You are on page 1of 20

Performance Evaluation of Distributed Computing Paradigms in Mobile Ad Hoc Sensor Networks

Yingyue Xu, Hairong Qi Department of Electrical and Computer Engineering The University of Tennessee, 319 Ferris Hall Knoxville, TN 37996-2100 {yxu4, hqi}@utk.edu

Abstract The emergence of mobile ad hoc sensor networks has brought new challenges to traditional network design. This paper focuses on the study at the application layer. In specic, it compares the performance of two distributed computing paradigms, the client/server-based paradigm and the mobile-agent-based paradigm, through mathematical modeling and simulation. The dierences between the two computing paradigms can be characterized by what is transferred over the network and correspondingly, where data processing is carried out. In the client/server-based computing, clients send data to a server and data processing is done at the server ; while in the mobileagent-based computing, the server dispatches mobile agents which carry executable codes to clients and data processing is done locally. Previous works have shown that the mobile-agent-based paradigm is more appropriate to handle computations in ad hoc sensor networks. However, no simulation work has done to quantitatively measure

This research was supported in part by DARPA under grant N66001-001-8946.

the performance. This paper rst describes how computing is accomplished in the mobile-agent-based paradigm. It then presents a modied mathematical model and uses the execution time as a metric to measure the performance. The simulation is written using GloMoSim. Eight experiments are designed to show the eect of dierent parameters to the performance of the paradigms. Experimental results show that the mobile-agent-based computing does not always perform better than the client/serverbased since the usage of mobile agents also brings overhead. However, in the context of mobile ad hoc sensor networks with hundreds or even thousands of nodes, unreliable communication links, and reduced bandwidth, the mobile-agent-based computing provides solutions to low network latency and reliable data processing.

Keywords: mobile ad hoc sensor networks, distributed computing paradigm, mobileagent-based paradigm, client/server based paradigm, performance evaluation, simulation

Introduction

The advancement in wireless communication and the emergence of economical and portable computing devices have made mobile computing possible. One research issue that has attracted a lot of attention recently is the design of mobile ad hoc network (MANET). MANET is dened as a collection of mobile platforms or nodes where each node is free to move about arbitrarily [12]. It is characterized by frequent host movement, frequent topology changes, easy and fast deployment, and less dependence on network infrastructure. MANET can be used in a wide spectrum of applications, including battleeld communication, emergency services, home networking, personal area networks, etc. Although MANET does not require any centralized control and allows node mobility, which have provided extraordinary exibility, these advantages, on the other hand, also cause problems to traditional network design. We summarize some important issues brought up by MANET in the design of the physical, medium access, network and application layers, as 2

well as possible solutions. At the physical layer, an important issue is power control. The transmission power of nodes needs to be regulated so that it is high enough to reach the intended receiver while minimizing the interference at other nodes [6]. Many iterative power control algorithms [2, 19] have been developed to achieve this goal. The diculty in the medium access scheme lies in the time-varying network topology and the lack of centralized control. Normal schemes like TDMA, FDMA and CDMA are inecient in such networks. Sharing channels in wireless networks also leads to new problems associated with the so-called hidden terminal and exposed node problems [18]. Various protocols based on collision avoidance handshakes have been developed, including MACA [9], MACAW [3], IEEE 802.11 [11], etc. At the network layer, the main problem is routing. Each node in MANET is both an end host and a router. The self-starting behavior and convergence characteristics of existing routing protocols are not good enough to t the needs of ad hoc networks. Various routing protocols have been designed specically for MANET, including DSDV (Destination Sequenced Distance Vector) [14], WRP (Wireless Routing Protocol) [17], AODV (Ad Hoc On-Demand Distance Vector Routing) [15], DSR (Dynamic Source Routing) [8] and TORA (Temporally Ordered Routing Algorithm) [13]. The issue at the application layer is about mobile computing. Computing in a mobile environment is dierent from conventional xed network computing. This is partially because of the movement of the mobile hosts that requires maintaining connection while moving. On the other hand, the relatively unreliable wireless link and the low communication bandwidth also make traditional client/server-based computing model inecient. New approaches such as the client/proxy/server model [10] and the mobile agent model [5, 16] are proposed. This paper focuses on the study at the application layer. In specic, it compares the performance of two distributed computing paradigms, the client/server-based paradigm 3

and the mobile-agent-based paradigm, through mathematical modeling and simulation. We choose a particular kind of MANET - ad hoc sensor networks (SNs) as the subject of our study. An ad hoc SN is a wireless network composed of a large number of sensor nodes. It presents some unique challenging problems because of the large number of nodes, more densely node deployment and more proneness to node failure [1]. In ad hoc SNs, raw data are collected at each sensor node and then processed for further usage, such as target detection, classication and tracking [20]. Because the number of sensor nodes is usually large, using the client/server-based computing will create huge network trac, where each sensor (the client) needs to transfer all the raw data to a centralized processing center (the server). The mobile-agent-based computing [16], on the other hand, transfers the executable code from the processing center to local nodes and the processing is done locally. This paper rst discusses the mobile-agent-based computing paradigm in Sec. 2. It then presents, in Sec. 3, a modied mathematical model that measures the performance using the execution time. Sec. 4 designs eight experiments that simulate dierent dierent scenarios in order to analyze conditions under which one computing paradigm performs better than the other. Sec. 5 draws the conclusions.

Mobile-agent-based Computing Paradigm for Ad Hoc Sensor Networks

An ad hoc sensor network (SN) forms a typical distributed environment and the client/server paradigm has been one of the most popular models adopted in distributed computing [4]. In this paradigm, a server oers a set of services, resources, and know-how needed for service execution. The client requests the execution of a service. As a response, server performs the requested service by executing the corresponding know-how and accessing the involved resources in the server. Figure 1(a) illustrates the client/server-based paradigm. Although

widely used, the disadvantages of this model are also dramatic, especially for ad hoc SNs [7, 16]. First of all, the mobile client may suer from a slow and unpredictable response time, especially when the number of client is huge or when large server replies such as query results are transmitted through wireless links. Secondly, in some particular applications, such as data processing or data fusing, large amounts of data are moved from the clients to the server. This will require many round trips over the network in order to complete one transaction. Each trip creates network trac and consumes bandwidth. In a system with many clients and/or many transactions, the total bandwidth requirements may exceed available bandwidth, resulting in poor performance of the system. Thirdly, the design of a client/server-based system requires precise consideration of the network trac, the number of clients and servers, transaction volumes, etc. If the estimates are inaccurate, the performance of the system will suer.

Processing Element

Processing Element

mobile agent A

mobile agent B

sensor

sensor

(a) Client/server-based computing.

(b) Mobile-agent-based computing.

Figure 1: Dierent computing paradigms.

On the contrary, in the mobile-agent-based paradigm, the service know-how is owned by the server, but most of the resources are located at the clients. The server sends out mobile agents carrying service know-how. The mobile agents complete the service using resources available at the clients. Figure 1(b) shows the framework of the mobile agent model. The mobile agent model has many advantages over the client/server model: (1) 5

Network bandwidth requirement is reduced. Instead of passing large amounts of raw data over the network through several round trips, only the agent with small size is sent. This is especially important for real-time applications and where the communication is through lowbandwidth wireless connections. (2) Stability. Mobile agents can be sent when the network connection is alive and return results when the connection is re-established. Therefore, the performance of mobile agent is not much aected by the reliability of the network. We dene mobile agent as an entity of four attributes: identication, itinerary, data space, and service know-how, as shown in Figure 2. Identication uniquely identies each mobile agent. Data space is the agents data buer which carries a partially integrated result. This result should provide progressive accuracy as the agent migrates from node to node. Itinerary is the route of migration. It can be xed or dynamically determined based on the current network status. Service know-how is the processing task (or execution code) carried with the agent.
service knowhow

itinerary

data space

(160.36.0.0, 1) identification

Figure 2: Mobile agent components.

A temporal and spatial presentation of the mobile agent life cycle is given in Figure 3. Let ttrans be the time used for transferring data from one node to the other, toh the overhead time used to create, dispatch, and receive the mobile agent, and tproc the data process time. A mobile agent is created at node Si using toh amount of time. It then migrates to Sj in ttrans amount of time. After Sj receives the mobile agent, it spends toh to read data from the mobile agent and tproc to do data processing. Then the agent migrates to Sk . After the same procedure, it nishes migration and returns to Si . 6

TIME S P A C E

  



  

 

"   



MOBILE AGENT CREATION

RESULT RETURNED

Figure 3: Mobile agent migration in time and space. Take target classication in an unattended ground sensor system as an example. Assume the system is used in a battleeld environment where some thousands of sensors are deployed along the roads to detect the type of vehicles running on the roads. Sensors may be xed or carried by soldiers or vehicles. Each sensor platform can collect seismic, acoustic or IR signals. By studying the signals collected at each sensor platform, one can classify a vehicle as a particular type using pattern recognition. We can use mobile agent to carry out the classication fusion task. The sensor network is partitioned into several clusters. Each cluster consists of several sensors. A mobile agent is sent from the server node and migrates around sensors in the same cluster. Once arriving at a sensor, it integrates the data of that sensor with the result it carries, generates the partially integrated results, and continues migrating. Field demo results show it can classify targets with very high accuracy [20].

Performance Evaluation

Even though the mobile-agent-based computing is promising and seems more appropriate for ad hoc SNs, there has not been any simulation work done to actually compare the performance of these two computing paradigms. The mobile-agent-based computing may

not always perform better since mobile agents also introduce overhead, which mainly comes from the agent creation and dispatch time. On the other hand, the client/server-based computing needs to transfer data les to the processing center, which also causes overhead due to le accesses. Therefore, a performance evaluation of these two computing paradigms under dierent conditions is essential.

3.1

Mathematical models

The execution time is used as a metric to evaluate the performance. It is the time spent to nish a processing task. In the mobile-agent-based paradigm, it starts from the time a mobile agent is created to the time the mobile agent returns with results. In the client/server-based paradigm, it is from the time the clients send out requests to the time the clients receive results. Execution time composes of three parts, data transfer time (ttrans ), overhead time (toh ), and data process time (tproc ). For the client/server paradigm, the execution time is determined by the network transfer rate vn , the data processing rate vd , the data le size sf (the size of the raw data each node collects), the overhead of le access of (the time used to read and write a data le), the number of sensor nodes p, and the balance between the number of agents m and the number of sensor nodes n that each agent migrates (the server is not included). Notice that p = m n. Thus, the data transfer time is ttrans = mnsf /vn ; the overhead time is toh = 2mnof (assuming the time used to read and write the data le is the same); and the data processing time is tproc = mnsf /vd . Therefore, the total execution time for the client/server paradigm is tcs = mnsf mnsf + 2mnof + vn vd (1)

For the mobile agent paradigm, the execution time is determined by the network transfer rate vn , the data processing rate vd , the mobile agent size sa , the overhead of mobile agent

oa , the number of sensor nodes p, number of agents m, and the number of sensor nodes n that each agent migrates. Again, p = m n. Thus, the time used to transfer agents is ttrans = (m + n)sa /vn , since mobile agents use nsa /vn to transfer simultaneously and they return to server after nishing the task serially, which takes msa /vn additional time for m mobile agents; the agent overhead time is toh = 2(m + n)oa , since it takes 2moa for the server to send and receive m mobile agents, and 2noa for the nodes to send and receive each mobile agent; and the the time used to run the executable code locally is: tproc = (m + n)sa /vd . Therefore, the total execution time for the mobile agent paradigm is: (m + n )s a (m + n )s a + 2(m + n)oa + vn vd

tma =

(2)

3.2

Simulation model

The mathematical model derived above characterizes the behavior of these two computing paradigms to some extent, but it is dicult to get any quantitative measurement, especially when modeling the data transfer time, where retransmission and error control are not considered. Unfortunately, these factors occur quite often in ad hoc SNs because of the wireless link. Therefore, we propose to use simulation tools for more accurate estimation of the data transfer time ttrans . The model used to calculate the execution time is: mnsf (m + n)sa , tma = ttransma + 2(m + n)oa + vd vd

tcs = ttranscs + 2mnof +

(3)

where ttranscs and ttransma are obtained from simulation.

3.3

Simulation tool

After careful study of the current existing simulation software, we choose to use GloMoSim [21], developed by UCLA, to simulate ad hoc SNs. GloMoSim is designed specically for

wireless network simulation. It is scalable. The usage of the PARSEC simulation environment [21] enables quick and accurate simulation of a network with thousands of nodes which is typical in an ad hoc SN. Following the OSI seven layer network architecture, GloMoSim divides a network into ve layers: application layer, transport layer, network layer, MAC layer, and radio and mobility layer. In each layer, GloMoSim incorporates dierent models and protocols.

3.4

Experimental parameter setup

In our simulation, the basic network consists of wireless nodes randomly distributed within a 100m by 100m square area. Node placement can vary depending on an intended task. The mobility model we use is the random waypoint model where nodes randomly choose a destination and move at a speed of 10m/s. Once it arrives at the destination, it pauses for 5 seconds and then continues moving. Other parameters that characterize the network include the MAC layer protocol, the routing protocol, the transport layer protocol, the transmission rate, the propagation model, the propagation fading model, and the transmission range which is determined by the transmission power, radio frequency, and radio receive threshold. Parameter setup in the basic network is shown in Table 1.
Network area Node placement Mobility Transmission Radio frequency range Transmission power Radio receive threshold Transmission rate Propagation model Propagation fading model MAC layer Routing protocol Transport layer protocol 100m by 100m Random Random waypoint, 10m/s with 5s pause 2500MHz 15dbm -81dbm 2Mbps Two ray Rayleigh IEEE 802.11, RTS/CTS enabled Bellman-Ford TCP

Table 1: Parameter setup for the basic network. 10

Parameters that determine the data processing time (tproc ) and the overhead (toh ) in dierent computing paradigms include the number of nodes (p) (in the mobile-agent based paradigm, p is the multiplication of the number of mobile agents (m) and the number of nodes migrated by each agent (n)), the data le size (sf ), the mobile agent size (sa ), the mobile agent overhead (oa ), the le access overhead (of ), the network transfer rate (vn ), and the data processing rate (vd ). The parameter values in the basic network are shown in Table 2.
n 10

m
1

sf
10Kb

sa 1Kb

of 0.0125s

vn 0.05s

vd 2Mbps

oa 100Mbps

Table 2: Computing paradigm related parameter setup for the basic network.

Experiments and Simulation Results

In each experiment, we change one parameter of the basic network and keep all the others unchanged. We design eight experiments to test the eect of dierent parameters to the execution time. The factors include the number of nodes p, the number of mobile agents m, the node speed, the transmission rate vn , the transmission range, the size of data le sf , the size of mobile agent sa , the overhead ratio of /oa , and dierent protocols and models used at dierent layers.

4.1

Experiment 1 - The eect of the number of nodes (p)

In this experiment, we change the number of nodes p from 2 to 30. The result is shown in Figure 4. It can be observed that the execution time using both paradigms grows as the number of nodes increases. But the client/server model grows much faster than the mobile agent model. This is because as the number of nodes increases, the server has to deal with more connections requested by the clients at the same time which elongates the execution 11

time. On the other hand, the mobile agent model is less inuenced by the number of nodes because there are much less connections at one time for the mobile agent model. The gure also shows that, for p < 8, the client/server model performs a little better than the mobile agent model. This happens when the mobile agent model needs more connections than the client/server model in order to send and receive mobile agents. It also happens when the overhead of the mobile agent surpasses the overhead of the client/server model. Therefore, in a network with fewer nodes, the client/server model may have less execution time than that of the mobile agent model. However, if the number of node is large, the mobile agent model will perform better.
70 client/server based mobileagentbased 60

50 Execution time (seconds)

40

30

20

10

10

15 Number of node

20

25

30

Figure 4: The eect of the number of nodes p.

4.2

Experiment 2 - The eect of the number of mobile agents (m)

In this experiment, we x the node number at 100 and observe the eect of the number of mobile agents to the performance of dierent computing paradigms. Without loss of generality, we assume each agent migrates the same number of nodes. We expect a constant prole from the client/server-based computing since it is irrelevant to the number of mobile agents. We can see from Figure 5 that the mobile agent model always performs better than the client/server model because the node number is large. Interestingly, the execution time 12

of the mobile agent model decreases as the number of mobile agents increases and reaches the lowest point when there are ve mobile agents. Then the execution time begins to increase. This is because more mobile agents will reduce the number of nodes each agent migrates, thus reduce the execute time. But more mobile agents also cause more connections and more overheads. So a proper number of mobile agents can make the model perform more eciently. We can see more clearly from Figure 6 that the mobile agent model can save at most 92% of execution time when m = 5.
900

95

800

90
700

Execution time (seconds)

600

Execution time reduction ratio(%)


client/server based mobileagentbased 0 5 10 15 20 25 30 Number of mobile agents 35 40 45 50

85

500

400

80

300

200

75

100

70

10

15

20 25 30 Number of mobile agents

35

40

45

50

Figure 5: The eect of the number of mobile Figure 6: The execution time reduction ratio agent. (tcs tma )/tcs .

4.3

Experiment 3 - The eect of the node speed

In this experiment, we change the mobility feature of the node from static (0m/s) to a moving speed of 100m/s and keep all the other parameters in the basic network unchanged. We observe from Figure 7 that the execution time for both paradigms grows as the speed of the node increases. This is because it is more dicult to maintain connection when nodes are moving fast, which is when more drops and retransmissions usually occur. However, no matter how fast the node moves, the mobile-agent-based paradigm always performs better.

13

4.4

Experiment 4 - The eect of transmission rate (vn )

In this experiment, we x all other parameters in the basic network, but change the transmission rate from 0.1Mbps to 2Mbps. Figure 8 shows the execution time decreases as the transmission rate increases. However the mobile agent model is less aected by the transmission rate than the client/server model. The mobile agent model performs much better in low transmission rate as well.
3.6 client/server based mobileagentbased 3.4

50 client/server based mobileagentbased

45

40
3.2 Execution time (seconds)

35 Execution time (seconds)

30

25

2.8

20

2.6

15

10
2.4

2.2

10

20 35 Mobility (m/s)

50

75

100

0.2

0.4

0.6

0.8 1 1.2 Transmission rate (bits/s)

1.4

1.6

1.8 x 10

2
6

Figure 7: The eect of mobility v .

Figure 8: The eect of transmission rate vn .

4.5

Experiment 5 - The eect of the node transmission range

In this experiment, we x all other parameters in the basic network, but change the transmission range from 6m to 59m. Figure 9 shows the eect of transmission range. The execution time decreases as the transmission range increases. The explanation is straightforward. As the transmission range increases, less hops are needed in routing, thus reducing the execution time. We also observe that when the transmission range goes beyond 15m, very small eect is done to the execution time. This is probably because the node density is less than 15m. Again, the mobile agent model is less aected than the client/server model.

14

20 client/server based mobileagentbased

18

16

Execution time (seconds)

14

12

10

10

20

30 Transmission range (m)

40

50

60

Figure 9: The eect of the node transmission range.

4.6

Experiment 6 - The eect of data size (sf ) and mobile agent size (sa )

In this experiment, we change the size of the data le sf , but x the other parameters and let sa = 1k . Figure 10 shows the result. We can see that the execution time using the mobileagent based computing is constant since the data are located at the local nodes, only a xed amount of results are transferred. When the data size is less than 7KB, the client/serverbased computing performs better than the mobile-agent based. However, the larger the data size, the more advantageous the mobile-agent based paradigm. This is because as sf increases, more data need to be transferred, thus increasing the time used by the client/server model. Next, we x sf = 10k and change sa , as shown in Figure 11. We can see that the execution time of the mobile agent increases as sa increases. When sa > 8k , the client/server model performs better than the mobile agent model.

4.7

Experiment 7 - The eect of the overhead ratio (of /oa )

In this experiment, we x all other parameters in the basic network, and observe the eect of the overhead ratio of /oa (between 0.1 and 4.0) to the performance of dierent comput15

10 client/server based mobileagentbased 9

3.5

Execution time (seconds)

Execution time (seconds)

2.5

1.5

client/server based mobileagentbased

10

15

20

25 30 Size of data (KB)

35

40

45

50

4 5 6 Size of mobile agent (KB)

10

Figure 10: The eect of the data size sf .

Figure 11: The eect of mobile agent size sa .

ing paradigms. We can see from Figure 12 that when the ratio is greater than 0.5, the client/server-based computing starts to perform worse than the mobile-agent model since the larger the of , the more the execution time.
7.5 client/server based mobileagentbased 7

6.5

Execution time (seconds)

5.5

4.5

3.5

0.5

1.5 2 2.5 Overhead ratio (Of/Oa)

3.5

Figure 12: The eect of the overhead ratio of /oa .

4.8

Experiment 8 - The eect of dierent protocols and models

In this experiment, we x the other parameters in the basic network, but change the protocols and models used in order to show their inuence to the execution time. Figure 13 shows the 16

eect of dierent MAC layer protocols. While the client/server model performs the best using the 802.11 protocol, the mobile agent model performs the best using the CSMA protocol. We also observe the eect of using dierent routing protocols. From Figure 14, we can see that both paradigms perform the best when using the WRP routing protocol. Figure 15 shows that a uniform node deployment saves more execution time than a random deployment since it is easier to maintain connection in the uniform deployment environment. Figure 16 demonstrates that the Ricean fading distribution performs better than the Rayleigh fading distribution. The reason is that the success rate to receive a radio signal is higher in Ricean fading distribution for the line-of-sight path.
3.5
5 4.5

3 2.5 2 1.5 1 0.5 0


CSMA MACA client/server based 802.11 mobileagentbased

Execution time (seconds)

3.5 3 2.5 2 1.5 1 0.5 0

Execution time (seconds)

AODV DSR LAR1 WRP BELLMANFORD client/server based mobileagentbased

Figure 13: Eect of MAC layer protocols.


3.5 3 Execution time (seconds) 2.5 2 1.5 1 0.5 0 Random client/server based Uniform mobileagentbased Execution time (seconds) 3.5 3 2.5 2 1.5 1 0.5 0

Figure 14: Eect of routing protocols.

Rayleigh client/server based Ricean mobileagentbased

Figure 15: Eect of node deployment.

Figure 16: Eect of fading distribution.

17

Conclusions

This paper focuses on the discussion of two distributed computing paradigms designed in ad hoc SNs. In specic, it compares the performance of the client/server-based paradigm and the mobile-agent-based paradigm. We present a modied mathematical model which provides more realistic measurement of the data transfer time through simulation. We design and implement several scenarios based on GloMoSim to investigate the eect of dierent parameters to the performance of the computing paradigms. We show that the mobile-agentbased model does not always perform better than the client/server-based model. However, in the context of ad hoc SNs with hundreds or even thousands of nodes, unreliable communication links, and reduced bandwidth, the mobile-agent-based computing provides solutions to low network latency. From experiments, we also nd that the mobile agent model is less aected by network parameter changes. Stability is its big advantage as well. The experiments provide us a good reference when designing computing paradigms in ad hoc SNs. If there are only several nodes, or the raw data size is not large, or we cannot increase overhead ratio between of and oa , then the client/server paradigm may be a better choice. On the other hand, if we decide to use the mobile agent model for the application, then we have to choose carefully the number of the mobile agents in order to minimize the execution time. If possible, we should reduce the node speed, increase the transmission rate and the transmission range. We may also want to use CSMA protocol in MAC layer and WRP as routing protocol. In some situation, we may deploy nodes uniformly or get line-of-sight path among nodes to further improve the network performance.

References
[1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks: a survey. Computer networks, 38:393422, 2002. 18

[2] N. Bambos, S. Chen, and G. Pottie. Radio link admission algorithms for wirelessd networks with power control and active link quality protection. In Proc. IEEE INFOCOM, volume 1, pages 97104, Apr. 1995. [3] V. Bharghavan, A. Demers, S. Shenkar, and L. Zhang. MACAW: a media access protocol for wireless lans. In Proc. SIGCOMM94 Conf. on Communications Architectures, Protocols and Applications, pages 212225, Aug. 1994. [4] A. Fuggetta, G. P. Picco, and G. Vigna. Understanding code mobility. IEEE Trans. on Software Engineering, 24(5):342361, 1998. [5] D. Gilbert et al. Intelligent agent strategy. Technical report, IBM Corporation, Research Triangle Park, NC, 1995. [6] P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Trans. on Information Theory, 46(2):388404, Mar. 2000. [7] A. Helal et al. Any Time, Anywhere Computing. Kluwer Academic Publishers, 1999. [8] D. B. Johnson and D. A. Maltz. Mobile Computing. Kluwer Academic Publishers, 1996. [9] P. Karn. MACA: a new channel access method for packet radio. In Proc. 9th

ARRL/CRRL Amateur Radio Computer Networking Conf., pages 134140, Sep. 1990. [10] M. Kojo, K. Raatikainen, and T. Alanko. Connecting mobile workstations to the internet over a digital cellular telephone network. The Kluwer international series in engineering and computer science. VLSI, computer architecture and digital signal processing, 353:253270, 1996. [11] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specications. IEEE Standard 802.11-1997. IEEE Computer Society LAN MAN Standards Committee, 1997.

19

[12] J. Macker and M. S. Corson. Mobile ad hoc networking and the IETF. ACM Mobile Networks and Applications Journal, 3(1):101119, Jan. 1998. [13] V. Park and M. S. Corson. A highly adaptive distributed routing algorithm for mobile radio networks. ACM Wireless Networks Jounal, 1:6181, Jan. 1995. [14] C. Perkins and P. Bhagwat. Highly dynamic destination-sequenced distance vector routing (DSDV) for mobile compute. In Proc. of ACM SIGCOMM, pages 234244, Sep. 1994. [15] C. Perkins and E. Royer. Ad-hoc on-demand distance vector routing. In Proc. of 2nd IEEE Workshop on Mobile Computing Systems and Applications, pages 17021706, Feb. 1999. [16] H. Qi, S. S. Iyengar, and K. Chakrabarty. Multi-resolution data integration using mobile agents in distributed sensor networks. IEEE Trans. on Syst., Man, and Cybern. Part C: Applications and Reviews, 31(3):383391, Aug. 2001. [17] J. Raju and J. J. Garcia-Luna-Aceves. A comparison of on-demand and table-driven routing for ad hoc wireless networks. In Proc. of IEEE ICC, pages 17021706, Jun. 2000. [18] C. K. Toh. Ad Hoc Mobile Wireless Networks: Protocols and Systems. PH, 2002. [19] S. Ulukus and R. Yates. Stochastic power control for cellular radio systems. IEEE Trans. on Communications, 46:784798, Jun. 1998. [20] X. Wang, H. Qi, and S. S. Iyengar. Collaborative multi-modality target classication in distributed sensor networks. Accepted by Information Fusion, Jul. 2002. [21] X. Zeng, R. Bagrodia, and M. Gerla. GloMoSim: a library for parallel siumulation of large-scale wireless networks. In Proc. of the 12th workshop on Parallel and Distributed Simulations (PADS 98), pages 154161, May 1998. 20

You might also like