The Impact of Packet Fragmentation and Reassembly in

Journal of Computing and Information Technology - CIT 21, 2013, 2, 97–107doi:10.2498/cit.1002195

97

The Impact of Packet Fragmentationand Reassembly in ResourceConstrained Wireless Networks

James Pope and Robert SimonDepartment of Computer Science, George Mason University, Fairfax, Virginia, United States

Low-power and lossy (LLN) networks, including Wire-less Sensor Networks, provide an important buildingblock for Internet of Things (IoT) technology. Theresource constraints associated with LLN devices neces-sitate the redesign of a number of basic Internet protocols.This paper evaluates the effect of packet fragmentationon the performance of protocols redesigned for LLNsystems. We focus on the important class of tree-basedLLNs that exhibit both one-to-many and many-to-onecommunication patterns. In particular, we measurethe impact of fragmentation on these communicationprotocols. Our results demonstrate that excessive packetfragmentation has a tremendously negative impact oncommunication within a tree-based LLN system. Thiswork provides a guideline for IoT engineers in techniquesfor avoiding these damaging effects.

Keywords: Internet of Things communication, Internetprotocols over embedded systems, low-power and lossynetworks

1. Introduction

The Internet of Things (IoT) refers to a collec-tion of technologies designed to interconnectphysical devices with the Internet [1]. A sys-tem or application is considered part of the IoTif it can run the IP protocol stack. One par-ticularly important type of IoT system is char-acterized by topologies populated by resource-constrained nodes. This system class, includingwireless sensor networks, are referred to as LowPower and Lossy (LLN) networks [2, 3, 4]. Weconsider an LLN to be tree-based if it has a base-station or gateway that coordinates the activitiesof the devices in the network [5]. Tree-basedLLNs have two types of communication pat-terns. The first is one-to-many, where the base

station sends control or management packets toeach of the devices in the system. This mode re-quires system-wide broadcasting [6]. The sec-ond is many-to-one, where each device trans-mits its packets up the tree to the gateway. Thismode is sometimes called convergecasting [5].Transforming tree-based LLNs to work withinthe IoT requires the redesign and reimplemen-tation of a number of basic Internet protocols.Our paper examines the performance impact ofthese redesigned protocols on tree-based LLNsystems.

In particular, we focus on the effect that packetfragmentation has on protocols redesigned forLLN applications. Packet fragmentation oc-curs when a device tries to transmit a packetthat is larger than the maximum transmissionunit size, or MTU, allowed by the link layer.When this occurs, the device can either dropthe packet or fragment the packet into severalsmaller packets so that each will be no largerthan the MTU. Using standard IP protocols, ifthe packet is fragmented and not dropped, thenthe original packet can be reassembled when allof the packets are received. The advantage offragmentation is that it offers a way around thelimitations imposed by a small MTU. There are,however, several potential disadvantages. Frag-mentation and reassembly protocols incur addi-tional complexity within the system. Further,since multiple packets now must be transmittedfor each original packet, the likelihood of packetloss increases. The impact of packet loss maybe particularly severe for wireless IoT systems,where resources are heavily constrained[7].

98 The Impact of Packet Fragmentation and Reassembly in Resource Constrained Wireless Networks

Our goal is to precisely quantify the differencesin metrics such as Packet Delivery Ratio (PDR)and network overhead between situationswherefragmentation occurs and where it does not oc-cur. We study this impact by deploying severaltypes of applications in Contiki, a popular andwidely used open source operating system forLLNs and IoT applications [8]. This testbed en-abled us to understand the dynamics and impactof packet fragmentation on both network man-agement and application level performance.

For the one-to-many application, we designedand implementedRadiate, a semi-reliable broad-cast protocol. Radiate has several different op-erating modalities, trading off reliability withmessage complexity and packet overhead. Forthe many-to-one application we modeled a ba-sic, non-aggregating data collection architec-ture, where each node periodically transmits avalue up the tree to the base station. We variedpacket rates at both the base station and interiornodes to study how fragmentation impacted thePDR performance.

The major contribution of this paper is to showthat, using a general purpose LLN operatingsystem, packet fragmentation causes a dramaticdecrease in packet delivery rates for both theone-to-many and the many-to-one communica-tion scenarios. This decrease in delivery ratesoccurs well before the network starts to get con-gested or reaches its transmission capacity. Forthe one-to-many broadcast communication pat-tern, our results show the need, depending on therequired message update rate, to change the re-liability mechanism. For the many-to-one datagathering communication pattern, our resultsshow that either different network fragmenta-tion and reassembly support is required or theapplication must alltogether forbid fragmenta-tion.

The paper proceeds as follows: Section 2 pre-sents background and related work. Section 3describesRadiate, ourmulti-modal semi-reliablebroadcast protocol. Section 4 discusses ourevaluationmethodology and presents the resultsof our experiments. Finally, Section 5 offerssome observations and conclusions.

2. Background and Related Work

We first present some background on our tar-geted network architecture, and then discuss re-lated work.

2.1. Tree-based Resource ConstrainedNetworks

Our work centers on tree-based wireless net-works using low power communication archi-tectures such as those supported by the IEEE802.15.4 specification[9]. This technology isdesigned to support applications such as theSmart Grid, data center power control, indus-trial networks or building and home automation[10, 11, 12, 13]. These LLN systems are pop-ulated by resource constrained devices, haveunreliable communication links and low datarates, and, as explained in the Introduction,are referred to as low-power lossy networks(LLNs)[14].

Interest in tree-based LLNs is partially mo-tivated by the architecture of the Internet-of-Things. IoT systems must support two-way andend-to-end IP enabled communication. Tradi-tional IP protocols failed to address the operat-ing characteristics within a LLN environment.This forced a redesign of several basic protocols[3, 15]. For instance, the 6LoWPAN protocolwas designed to allow the use of the IP layerfrom the Internet protocol stack to be run di-rectly on LLN devices [16].

The 6LoWPAN protocol is designed to be inde-pendent of the data link layer, but is often as-sumed to be running over an 802.15.4 link. Inorder to effectively run in this environment, theprotocol allows a number of options for headercompression. It also supports packet fragmen-tation and reassembly. Packet fragmentationis the process of breaking packets into smallerfragments, and then sending each fragment as aseparate packet. Fragmentation occurs becausedifferent networks connected through the Inter-net can have different maximum sizes or MTUs.For instance, the maximum size of the 802.11WiFi data payload is 2312 bytes, but the maxi-mum size of the data payload carried by manyvariants of the 802.15.4 standard is 104 bytes.Packets are fragmented when they arrive on alink that has a smaller maximum size than the

The Impact of Packet Fragmentation and Reassembly in Resource Constrained Wireless Networks 99

packet itself. In IP, each fragment is retransmit-ted in a separate packet, and the entire packetis reassembled at the end-host, once all of thefragments are received.

The 6LoWPAN protocol is an adaptation layertypically sitting between the 802.15.4 link layerand the network layer performing fragmenta-tion and reassembly [16]. One major differ-ence between fragmentation in IPv4 and IPv6versus fragmentation in 6LoWPAN is that thelatter does not copy IP header information intoeach packet. This means that each device mayonly be able to reassemble fragments originat-ing from one original packet at a time. Frag-ments from other packets would need to bedropped.

Supporting packet fragmentation in a resourceconstrained network is a difficult problem. Inaddition to increased physical layer impairments,higher layers encounter problems handling frag-mentation/reassembly. Considering the many-to-one mode, only the sink would need to re-assemble the packets. Since all messages aredirected towards the sink, simultaneous packetreassembly from multiple senders is necessary.Increasing buffer size or adding additional buf-fers may not be viable solutions for resourceconstrained networks. Careful buffer manage-ment is necessary for devices in these networksto support this paradigm. For instance, Con-tiki’s 6LoWPAN implementation keeps one re-assembly buffer.

The one-to-many mode is effected by fragmen-tation/reassembly in a different way. Sincethe message is intended for all devices in thenetwork, each device needs to reassemble thepacket. Once reassembled, the device can thenretransmit the message to neighbors, eventu-ally flooding the message to all devices. Inresource constrained networks, this is very ex-pensive, particularly from an energy perspec-tive. Contiki’s Rime Network Flood [17] proto-col (one-to-many) does not provide fragmenta-tion/reassembly capability. TheContiki 6LoW-PAN implementation provides a local neigh-bor broadcast fragmentation/reassembly capa-bility, but does not provide the one-to-many ca-pability.

2.2. Performance Impact of Packet Lossin Tree-based Networks

The performance impact, along with mitigationapproaches to the problem of packet loss in LLNwireless networks, has been extensively studied.Thiswork has largely focused on the general im-pact of lossy links and low transmission speedson application level performance, alongwith thetechniques for eliminating some of these effects[18, 19, 20, 21, 22, 23]. Much of this centers oncurrent hardware designs or generic LLN pro-tocols and algorithms and does not address ourconcern about redesigned IP-friendly protocolsoperating in an LLN.

We focus on the two basic communication pat-terns in tree-based LLNs, one-to-many and ma-ny-to-one [5, 6]. There have been a very largenumber of research papers devoted to one-to-many communication in the context of reliableor semi-reliable broadcasting. Broadcasting isan important network activity, since it is used bybase stations to deliver commands or networkmanagement information to the devices in thesystem. Approaches to broadcasting in LLNsinclude methods for fully reliable or probabilis-tically reliable techniques [24, 25]. These tech-niques generally aim to improve the deliveryrate of broadcasting to all nodes in the systemwithout requiring excessive flooding. Althoughthese techniques are often quite powerful, theyrequire varying levels of protocol complexityand message overhead.

To study the impact of fragmentation on one-to-many broadcasting, we designed and im-plemented a semi-reliable protocol called Ra-diate. The principles underlying this proto-col are heavily influenced by earlier work inlightweight broadcast and data dissemination,including protocols such as Rime [17] and Tric-kle [26]. In particular, we use techniques suchas eavesdropping and selective retransmissionto improve reliability and reduce message over-head while minimizing complexity. As will beseen in Section 3, Radiate is designed to tradeoffreliability with overhead, and can be used effec-tively in different modalities, depending on therate of packet broadcast.

This paper extends the results presented in [27].That earlier work focused primarily on the per-formance impact of many-to-one communica-


tion and did not address one-to-many rebroad-cast paradigms.

3. Reliable Broadcasting with Radiate

Broadcasting is the basic communication op-eration to support tree-based one-to-many com-munication. To improve reliability and decreasethe overhead associated with a purely floodingstrategy, we designed the Radiate protocol. Ra-diate combines elements from IPv6 with care-fully tuned retransmission timers. In particular,Radiate uses the multicast capability of the IPv6protocol. The sender broadcastsmessages to thelink-local, all nodes multicast address, whichis ff02::1. Radiate adds a small data structureto these messages. This structure contains thesequence-number, origin, source, and length ofthe payload.

When a device has data to send, it incrementsthe sequence-number and the data is copied intothe payload. The device transmits the messageand then sets a timer to send the message againat some random time in the next second. Ran-domization is a common technique designed toreduce the chance of collisions. When the timerexpires, the message is again randomly broad-cast, doubling the timer interval (e.g. 1, 2, etc.)until a maximum of 4 seconds is reached. Typ-ically, a total of 4 broadcasts are performed foreach send execution.

Normally, this broadcastingwould be excessive,so Radiate mitigates this by increasing the timerif a device overhears the same sequence-number

that it is currently broadcasting, effectively can-celling the current timer. Thus, if a devicebroadcasts a message and three neighbors suc-cessfully receive and then also broadcast, thedevice will have increased beyond the maxi-mum and no longer rebroadcasts. The heuris-tic causes devices in dense topologies to likelyonly broadcast the message once. The heuristiccauses devices in sparser topologies, includingthose at the edge of the network, to rebroadcastcloser to the maximum number of times.

Devices receiving the broadcast check the se-quence-number to ensure whether it is more re-cent and, if so, start broadcasting in a similarfashion. Eventually, all devices will receive thebroadcast with the latest sequence number andthe re-broadcasting stops.

Figure 1 illustrates a four node topology andtime line using the Radiate protocol. Usingnode B as an example, it receives a messagewith sequence-number 1 from A, so B sched-ules a broadcast of the same message to betransmitted (note that this broadcast cannot becancelled). After the broadcast is sent, nodeB overhears C broadcast the same message, soB cancels its next scheduled broadcast. NodeB then overhears node A broadcast the samemessage and again cancels the next scheduledbroadcast. Finally, node B overhears the samemessage from node C and node B stops schedul-ing rebroadcasts, ignoring any subsequent mes-sages with the same sequence-number. Notethat each node transmits at least once and nomore than four times.

Figure 1. Radiate example showing selective cancelation.


3.1. Radiate Protocol

Following the rebroadcasting process describedabove, the Radiate algorithm is next detailedusing the notation introduced in Table 1.

m Radiate Message

mseqSequence numberof Radiate Message

msrcNeighbor sourceid of Radiate Message

n Node in network

n.m Current message forNode n (i.e. last m set)

nid Identifier of Nodewait Current wait

MAX WAIT Maximum wait, default 4000 ms

REBROAD INT Time interval firstrebroadcast, default 1000 ms

WAIT START Initial time to startwaiting, default 1000 ms

Table 1. Radiate Protocol Notation.

In algorithm 1, the sink initiates the flooding byincrementing the sequence-number and broad-casting to its neighbors. The radiate messageorigin and sequence-number fields remain con-stant along each hopwith the source changing tothe node broadcasting. Algorithm 2 is invokedwhen a radiate message is received by a nodefrom a neighbor.

In algorithm 2, line 2, the receiving node com-pares the sequence-number number of the mes-sage to its own to determine if the messageis newer. If not, the message is simply ig-nored. If newer, the message is set on thenode, the wait timer is reset/initialized to astarting value, and rebroadcast timer is sched-uled to expire in a random time (uniform dis-tribution) between 0 and REBROAD INT . Therandomness avoids collisions that might resultfrom neighboring nodes rebroadcasting at thesame time. The REBROAD INT defines anupper bound on how long it takes the flood-ing to move through the network. In our case,REBROAD INT = 1000ms so assuming thenetwork diameter to be 12 hops, the maximumamount of time for the radiate message to tra-verse the network is 12 seconds.

Algorithm 3 is executed whenever a previouslyscheduled timer has expired. It simply checks

Algorithm 1 Radiate Originator Send1: //Input: Message m from application2: set n.m = m3: increase n.mseq

4: set wait = WAIT START5: call radiateBroadcast

Algorithm 2 Radiate Receive1: //Input: Message m from neighbor2: if n.mseq < mseq then3: set n.m = m4: notify application new message received5: set wait = WAIT START6: randomWait=rand( REBROAD INT)7: scheduleBroadcast(randomWait)8: else if n.seq == m.seq then9: increaseWait(wait)10: else11: // message is ignored12: end if

Algorithm 3 Scheduled Timer Expired1: if wait ≤ MAX WAIT then2: call radiateBroadcast3: else4: // rebroadcasting stops5: end if

Algorithm 4 Radiate Re/Broadcast1: set n.msrc = nid

2: transmitMessage(n.m) to neighbors3: randomWait=rand(wait)4: scheduleTimer(randomWait)5: increaseWait(wait)

to determine if the maximum wait time has beenexceeded. If not, then a broadcast is sent. If ithas been exceeded, the timer scheduling ceasesand no more broadcasting occurs until anothernewer message is received.

Algorithm 4 ensures that the n.mseq is set to thenode’s id and transmits the message to neigh-bors. It then schedules the timer to potentiallysend again at the current wait time and increasesthe wait time.

The increaseWait function doubles the passedvalue each time it is invoked.

Line 9 in algorithm 2 increases the wait value ifthe message sequence-number overheard is the


same as the node’s current message sequence-number. This increases the wait without trans-mitting a message, therefore, effectively can-celling a pending rebroadcast. This is called thecancelling heuristic. Nodes always rebroad-cast at least once the first time. Subsequently,while waiting to rebroadcast, nodes may over-hear the same message enough times to exceedthe MAX WAIT . In this case, when the timerexpires no additional transmission is sent be-cause of the check in algorithm 3. If no mes-sages are overheard, then the timer expires call-ing algorithm 4 which transmits the messageand increases the timer. In all cases, eventu-ally the maximum wait time is exceeded andthe rebroadcasting ceases.

4. Evaluation

The goal of our evaluation was to conduct per-formance studies using a realistic set of appli-cation level programs. We constructed a simu-lation environment using the ContikiOS CoojaSimulator [28]. For devices we used the ZolertiaZ1 series built using a MSP430 processor with8kB of RAM and using the CC2420 radio. Thesystem we programmed used a simple data re-porting application for the many-to-one-mode,and the Radiate protocol carrying periodic con-trol messages for the one-to-many mode. Theevaluation first examines the impact of rebroad-casting for the many-to-one mode using Radi-ate. The evaluation then compares the impactthat fragmentation has for the many-to-one andone-to-many modes.

The configured network stack for the many-to-one modality was as follows. The mesh net-working protocol was not necessary for one-to-many, however, for realism, it was allowed torun during the Radiate experiments.

• IEEE 802.15.4

• Contiki’s 6LoWPAN

• Contiki’s mesh networking protocol

• IPv6

• UDP

The Unit Disk Graph Medium Distance Lossmodel with a 50-meter transmission range anda 100 meter interference range was used in thesimulations. We created three different network

sizes, 16 devices, 36 devices and 64 devices.The 64 device topology is shown in Figure 2.The topology is based loosely on a grid struc-ture, with an average of 30 meters between de-vices. Within this structure each devicewas ran-domly placed within a proportionally smallersquare area. This introduces some variabilitywhile still maintaining a connected network.The sink was placed in the middle of the net-work.

Figure 2. 64 device topology with the gateway in thecenter.

For each network size, a set of experiments wasperformed to determine the PDR for applicationmessages using varying rates for both many-to-one (unicast) and one-to-many (broadcast)communication modes. The PDR for each de-vice was calculated by taking the number ofapplication messages received, divided by thenumber of application messages sent. The aver-agewas then taken over the devices to determinethe PDR for the scenario.

Note that during this time normal network rout-ing and management traffic continued to betransmitted. Each rate is run for three minutesat which point the rate in increased by five andagain run for three minutes. Therefore, a rateof 35 runs for a total of 21 minutes.

Fragmentation is induced by setting the payloadthreshold to 75 bytes instead of 100 with ap-proximately 25 bytes reserved for lower layerheaders. The unicast and broadcast messages


are 85 bytes. The fragmentation scenario re-sults in two fragments/message for the samplemessages versus only requiring one for the non-fragmentation scenario.

4.1. Radiate Evaluation

The Radiate protocol was used to compare theimpact that one-to-many rebroadcast mecha-nisms have on PDR. The rates were measuredup to 60 messages/minute. In addition to PDR,overhead results were determined. The over-head for each node was determined by countingthe number of times each node broadcasts (orrebroadcasts) a Radiate message. The averageof the nodes was then taken for each rate.

The Radiate broadcast protocol is modifiedwiththe following variants, according to the numberof repeated broadcasts.

• radiate one

• radiate can

• radiate all

The radiate one simply rebroadcasts the receivednewer message once. The radiate can is thecancelling heuristic described previously and isexpected to rebroadcast between one and fourmessages. The radiate all does not attempt tominimize the overhead and always rebroadcastsfour messages.

Rebroadcasts should result in a higher PDR.However, as the rate increases, the rebroad-casts’ additional congestion may diminish thePDR improvements.

Figures 3 and 4 show the PDR and overhead for16 nodes. The radiate all PDR does well from10-20 rate, but thereafter does worse than bothother protocols. The radiate one drops to 90%PDR at rate 15, however, performs well withregard to the other protocols. The overhead be-tween the protocols is somewhat expected withradiate all incurring the most, radiate can in be-tween, and radiate one with the least overhead.At rate 30 radiate all has approximately 44%more overhead than radiate can. At rate 60 radi-ate can has approximately 39% more overheadthan radiate one.

Figures 5 and 6 show the PDR and overheadfor 36 nodes. The radiate one initially does

Figure 3. PDR Radiate 16 Devices.

Figure 4. Overhead Radiate 16 Devices.




worse from the 10-20 rate, but is approximately5% better at rate 60 than either protocol. Theradiate all again does well early, but degradeswith almost 10% worse PDR at rate 30. Theoverhead is 60% more for radiate all versus ra-diate can at rate 15. The radiate can protocolhas approximately 38% more overhead than ra-diate one at rate 40.

Figures 7 and 8 show the PDR and overhead for64 nodes. The PDR and overhead results aresimilar to the 16 and 36 network size results.



The PDR results suggest that radiate one doespoorer for lower rates 10-20, but generally doeswell thereafter. The radiate all variant initiallyperforms well for rates 10-15, but drops offabruptly at rate 20. Interestingly, radiate canappears to first correlate with radiate all andthen switches, roughly between rates 15-25, tocorrelating more with radiate one. At rate 60it appears that the variants somewhat convergeto the same PDR. The results suggest that forlower rates rebroadcasts do improve PDR.How-ever, for higher rates, the additional rebroad-casts actually become counterproductive, pro-ducing worse PDR results.

The relative overhead results between the pro-tocols is as expected. There appears to be a di-minishing of the overhead as the rate increases.This is believed to be in conjunction with the re-duced PDR; dropped packets will result in fewerrebroadcasts/overhead.

4.2. Fragmentation Evaluation

The many-to-one unicast scenario results aredepicted in Figures 9, 10, and 11. The frag-mentation scenario is labeled as fragunicast andfragbroadcast; the non-fragmentation scenariois labeled nofragunicast and nofragbroadcast.All three graphs clearly show that fragmentationseverely degrades the PDR for the different net-work sizes. As the rate increases, the PDR forboth scenarios appear to be effected proportion-ally. Figure 9 shows an increasing differencebetween fragmentation and non-fragmentationfrom approximately 18% for a rate of 5 mes-sages/minute to over 30% difference for rate35 messages/minute.

Figure 9. PDR Unicast 16 Devices.



The non-fragmentation scenario appears to beless affected by the rate increases.

Figures 10 and 11 both show the stress bothscenarios experience as the rate increases inlarger networks. Figure 10 indicates that thenon-fragmentation scenario consistently deliv-ers 20% more messages for all rates. Figure 11shows that eventually both scenarios essentiallyfail to deliver any many-to-one messages to thesink. We believe that this is due to feeder routesnear the sink becoming expectedly congested.


The unicast results clearly show the tremen-dously negative impact of fragmentation onpacket delivery rates. This strongly suggeststhat network engineers should modify their pro-tocols to either avoid excessive fragmentation orprovide additional reliability mechanisms.

The broadcast results, using radiate one, areshown in Figures 12, 13, and 14. Surprisingly, itappears that fragmentation has less of an effecton the PDR.

Figure 12 shows that both scenarios success-fully deliver 90% or more of the messages forall rates except the last which still achieves ap-proximately 86% at 35 messages/minute. Forlarger networks and higher rates, Radiate stilldelivers good performance. For rates up to 25messages/minute, the PDR is at or greater than85%. However, Figures 13, and 14, show thatfragmentation begins to denigrate the PDR to-wards 70% for the rates of 30 messages/minute.

We investigated the log files for both fragmen-tation and non-fragmentation scenarios and de-termined that, indeed, fragmentation was occur-ring when expected and messages were beingdropped due to reassembly errors. We believe

Figure 12. PDR Broadcast 16 Devices.



the difference between the unicast results andthe broadcasts results are caused by two rea-sons:

1. Broadcast messages reassembled each hop

2. Neighbors rebroadcast packets allowing cor-rect reception of previously missed frag-ments

Because the messages are reassembled by theneighboring devices, for each neighbor the fullmessage can then be transmitted again. Evenif a message is not received due to a fragmen-tation reassembly error, the device very likely


has several more opportunities to receive themessage.

Finally, we note that the rate appears to have amuch greater effect on PDR than the networksize. Network size and topology certainly playa role, however, as Figures 13 and 14 show,there is little difference between their PDR forthe varying rates.

5. Conclusion

This paper examined the impact of packet frag-mentation at the data link level on the end-to-end performance of IP-based protocols de-signed to support IoT systems. Using a mix-ture of applications and network layer routingfunctions, our results show that fragmentationcan seriously degrade the performance of thetypical IoT device to gateway communicationmodality. The results show that better fragmen-tation/reassembly support is required for IoTsystems or, alternatively, implementations sim-ply forbid fragmentation in the first place. Forone-to-many communication, we presented theRadiate broadcast protocol. Our results showedthat with proper design of data broadcast mech-anisms, gateway-to-device communication canmaintain relatively high performance levels.

Acknowledgments

This work is supported by NSF under grantsCNS-1116122 and CNS-1205453.

References

[1] L. ATZORI, A. IERA, G. MORABITO, The internetof things: A survey. Computer Networks, 54(15)(2010), 2787–2805.

[2] S. DUQUENNOY, F. OSTERLIND, A. DUNKELS, LossyLinks, Low Power, High Throughput. In Pro-ceedings of the ACM Conference on NetworkedEmbedded Sensor Systems, ACM SenSys 2011,(2011), Seattle, WA, USA, Nov. [Online]. Avail-able: http://dunkels.com/adam/duquennoy-11lossy.pdf

[3] P. LEVIS, A. TAVAKOLI, S. DAWSON-HAGGERTY,Overview of existing routing protocols forlow power and lossy networks, IETFDraft Report, (2009). [Online]. Avail-able: http://tools.ietf.org/html/draft-ietf-roll-protocols-survey-07

[4] J. VASSEUR, M. KIM, K. PISTER, Routing MetricsUsed for Path Calculation in Low Power and LossyNetworks, IETF Draft Report, (2011).

[5] Y. ZHANG, S. GANDHAM, Q. HUANG, Distributedminimal time convergecast scheduling for small orsparse data sources. In Real-Time Systems Sympo-sium, 2007. RTSS 2007. 28th IEEE International,(2007), pp. 301–310.

[6] L. MOTTOLA, G. P. PICCO, Programming wire-less sensor networks: Fundamental concepts andstate of the art. ACM Comput. Surv., 43(3)(Apr. 2011), 19:1–19:51. [Online]. Available:http://doi.acm.org/10.1145/1922649.1922656

[7] Y.-T. PARK, P. STHAPIT, D.-H. LEE, Y.-S. CHOI, J.-Y.PYUN, Data fragmentation scheme with block ackin wireless sensor networks. In Multimedia andUbiquitous Engineering (MUE), 2011 5th FTRAInternational Conference on, (2011), pp. 79–83.

[8] J. KO, J. ERIKSSON, N. TSIFTES, S. DAWSON-HAGGERTY, A. TERZIS, A. DUNKELS, D. CULLER,ContikiRPL and TinyRPL: Happy Together. In Pro-ceedings of the Workshop on Extending the Internetto Low Power and Lossy Networks (IP+SN 2011),(Apr. 2011), Chicago, IL, USA. [Online]. Avail-able: http://dunkels.com/adam/ko11conti-kirpl.pdf

[9] Part 15.4: Wireless Medium Access Control (MAC)and Physical Layer (PHY) Specifications for Low-Rate Wireless Personal Area Networks (WPANs),IEEE Computer Society Std. 802.15.4, 2006.

[10] D.-M. HAN, J.-H. LIM, Smart home energy manage-ment system using ieee 802.15.4 and zigbee. Con-sumer Electronics, IEEE Transactions on, 56(3)(2010), 1403–1410.

[11] S. DEPURU, L. WANG, V. DEVABHAKTUNI, N. GUDI,Smart meters for power grid – challenges, issues,advantages and status. In Power Systems Conferenceand Exposition (PSCE), 2011 IEEE/PES, (2011),pp. 1–7.

[12] V. GUNGOR, B. LU, G. HANCKE, Opportunities andchallenges of wireless sensor networks in smartgrid. Industrial Electronics, IEEE Transactions on,57(10) (2010), 3557–3564.

[13] A. WILLIG, Recent and emerging topics in wirelessindustrial communications: A selection. IndustrialInformatics, IEEE Transactions on, 4(2) (2008),102–124.

[14] T. WATTEYNE, A. MOLINARO, M. RICHICHI, M.DOHLER, From MANET To IETF ROLL Standard-ization: A Paradigm Shift in WSN Routing Proto-cols. Ieee Communications Surveys and Tutorials,2010.

[15] J. KO, A. TERZIS, S. DAWSON-HAGGERTY, D.CULLER, J. HUI, P. LEVIS, Connecting low-powerand lossy networks to the internet. CommunicationsMagazine, IEEE, 49(4) (2011), 96–101.


[16] J. HUI, D. CULLER, Ipv6 in low-power wireless net-works. Proceedings of the IEEE, 98(11) (2010),1865–1878.

[17] A. DUNKELS, F. OSTERLIND, Z. HE, An AdaptiveCommunication Architecture for Wireless SensorNetworks. In Proceedings of the Fifth ACM Con-ference on Networked Embedded Sensor Systems(SenSys 2007), (November 2007).

[18] D. S. J. DE COUTO, D. AGUAYO, J. BICKET,R. MORRIS, A high-throughput path metricfor multi-hop wireless routing. Wirel. Netw.,11(4) (Jul. 2005), 419–434. [Online]. Avail-able: http://dx.doi.org/10.1007/s11276-005-1766-z

[19] H. ZHANG, A. ARORA, Y.-R. CHOI, M. G. GOUDA,Reliable bursty convergecast in wireless sensor net-works. In Proceedings of the 6th ACM InternationalSymposium on Mobile Ad-hoc Networking andComputing, ser. MobiHoc ’05. (2005), New York,NY, USA: ACM, pp. 266–276. [Online]. Available:http://doi.acm.org/10.1145/1062689.1062724

[20] M. Z. N. ZAMALLOA, B. KRISHNAMACHARI, Ananalysis of unreliability and asymmetry inlow-power wireless links. ACM Trans. Sen.Netw., 3(2) (Jun. 2007). [Online]. Available:http://doi.acm.org/10.1145/1240226.1240227

[21] C.-Y. WAN, S. B. EISENMAN, A. T. CAMP-BELL, J. CROWCROFT, Overload traffic manage-ment for sensor networks. ACM Trans. Sen.Netw., 3(4) (Oct. 2007). [Online]. Available:http://doi.acm.org/10.1145/1281492.1281493

[22] B. LATRE, P. D. MIL, I. MOERMAN, N. V. DIER-DONCK, B. DHOEDT, P. DEMEESTER, Maximumthroughput and minimum delay in IEEE 802.15.4.Proceedings of the International Conference onMobile Ad-hoc and Sensor Network (MSN 05),2005.

[23] J. KIM, X. LIN, N. SHROFF, P. SINHA, Minimizingdelay and maximizing lifetime for wireless sensornetworks with anycast. IEEE/ACM Transactions onNetworking, 18(2) (April 2010), 515–528.

[24] N. RAHNAVARD, F. FEKRI, Crbcast: a collabo-rative rateless scheme for reliable and energy-efficient broadcasting in wireless sensor networks.In Proceedings of the 5th international Confer-ence on Information Processing in Sensor Net-works, ser. IPSN ’06. (2006), New York, NY,USA: ACM, pp. 276–283. [Online]. Available:http://doi.acm.org/10.1145/1127777.1127820

[25] F. WANG, J. LIU, On reliable broadcast in low duty-cycle wireless sensor networks. Mobile Computing,IEEE Transactions on, 11(5) (2012), 767–779.

[26] P. LEVIS, E. BREWER, D. CULLER, D. GAY,S. MADDEN, N. PATEL, J. POLASTRE, S.SHENKER, R. SZEWCZYK, A. WOO, The emer-gence of a networking primitive in wire-less sensor networks. Commun. ACM, 51(7)

(Jul. 2008), 99–106. [Online]. Available:http://doi.acm.org/10.1145/1364782.1364804

[27] J. POPE, R. SIMON, The impact of packet frag-mentation on internet-of-things enabled systems. InProceedings of the 35th International Conferenceon Information Technology Interfaces, ser. ITI ’13,(2013), pp. 13–18.

[28] F. OSTERLIND, A. DUNKELS, J. ERIKSSON, N. FINNE,T. VOIGT, Cross-level Sensor Network Simulationwith Cooja. Proceedings of the Local ComputerNetworks, 2006 31st IEEE Conference, (November2006).

Received: August, 2013Revised: September, 2013

Accepted: September, 2013

Contact addresses:

James PopeDepartment of Computer Science

George Mason UniversityFairfax

VirginiaUnited States

e-mail: [email protected]

Robert SimonDepartment of Computer Science

George Mason UniversityFairfax

VirginiaUnited States

e-mail: [email protected]

JAMES POPE ia a Ph.D. student of information technology at GeorgeMason University, Fairfax, Virginia. He received his M.S. in telecom-munications from George Mason University and a B.S. in businessadministration from Auburn University at Montogomery. He is a mem-ber of The Honor Society of Phi Kappa Phi. His research areas includeembedded wireless systems and developing routing algorithms and pro-tocols that effectively support application requirements for sensing andinformation quality.

ROBERT SIMON is an Associate Professor in the Department of Com-puter Science at George Mason University, Fairfax, Virginia. He re-ceived his Ph.D. in computer science from the University of Pittsburgh.He also has a B.A. in history and political science from the Universityof Rochester. His research specialization is in the field of networks,mobile and wireless computing, performance modeling and simulation.He teaches undergraduate and graduate courses in these areas.

Documents

The Impact of Packet Fragmentation and Reassembly in