patient’s body temperature, blood pressure, and heart rate. These sensor data need to be collected to a remote patient record server for further analysis. When the patient goes outdoors where there are no infrastructural sinks, the sensor data will need to be transmitted to a resource-limit-ed personal device such as the patient’s PDA. Through the GPRS interface, the sensor data can be transmitted to the patient record server. When the patient comes back home, or wherever there are more resourceful sinks such as a full-ﬂedged PC with low-cost ADSL connection, one can assign a stronger magnetic charge to the PC. Then the data, aware of the stronger attraction, will go through the PC rather

than the PDA (Fig. 4). Such resource-aware feature brings

not only convenience but also optimized utilization of the resources in the environment.

MD operates in two phases: the interest broadcast and data propagation. The magnetic ﬁeld is established in the interest broadcast phase such that the data can be dissem-inated towards the sink in the data propagation phase. In the subsections below, we describe the details of the MD operation and the implementation options.

3.1. Interest broadcast

We first introduce two data entities involved in the inter-est broadcast phase: the interinter-est and the entry. The interinter-est is a message used to establish the magnetic field to facilitate the proper flow of data toward the sink. An interest mes-sage consists of two elements, the data type and magnetic charge. The former records what data type the sink wants to collect. The latter is an integer value recording the strength of the magnetic influence from the sink. The entry is where a node records the data type and the magnetic charge in each node.

When a sink wants to collect data, it sends an interest message to its neighbors. When a node receives an interest message for the ﬁrst time, it creates an entry for this inter-est. Then, the node decrements the magnetic charge in the interest by one. The node records the data type and the magnetic charge into its entry and then forwards the inter-est message to its neighbors. The decrementing magnetic

charges from the sink to source will guide the ﬂow of data in the reverse direction, mimicking the traverse of metallic nails from the low-charge to high-charge points in the mag-netic ﬁeld.

If a node receives an interest message and the corre-sponding entry exists, it will compare the decremented magnetic charge of the interest with the one in the entry. If the magnetic charge of the interest is greater than the one in the entry, the node will update the value in the entry to that of the interest and forward the interests to its neigh-bors. Otherwise, the node will discard the interest knowing the interest is not from a node closer to the current sink.

The sink node will broadcast interest messages periodi-cally to re-establish proper magnetic charges in case of network dynamics. It provides a robust environment espe-cially for the dynamic network.

InFig. 5(a), the sink node broadcasts the interest to its

neighbor nodes. InFig. 5(b), both nodes, A and B, receive

the interest, create corresponding entries, set the data type and proper magnetic charges in the entries, and then

prop-agate the interests to their neighbors. InFig. 5(c), node C,

D, and E receive the interests from node A and B and

repeat the actions taken by A and B inFig. 5(b). Each node

may receive and discard duplicate interests or interests of weaker magnetic charges.

3.2. Data propagation

We next describe how the data are disseminated through a network. We have two implementation strategies for data propagation. One is gradient-based and the other is broad-cast-based.

3.2.1. Gradient-based

In the interest broadcast phase, when a node receives an interest from the neighboring nodes with a stronger mag-netic charge, the node will establish a gradient toward the interest-sending node. This gradient will lead the data through the network from the source to the sink.

When a node senses data, it checks if it has an entry matching the data type. If it does, the node sends the data to the nodes pointed by the existing gradients. The receiv-ing nodes check for the matchreceiv-ing entry and then send the

Fig. 4. Resource awareness: larger magnet out-attracting the small magnet. (a) Small magnet (resource-limited sink, e.g., PDA) attracts data and transmits via resource-limited network (e.g., GPRS) to Internet. (b) Large magnet (more resourceful sink, e.g., PC) which overwhelms the small magnet, attracts data, and transmits via wider bandwidth network (e.g., ADSL) to Internet.

(5)

data according to the gradients as well. The forwarding process goes on and the data will reach the sink eventually. Fig. 6(a) shows the gradients established by the interest

broadcast in Fig. 5. In Fig. 6(b), the source sends data to

the two neighbor nodes pointed by the gradients from

the source. In Fig. 6(c), the receiving nodes continue to

send data to their neighbor nodes based on their gradients. The sink ﬁnally receives the data.

3.2.2. Broadcast-based

In the broadcast-based mode, the nodes do not establish gradients in the interest broadcast phase. Instead, the mag-netic charge is included in the data being disseminated. The receiving node can tell from the charge carried in the data where the data are coming from and whether to forward the data further down the stream.

More speciﬁcally, when a node receives the data, the node checks if it has any matching entry. If so, it compares the magnetic charge in the entry with the magnetic charge of the data. If the magnetic charge in the entry is greater than that of the data, it sets the magnetic charge of the data to the charge in the entry and then broadcasts the data. This means the data are sent from the node whose magnet-ic charge is lower than the intermediate node and the inter-mediate node should continue to broadcast them toward the sink. If a node receives a duplicate data or the data whose magnetic charge is greater than that in the entry, the data will be discarded.

This dissemination mode is easy to implement and we do not need to maintain the gradients. The node simply com-pares the magnetic charges to decide whether to broadcast the data. The drawback is that every data message will have to carry the magnetic charge of the sending node.

InFig. 7(a), the magnetic charge of every node is

estab-lished with the sink of change strength 7. InFig. 7(b), when

the source wants to send data to the sink, it will broadcast

the data to its neighbor nodes. InFig. 7(c), the nodes with

charge strength 6 broadcast the data because the magnetic

charge of the nodes is greater than that of the data. Thus, the sink receives the data.

3.3. Random wait

When the interests or data are being broadcast hop by hop through the network, the messages might collide to each other. To avoid collisions, the node waits a random period of time before sending a message. This technique might decrease the probability of collision and, in the meantime, increase the transmission delay.

4. Evaluation

We have conducted an extensive set of simulations that examines the scalability of MD to the size of the network, the change of the data rate, and the level of the network

dynamics. In Section4.1, we describe the simulation setup

and the metrics examined for the performance evaluation.

In Section4.2, we provide the comparisons of two diﬀerent

mechanisms of MD. Section4.3examines the scalability of

MD. The impact of random wait to the reliability of MD is

discussed throughout Section4.4. In Section4.5, we focus

on the quality of data delivery and the message overhead of

MD in the presence of network dynamics. In Section4.6, a

set of selection guidelines for data dissemination mecha-nisms is provided.

4.1. Simulation setup and metrics

In this section, we describe the simulation setup and the metrics examined for the performance evaluation. We

implemented MD in the ns-2 simulator [30]. In order to

get the average behavior, there are 10 distinct runs for each setup. In each run, there are 50 nodes randomly placed in a 160 m by 160 m square. Each node has a radio range of 40 m. We use one source and one sink randomly selected from 50 nodes in our simulation. The source sends data every 5 s, and the interests are periodically sent every 30 s. The simulation time is 600 s, but the source will stop

sending data after 300 s.Table 1summarizes the

parame-ters of our choice.

The above scenario serves as the base setup. For the comparisons of gradient-based and broadcast-based MD, we use the exact base setup. To examine the scalability of the mechanisms, we increase the number of nodes from 10 to 200. The size of the simulated area is increased to 5 5 5 6 6 7 _sink A B E D C 6 6 7 _sink A B E D C 7 _sink A B E D C a b c

Fig. 5. An example of interest propagation.

5 5 5 6 6 7 _Sink Source 5 5 5 6 6 7 _Sink Source 5 5 5 6 6 7 _Sink Source a b c

Fig. 6. An example of data propagation with gradient-based mechanism.

5 5 5 6 6 7 _Sink Source 5 5 5 6 6 7 _Sink Source 5 5 5 6 6 7 _Sink Source a b c

Fig. 7. An example of data propagation with broadcast-based mechanism.

(6)

keep the node density constant. To study the impact of ran-dom wait, we keep all parameters the same as in the base setup except for the range of the random wait intervals. To simulate network dynamics, the nodes are scheduled to move around in the simulated area and the nodes may go on and oﬀ in random patterns.

We choose three metrics to evaluate the performance of MD and to compare it to other schemes: Overhead mea-sures the amount of interest and data packets transmitted. The metric is closely related to the energy consumption of the system in wireless sensor networks when an energy-ef-ﬁcient MAC protocol is applied. Reachability measures the probability the sink receives the data packet successful-ly. This metric is important for medical applications in that the sensor data are mission critical and data losses can be life-or-death matters. Latency measures the data transmis-sion time from the source to the sink. For medical applica-tions, the metric represents the timeliness and temporal reliability of the data.

4.2. Gradient vs. broadcast

To compare the performance of two forwarding strate-gies, we simulate both the MD gradient (MDG) and broad-cast (MDB) modes using the basic setup. Simulation results

are shown inTable 2andFig. 8.Table 2shows the

over-head in terms of the data, interest, total packet counts,

and the average reachability.Fig. 8depicts the cumulative

distribution of the data delivery latency. The data latency of MDB is generally better than that of MDG. As the results suggest, there are several advantages of the MDB mode over the MDG mode. MDB is more energy efficient. The interest packet overhead of the two modes is the same because interests are transmitted in an identical way in both modes. The difference lies in the data packet over-head. In MDG, if a node has five gradients to its neighbors,

there are ﬁve packets sent, one for each gradient. But in MDB, it broadcasts only one data packet. Therefore, the overall overhead of MDB is much smaller than MDG. Additionally, the data latency is small in MDB. It does not need to send the handshake packets, e.g., the RTS/ CTS in IEEE 802.11, before sending a data packet, but MDG does. As a result, MDB behaves better in latency. However, without the handshaking, there are more colli-sions observed in MDB simulations, and this results in a slight lower reachability in some cases. MDB also performs better in the presence of network dynamics. Given the var-ious advantages of the MDB over MDG, we use MDB for the rest of the comparison.

4.3. Scalability

To compare the scalability of DD, MD, and ﬂooding, we change the number of nodes in the simulated network from 10, 50, 100, 150, to 200. We change the size of the sim-ulated area accordingly to keep a ﬁxed node density. 4.3.1. Overhead

Fig. 9shows the overhead for diﬀerent number of nodes. InFig. 9(a), the interest packet overhead of TPP is slightly higher than that of OPP and MD in all cases. This is because TPP has to disseminate additional control packets such as the positive and negative reinforcement messages, whereas MD and OPP do not.

Note that we set the interval of periodic interest to 30 s in all mechanisms. That is the reason that the interest packet overhead of MD and OPP is almost identical.

Fig. 9(b) shows that MD has a lower data packet over-head than that of the TPP and a higher data overover-head over that of the OPP in all network sizes except the 10-node case. OPP selects only one way to disseminate data, and thus has the least amount of data packet overhead. This result suggests that, independent of the network sizes, the resulting data packet overhead of TPP’s exploratory data is higher than that of MD’s multi-path delivery, and that MD is more energy eﬃcient than TPP. In the 10-node case,

Table 1 Simulation settings Nodes 50 Size 160· 160 Radio range 40 m Data rate 5 s Periodic interest 30 s MAC 802.11 Antenna OmniAntenna Simulation time 600 s Table 2

Comparisons of overhead and reachability between MDG and MDB

Gradient Broadcast Interest packets 4,970 4970 Data packets 5,162 3490 Total packets 10,132 8460 Reachability 100% 98.6441% 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 10000 Cumulative probability delay (ms) Latency MDG MDB

(7)

MD’s multiple paths might cover the whole network and thus the data packet overhead of MD and TPP is too close to set the results apart. It seems that MD is worse than OPP in terms of data packet overhead. However, the slight-ly higher data overhead, in return, provides better data reachability and latency. The reachability and latency of the mechanisms will be discussed next.

Fig. 9(c) shows that ﬂooding has the largest total packet overhead even though the mechanism does not require any control packets. More importantly, we observe that (1) ﬂooding scales the worst with the increasing network size, and (2) the multi-path MD scales as well as the single-path OPP.

4.3.2. Reachability

Fig. 10 shows the reachability of the mechanism in different network sizes. OPP and TPP experience an 100% reachability, MD and flooding’s reachability is at about 95%. The reason of the difference is at the way the data are propagated. OPP and TPP are gradient-based, whereas MD and flooding are broadcast-based. If data packets collide, the gradient-based mechanisms will retransmit assuming IEEE 802.11 as the underlying MAC protocol. The broadcast-based mechanisms, how-ever, do not detect, nor recover for errors, and thus the slightly lower reachability observed in MD and flooding. Reachability of the mechanisms seems to be independent of the network size.

4.3.3. Latency

The latency results are similar with varying network sizes. We compare the mechanisms ﬁrst using the 50-node result and then discuss their scalability with increasing

net-work sizes. InFig. 11, MD has the least latency because it

broadcasts data and bypasses the RTS/CST handshake in IEEE 802.11. This is also why OPP performs much worse than MD, even if it always finds the fastest path. Flooding is better than OPP and TPP only in the preceding 60%. We observe an unusual amount of traffic and collisions in the flooding set of simulations. As a result, some data packets get retransmitted repeatedly and routed for a long way to finally reach the sink. This explains why flooding has a lot of data packets with long latency. TPP is very close to OPP in the preceding 90%, but there is a long tail caused by the random wait mechanism in disseminating explorato-ry data in the later 10%. This plot shows that the broad-0 5000 10000 15000 20000 25000 30000 35000 40000 45000 200 150 100 50 10 Packet # Node# Interest Overhead OPP TPP MD Flooding 0 20000 40000 60000 80000 100000 120000 200 150 100 50 10 P a cke t # Node # Data Overhead OPP TPP MD Flooding 0 20000 40000 60000 80000 100000 120000 200 150 100 50 10 Packet # Node # Total Overhead OPP TPP MD Flooding a b c

Fig. 9. Overhead with increasing network sizes. (a) Interest packet overhead. (b) Data packet overhead. (c) Total (interest + data) packet overhead. 0 10 20 30 40 50 60 70 80 90 100 200 150 100 50 10 Reachability Node # OPP TPP MD Flooding

(8)

cast-based mechanisms work better for applications with latency requirements.

For other network sizes, we find all curves shift to the right. The reason is that the transmission paths are gen-erally longer when the networks are larger. The advan-tage of flooding over OPP and TPP diminishes with the increasing number of nodes. More nodes generate more traffic. This results in more collisions consequently. Thus, there are gradually more packets experiencing long delays in the flooding case. MD scales better than flood-ing by propagatflood-ing data over a selective set of paths. 4.4. Impact of random wait

The slight lower reachability in MD and flooding is due to collisions. As opposed to the gradient-based DD where point-to-point packet losses are detected and recovered by the underlying MAC protocol, broadcast-based MD and flooding will have to avoid collisions to reduce the amount of packet losses. In view of improving the reachability, we implement a random wait mecha-nism. Every node waits for a random amount of time, uniformly selected within a pre-defined range, before sending any broadcast-based packets. This mechanism avoids the sending of packets simultaneously and thus will be able to reduce the amount of collisions experi-enced by MD and flooding.

4.4.1. Collision

Fig. 12shows the number of collisions for diﬀerent ran-dom wait ranges. The number of collisions drops drastical-ly as the range increases, but the drop slows down when the range goes beyond 4–6 ms. A larger range implies a larger average waiting time. Having the packets waiting longer in the network will result in a higher degree of end-to-end latency. That is, a smaller range is preferred for a lower overall packet delay. This concave curve suggests that most of the collisions can be avoided with a small range, and there is little marginal beneﬁt to use a large range.

4.4.2. Reachability

Fig. 13 shows the reachability for applying different ranges to the random wait mechanism. The reachability ris-es rapidly within 4 ms, and then remains steady in between 99.8% and 100%. The result echoes the discussion about collisions. We see the number of collisions drops signifi-cantly for the first 4 ms, and this is exactly where the reach-ability rises most evidently. This convex curve suggests that the reachability can be improved to pseudo 100% with a small random wait range, and there is no need of using a larger range. Four milliseconds is empirically the optimal range to use for this set of simulated networks. Given the rate of collision depends on the number of nodes sharing the wireless medium, we think the optimal range value depends on the network density.

4.4.3. Data with acceptable delay

Fig. 14shows the percentage of data packets with lower than 50-ms end-to-end delay. The reason that we choose 50 ms to be the threshold is that 50-ms delay is too short to be distinguishable by human. In other words, data arriv-ing with an extra 50-ms delay are acceptable. The 50-ms 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 10000 Cumulative probability Delay (ms) Latency (static) OPP TPP MD Flooding

Fig. 11. Cumulative distribution of delay time for DD, MD, and ﬂooding.

0 200 400 600 800 1000 1200 1400 0 2 4 6 8 10 12 14 16 Number of Collisions

Random Wait Range Collisions (50 nodes)

MD Random

Fig. 12. Number of collisions with increasing random wait range.

97.5 98 98.5 99 99.5 100 0 2 4 6 8 10 12 14 16 Reachability

Random Wait Parameter Reachability (50 nodes)

MD Random

(9)

data arrival rates are all very close to 100% using the ran-dom wait mechanism. This suggests that the ranran-dom wait mechanism is eﬀective in improving the reachability of MD and the extra delay is too small to be perceptible. 4.5. Impact of dynamics

To be realistic, we simulated two kinds of dynamics in our experiments, mobile and node on–oﬀ. For the mobile simulations, each node moves at 120 s. After a node moves to a new random position, it waits another 120 s to move again. The amount of time that a node takes to move to its new position varies from node to node because of the random new position selection. As a result, the movements of the nodes are not simultaneous.

In the node on–off simulations, each node is turned on and off according to the following probabilistic distribu-tion. At the beginning, all nodes are in the on state for a random period between 5 and 65 s. Then, each node goes into the off state for 25–35 s, and wakes up again for 55– 65 s. All the on or off time durations are uniform randomly selected and this process will continue until the end of the simulations.

These sets of simulations are necessary because the two types of dynamics are common in reality. The mobile type emulates people maneuvering in office buildings where we move, stop to mind some business, and then move again. The on–off type represents sensor nodes operating contin-uously for some time, being stopped for some time to change for new batteries, and then resuming the operation. The choices of the parameters represent the extreme cases to highlight the distinct properties in MD to the other two mechanisms, DD and flooding.

4.5.1. Overhead

Fig. 15 shows the amount of overhead for the static, mobile, and on–oﬀ cases. Note that in the static case, all nodes remain static and on for the entire duration of the

simulations. Figs. 15(a)–(c) show the amount of interest,

data, and total packets, respectively. FromFig. 15(c), we

observe the general trend of ﬂooding experiencing a higher amount of packet overhead in all three cases. The total packet overhead of DD and MD is higher in the mobile case due to the increase in the interest packet overhead (Fig. 15(a)). The DD- or MD-related states in the mobile nodes may be changed as the nodes move. These changes will be propagated by means of the interest or other control 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 2 4 6 8 10 12 14 16 Percentage

Random Wait Range Data Percentage (50 nodes)

MD Random

Fig. 14. Percentage of data packets arriving within 50 ms with increasing random wait range.

0 2000 4000 6000 8000 10000 12000 on-off mobile static P a cke t # Interest Overhead OPP TPP MD Flooding 0 5000 10000 15000 20000 25000 30000 35000 on-off mobile static Packet # Data Overhead OPP TPP MD Flooding 0 5000 10000 15000 20000 25000 30000 on-off mobile static Packet # Total Overhead OPP TPP MD Flooding a b c

Fig. 15. Overhead for static, mobile, and on–oﬀ cases. (a) Interest packet overhead. (b) Data packet overhead. (c) Total (interest + data) packet overhead.

(10)

packets which results in a higher amount of interest

over-head. Fig. 15(b) shows that the data packet overhead of

the on–off case is lower. This is because certain critical nodes in the network might be turned off and a significant amount of data fails to traverse through the network. 4.5.2. Reachability

Fig. 16 shows the reachability results of the static, mobile, and on–oﬀ cases. In the mobile case, the abilities of OPP and TPP drop signiﬁcantly. MD’s reach-ability is reduced to 85%, which is about 20% higher than TPP (65%) and 10% higher than OPP (75%). This is because MD sends data in multiple paths. When the nodes are mobile, the probability is higher to get the data through the network to the sink with multiple paths.

Note that TPP performs worse than OPP in the mobile case. The difference lies in the path update frequency. OPP reconfigures data path at the time of periodic interest, but TPP does so at the time of periodic exploratory data. The interval of periodic exploratory data is twice as much as the periodic interest. Therefore, TPP results in a lowest reachability in this case. If we shortened the time of periodic exploratory data, there will be a higher reachability, but the data packet overhead will also increase. From this set of results, we can see the advantages of MD over OPP and TPP in terms of reachability in mobile sensor networks. Note that the reachability of flooding is close to 100%. The result suggests that if the data are absolutely critical and energy is not a constraint, flooding is the best choice for reliable data dissemination. Otherwise, MD provides as a reliable solution for environment that energy resource is limited.

In the on–off case, we observe similar results. But the performance is unsatisfactory. Even the flooding mecha-nism manages only 85% reachability. This is because of the potential of broken paths, or even a disconnected network, when certain bottleneck nodes are turned off. Such conditions result in a lower overall reachability.

However, we think that a good deployment strategy may compensate for such situations. We find that OPP performs the worst in the on–off case. In some sense, the on–off type dynamic presents greater challenge to reliable delivery of data and OPP, or single-path

mecha-nisms, is less suitable for networks with extreme

dynamics. 4.5.3. Latency

Fig. 17 shows the cumulative probability of data delivery latency for the static, mobile, and on–oﬀ

cases. The static case result, Fig. 17(a), is the same

as Fig. 11 shown in Section 4.3. The ﬁgure is pre-sented again here only for the reference. We observe, in Fig. 17(b), that MD has the least degree of latency yet. Note that ﬂooding performs better than OPP com-pletely. This is because OPP has to retransmit a packet many times as a result of high frequency of packet loss and collision in mobile networks. The same reason in OPP and lost data packets result in the poor

perfor-mance of TPP. Fig. 17(c) shows similar result as well.

These results indicate that MD is a better solution for applications with requirements of restricted latency in dynamic network.

4.6. Summary and selection guidelines

Tables 3 and 4summarize our comparison of MD, DD, and ﬂooding in the static and dynamic cases. Note that the term DD here refers to the best of OPP and TPP in each case. We ﬁnd that the choice of data dissemination mechanism is application requirement and deployment environment dependent. Based on the experimental results, we are able to derive the following selection guidelines:

• For static and stable wireless sensor networks, DD is the best choice.

• For dynamic wireless sensor networks with energy con-straint, MD is a more plausible solution with a balanced performance in energy eﬃciency, reliability, and data latency.

• For networks that require absolute reliability, ﬂooding with continuous power supply is a better option. 5. Conclusion

Inspired by the physics of magnets and nails, we pro-pose magnetic diffusion, a simple and yet efficient dissem-ination mechanism for mission-critical data. MD is able to identify all the shortest paths available from the data source to sink. By transmitting data over the multiple shortest paths, MD performs well in the timeliness and reli-ability of data delivery while the overhead and energy con-sumption of MD is kept low. These properties are confirmed by the simulations. Therefore, we conclude that MD is particularly suitable for sensor network applications 0 10 20 30 40 50 60 70 80 90 100 on-off mobile static Reachability OPP TPP MD Flooding

(11)

in healthcare and workplace safety. For these applications, the timeliness and reliability of data, and the energy eﬃciency of the system are all required properties.

Acknowledgments

This work is sponsored by National Science Council of Taiwan under Contract 2218-E-002-147, NSC-93-2218-E-002-144, and NSC-93-2213-E-002-121. The authors would also like to acknowledge the support from the Intel Digital Health Center grant.

References

[1] Chalermek Intanagonwiwat, Ramesh Govindan, Deborah Estrin, Directed diﬀusion: a scalable and robust communication paradigm for sensor network, in: Proceedings of the Sixth Annual International Conference on Mobile Computing and Networking (MobiCOM ’00), August 2000, Boston, MA.

[2] John Heidemann, Fabio Silva, Deborah Estrin, Matching data dissemination algorithms to application requirements, in: Proceedings of the ACM SenSys Conference, Los Angeles, California, USA, ACM, New York, 2003, pp. 218–229.

[4] Wendi Rabiner Heinzelman, Anantha Chandrakasan, Hari Bala-krishnan, Energy-eﬃcient communication protocol for wireless mic-rosensor networks, in: Proceedings of the 33rd Hawaii International Conference on System Sciences, vol. 8.

[5] Seema Bandyopadhyay, Edward Coyle, An energy-eﬃcient hierar-chical clustering algorithm for wireless ad hoc networks, in: Proceedings of the 22nd Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), 2003. [6] Ossama Younis, Sonia Fahmy, Distributed clustering in ad-hoc

sensor networks: a hybrid, energy-eﬃcient approach, in: Proceedings of the 23nd Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), 2004.

[7] S.D. Servetto, G. Barrenechea, Constrained random walks on random graphs: routing algorithms for large scale wireless sensor networks, in: Proceedings of the 1st AC International Workshop on Wireless Sensor Networks and Applications (WSNA), Atlanta, GA, September 2002. [8] David Braginsky, Deborah Estrin, Rumor routing algorithm for

sensor networks, in: Proceedings of the 1st AC International Workshop on Wireless Sensor Networks and Applications (WSNA), Atlanta, GA, September 2002.

[9] Yan Yu, Ramesh Govindan, Deborah Estrin, Geographical and Energy Aware Routing: a recursive data dissemination protocol for wireless sensor networks, Technical Report UCLA/CSD-TR-01-0023, UCLA Computer Science Department, May 2001.

[10] Ananth Rao, Christos Papadimitriou, Scott Shenker, Ion Stoica, Geographic routing without location information, in: Proceedings of the 9th Annual International Conference on Mobile Computing and Networking, San Diego, CA, USA, September 14–19, 2003. [11] Hyung Seok Kim, Tarek F. Abdelzaher, Wook Hyun Kwon,

Minimum-energy asynchronous dissemination to mobile sinks in wireless sensor networks, in: Proceedings of the 1st International Conference on Embedded Networked Sensor Systems, Los Angeles, California, USA, November 05–07, 2003.

[12] CFan Ye, Haiyun Luo, Jerry Cheng, Songwu Lu, Lixia Zhang, A two-tier data dissemination model for large-scale wireless sensor networks, in: Proceedings of the 8th Annual International Conference on Mobile Computing and Networking, Atlanta, Georgia, USA, September 23–28, 2002. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 10000 Cumulative probability Delay (ms) Latency (static) OPP TPP MD Flooding 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 10000 Cumulative probability Delay (ms) Latency (mobile) OPP TPP MD Flooding 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 10000 Cumulative probability Delay (ms) Latency (on-off) OPP TPP MD Flooding b a c

Fig. 17. Cumulative distribution of delay time for (a) static, (b) mobile, and (c) on–oﬀ cases.

Table 3

Comparison of MD, DD, and ﬂooding in static case

Best Fair Worst

Overhead DD MD Flooding

Reliability DD Flooding MD

Latency MD DD Flooding

Table 4

Comparison of MD, DD, and ﬂooding in dynamic case

Best Fair Worst

Overhead DD MD Flooding

Reliability Flooding MD DD

(12)

[13] Chieh-Yih Wan, Andrew T. Campbell, Lakshman Krishnamurthy, PSFQ: a reliable transport protocol for wireless sensor networks, in: Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications, WSNA 2002, September 2002. [14] Y. Sankarasubramaniam, O.B. Akan, I.F. Akyildiz, ESRT:

Event-to-sink reliable transport in wireless sensor networks, in: Proceedings of the ACM MobiHoc’03, Annapolis, Maryland, USA, June 2003. [15] Omprakash Gnawali, Mark Yarvis, John Heidemann, Ramesh

Govindan, Interaction of retransmission, blacklisting, and routing metrics for reliability in sensor network routing, in: Proceedings of the First IEEE International Conference on Sensor and Ad hoc Com-munications and Networks, SECON, 2004.

[16] Sukun Kim, R. Fonseca, D. Culler, Reliable transfer on wireless sensor networks, in: Proceedings of the First IEEE International Conference on Sensor and Ad hoc Communications and Networks, SECON, 2004.

[17] Bret Hull, Kyle Jamieson, Hari Balakrishnan, Techniques for mitigating congestion in sensor networks, in: Proceedings of the ACM SenSys 2004, Baltimore, MD, November 2004.

[18] Chieh-Yi Wan, Shane B. Eisenman, Andrew T. Campbell, CODA: COngestion detection and avoidance in sensor networks, in: Proceedings of First ACM Conference on Embedded Networked Sensor Systems, SenSys 2003, Los Angeles, November 5–7, 2003, pp. 266–279.

[19] Yaling Yang, R. Kravets, Distributed QoS guarantees for realtime traﬃc in ad hoc networks, in: Proceedings of the First IEEE International Conference on Sensor and Ad hoc Communications and Networks, SECON, 2004.

[20] Douglas De Couto, Daniel Aguayo, John Bicket, Robert Morris, A high-throughput path metric for multi-hop wireless routing, in: Proceedings of the 9th Annual International Conference on Mobile Computing and Networking, MobiCom 2003, September 2003, pp. 134–146.

[21] Baruch Awerbuch, David Holmer, Herbert Rubens, High throughput route selection in multi-rate ad hoc wireless networks, in: Proceedings of the Wireless On-Demand Network Systems, First IFIP TC6 Working Conference, WONS 2004, January 21–23, 2004.

[22] S.R. Medidi, K. Vik, QoS aware source-initiated ad-hoc routing, in: Proceedings of the First IEEE International Conference on Sensor and Ad hoc Communications and Networks, Santa Clara, CA, USA, October 2004.

[23] Budhaditya Deb, Sudeept Bhatnagar, Badri Nath, ReInForM: Reliable information forwarding using multiple paths in sensor networks, in: 28th Annual Conference on Local Computer Networks, LCN, 2003.

[28] J.M. Kahn, R.H. Katz, K.S.J. Pister, Mobile networking for smart dust, in: Proceedings of the ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom99), Seattle, WA, August 17–19, 1999.

[29] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, K. Pister, System architecture directions for networked sensors, in: ACM SIGOPS Operating Systems Review, vol. 34, No. 5, 2000, pp. 93–104. [30] L. Breslau, D. Estrin, K. Fall, S. Floyd, J. Heidemann, A. Helmy, P.

Huang, S. McCanne, K. Varadhan, Y. Xu, H. Yu, The VINT Project, Advances in network simulation, IEEE Computer 33 (5) 2000 59–67.

Hsing-jung Huang received the B.S. degree in Computer Science and Information Engineering from National Taiwan University, Taiwan, in 2004. He is currently a graduate student of Department of Electrical Engineering of National Taiwan University. His research activity mainly focuses on routing protocols in wireless sensor networks.

Ting-Hao Chang received the B.S. degree in Computer Science and Information Engineering from National Taiwan University, Taiwan, in 2004. He is currently a graduate student of Department of Electrical Engineering of National Taiwan University. His research activity mainly focuses on reliable transfer over wireless sensor networks.

Shu-Yu Hu received the B.S. degree in Computer Science and Information Engineering from National Taiwan University (NTU), Taiwan, in 2004. He is currently a graduate student of Department of Electrical Engineering of the National Taiwan University. His research activity mainly focuses on sensor network deployment issue.

Polly Huang received her Ph.D. (1999) and M.S. (1997) in Computer Science from University of Southern California, and her B.S. (1993) in Mathematics from National Taiwan University. In February 2003, she joined the Department of Electrical Engineering of the National Taiwan University at which she currently holds an assis-tant professor position. Prior to joining NTU, she worked as a post-doctoral research scientist at the Computer Engineering and Networks Laboratory (TIK) of the Swiss Federal Institute of Technology (ETH) Zurich. Dr. Huang is on the editorial board of Journal of Communications and Networks. She is a member of the ACM and IEEE.