Traffic analysis in optical burst switching networks: a trace-based … · 2009-11-10 · TRAFFIC...

EUROPEAN TRANSACTIONS ON TELECOMMUNICATIONSEur. Trans. Telecomms. (2009)Published online in Wiley InterScience(www.interscience.wiley.com) DOI: 10.1002/ett.1393

Invited Paper

Traffic analysis in optical burst switching networks: a trace-based case study

Ahmad Rostami*, Adam Wolisz and Anja Feldmann

Institute for Telecommunication Systems, Faculty of Electrical Engineering and Computer Science, Technical University of Berlin, Germany

SUMMARY

Optical burst switching (OBS) appears as a promising technology for building dynamic optical transportnetworks. The main advantage of OBS is that it allows for dynamic allocation of resources at sub-wavelengthgranularity. Nevertheless, the burst contention problem, which occurs frequently inside the network, has tobe addressed before OBS can be really deployed as the next generation optical transport network. Recentlya lot of attention is devoted to different approaches for resolving contentions in OBS networks. Althoughperformance analysis of these approaches is strongly dependent on the traffic characteristics in the network,the majority of the studies is so far based on very hypothetical traffic assumptions. In this study we usetraces of real measurements in the Internet to derive realistic data about the traffic that is injected into theOBS network. Specifically, we investigate the marginal distributions of burst size, burst interdeparture time,assembly delay and number of packets per burst as well as the burstiness of the burst traces. We demonstratethat the performance of an OBS core node using the real traces is pretty similar to the results obtained whenthe traffic arriving to the core node is assumed to be Poisson. In fact, usage of the Poisson as the process ofburst arrival to the core node leads in all the investigated cases to an upper bound on the burst drop rate atthat node. Copyright © 2009 John Wiley & Sons, Ltd.

1. INTRODUCTION

Advances in optical technology and wavelength divisionmultiplexing (WDM) have resulted in a huge amountof capacity available for transmission over optical links.Unfortunately the possibilities of efficient usage of thiscapacity for data and multimedia traffic are frequentlylimited by the fact, that switching techniques with fineswitching granularity—like packet switching—cannot beefficiently realised in the optical domain using the currentlyavailable technology. The reason is mainly due to the factthat switching times of optical devices are still too largecompared with the average packet length and that there isno equivalent to random access memory (RAM) in opticaldomain to realise the so-called store-and-forward switch-ing technique. To address the issue, optical burst switching(OBS) has recently been proposed [1, 2] and attracted muchattention from the networking research community.

* Correspondence to: Ahmad Rostami, Telecommunication Networks Group (TKN), Technical University of Berlin, Sekr FT5, Einsteinufer 25, 10587Berlin, Germany. E-mail: [email protected]

OBS is a switching paradigm that allows for dynamicallocation of resources in the optical domain at sub-wavelength granularity. It does so by combining threeprinciples, namely burst assembly at the edge, one-wayout-of-band signalling and cut-through switching in thecore. Burst assembly refers to the process of aggregatingsmall-size packets into bursts at ingress edges of the opticalnetwork. By increasing the size of data units this aggre-gation makes it possible to relax the requirements on thespeed of optical switching. Once a new data burst is readyfor transmission at an ingress node, a signalling messageis generated and released to the network ahead of the bursttransmission. The data burst then follows the message anoffset time after it has been sent without waiting for anacknowledgment, i.e. one-way reservation. The role of thesignalling message is to inform all switching nodes alongthe path of arrival time of the burst so they can config-ure their ports to switch the burst in a cut-through fashion

Received 20 August 2009Copyright © 2009 John Wiley & Sons, Ltd. Accepted 28 August 2009

A. ROSTAMI, A. WOLISZ AND A. FELDMANN

upon its arrival. To that end, each signalling message isprocessed electronically at every intermediate node afterpassing through an opto-electrical conversion. However,in order that data bursts can bypass such conversions atintermediate nodes, signalling messages are transmitted ondedicated WDM channels, i.e. out-of-band signalling.

The role of properly designing the burst assembly processis crucial with respect to the operation and performanceof the OBS network. Consequently, in order to design anoperational OBS network, the first step is to study the burstassembly algorithms. In our previous work in Reference[3], making some simplifying assumptions about the inputpacket arrivals to the assembly unit, we modelled themost common assembly algorithms and presented exactanalytical formulae for the characteristics of the assembledtraffic. In this paper we put our focus on the analysis ofthe assembled traffic using the real measurement-basedIP traces as the input to the assembly buffer. The objectiveof this case study is first to gain a clear insight into thenature of the traffic injected into the OBS network, andabove that to investigate the importance of using a realistictraffic model in design and performance evaluation ofan OBS network. That is, we would like to examine thevalidity of the models that are commonly assumed inanalysing or optimising the performance of the network.

The rest of this paper is structured as follows. In Section 2related works are reviewed and then in Section 3, contribu-tions of the paper are introduced. In Section 4, we introducethe burst assembly process and discuss the various burstassembly algorithms as well as the main characteristicsof the assembled burst traffic that are investigated in thiswork. In Section 5, we introduce the traffic traces usedin our experiments and discuss preparation of the tracesfor the experiments including the packet classificationfollowed by the demonstration of the characteristics of theselected traces. Then in Section 6, we present our analysisof burst-level traffic followed by the discussion on theimpact of the achieved results on the performance of thenetwork in Section 7. Finally we conclude the work inSection 8 by summarising the achieved results.

2. RELATED WORK

Several studies have dealt with traffic characteristics andassociated performance implications inside OBS networks.The existing works in this area are in general classifiedinto two categories. First, studies that take an analyticalapproach or use simulation experiments based on synthetic

traffic models. The examples of the works belonging tothis category are References [3–10]. The second categoryincludes simulation studies based on real traces collectedfrom the Internet. The works presented in References[11–14] are the major existing literature of the secondcategory. In the following we briefly review the resultspresented in these works.

The probability generating function (pgf ) of burst lengthis derived in Reference [4] using a realistic packet sizedistribution and under the assumption that the arrivalprocess is the time-slotted Bernoulli. Then based on thepgf, the distribution of burst length is approximated bythe standard distributions like the Gamma distribution. Theanalysis, however, does not include the distribution of burstinterdeparture time. In Reference [5], analytical modelsare given for distributions of both burst length as well asburst interdeparture time for different burst assembly strate-gies, nevertheless, the work does not consider the hybridassembly algorithm. The hybrid algorithm is studied inReference [6], where it is claimed that the burst lengthdistribution would have a very narrow range, therefore onlythe distribution of burst interdeparture time is developedunder the assumption that the burst length is fixed. In otherwords, the hybrid algorithm is approximated by the volume-based algorithm. Authors in Reference [7] characterise thetraffic at the output of a time-based burst assembler assum-ing that the input packet-level traffic can be modelled by thefractional Gaussian noise (FGN) process. They also demon-strate that self-similarity does not have an adverse effect onthe blocking probability of a bufferless OBS core switch,as compared to the case where the burst traffic is assumedto be Poisson.

In Reference [8], a variation of the time-based assemblyalgorithm is studied, in which at the end of each aggregationperiod a length threshold is applied and contents of thebuffer are accordingly assembled into a few bursts of agiven length. The analysis provides the distribution of num-ber of bursts that are generated at the end of an assemblyperiod. In Reference [3], we analyse the assembly processand develop exact analytical models for the distributionof size and interdeparture time of bursts generated by theassembler. The analysis is carried out under the assumptionthat the process of input packet arrivals is Poisson andpacket sizes are exponentially distributed. In Reference [9]using synthetic self-similar input traffic it is claimed thatthe self-similarity in the traffic injected to the OBS networkcan be reduced through the assembly process. However,in Reference [10] it is demonstrated that this reduction isonly related to short time scales and also depends on theload offered to the assembly buffer.

Copyright © 2009 John Wiley & Sons, Ltd. Eur. Trans. Telecomms. (2009)DOI: 10.1002/ett

TRAFFIC ANALYSIS IN OPTICAL BURST SWITCHING NETWORKS

In Reference [11] a trace-based simulation is conductedto investigate the impact of the time threshold in atimer burst assembler on the fraction of bursts that aregenerated with a length smaller than the minimum lengthrequirements of the system. In Reference [12], the effect ofassembly parameter on the scaling properties of multifrac-tal traffic is analysed. The authors use real Internet tracesas input and conclude that the key factor in characterisingthe output of the assembler is the relationship betweenthe burst assembly time scale and the cut-off time scaleof the input traffic. In Reference [13], authors use recordsof MPEG4 video frames to verify the analytical modelsthat they have developed for the burst assembly process. InReference [14], authors study the statistical characteristicsof traffic at the output of burst assembly using tracescollected form the Internet. The work in Reference [14],to our knowledge, is the only available literature thatinvestigates the main characteristics of the burst-leveltraffic such as burst length and burst interdeparture timeusing real Internet traces. Nonetheless, a clear and thoroughpicture of the traffic statistics for the hybrid burst assemblyis missing in this work. Furthermore, the authors ignore theclassification of input IP packets according to their desti-nation address that, as we explain in the next section, is anintegral part of the burst assembly process. That is, the burstassembly algorithm is applied to all the packets travellingover a link irrespective of their destination address.

3. CONTRIBUTIONS

Appropriately designing and setting the burst assemblyunit is an important step in design and performanceoptimisation of the OBS network. Nevertheless, if welook at the assumptions usually made for the performanceevaluation and improvements of OBS networks in hundredsof published papers, we observe that this important issuehas been to some extent overlooked. In fact, many of theexisting performance studies use either the Poisson or thesynthesised self-similar to model the traffic at the output ofthe burst assembly unit and input of the core OBS nodes.While the former one might look too simplistic, the latterone usually does not have a reasonable accuracy when itcomes to the short time scale characteristics of traffic asthe focus of the model is on the long range characteristicsof the traffic. Nevertheless, in the OBS network, whichenjoys either no or limited buffering capacity, the accuracyin the short time scale is of great importance.

On the other hand—as explained above—the major exist-ing traffic models are analytical studies and, in fact, none of

them provides a comprehensive characterisation of trafficat the output of the burst assembler at different traffic ratesand for different assembly parameters using real packettraces. Therefore, our goal in this paper is to address theseshort comings through a comprehensive case study. Thework presented here enjoys several unique features as com-pared to the available literature. First and foremost, in thisstudy we take into account the destination-based classifica-tion of packets at the input of the burst assembler. Further,we evaluate the impact of the assembly parameters andtraffic intensity on the generated traffic through an inter-mediate utility function called PTout . More specifically, theinput parameters are first mapped into PTout , which assumesvalues in [0, 1] and gives the fraction of bursts that are gen-erated due to the expiry of the burst assembly timer. A largeset of experiments are designed and carried out to charac-terise PTout and its relation to the input parameters. Then,the burst level traffic is characterised at different values ofPTout . Since the analysis involves several input parameters,taking this approach brings some structures into the analysisand improves its tractability.

4. BURST ASSEMBLY IN OBS

Design and implementation of a properly tuned burst aggre-gation mechanism is indispensable to the realisation of anyOBS network. In order to generate bursts, each OBS ingressnode must contain a burst assembly unit. The unit receivespacket-level traffic at its input and aggregates them intobursts. In its simplest form, an assembly unit is composedof a packet classifier, an assembly controller and severalassembly buffers, see Figure 1. Once a packet arrives to theunit, its destination address (and possibly the QoS class thatit belongs to) is checked by the packet classifier. Accordingto the result of the classification, the packet is enqueuedin one of the available virtual destination queues (VDQs),where each VDQ is a buffer dedicated to packets destinedto a certain egress node and is referred to as the assembly

Input PacketTraffic

VDQ 1

VDQ 2

VDQ N

Assembly ControllerAssemblyParameters

AssembledBurst Traffic

Packet

Classifier

Figure 1. Burst assembly unit in an ingress OBS node.



buffer. Therefore, the unit should contain one assemblybuffer per destination (and per QoS class). Note that the termdestination here refers to an egress node in the OBS domain,thus packets of the same attributes from different microflows can be multiplexed into the same assembly buffer.

The assembly controller is responsible for making thedecisions regarding when contents of each assembly buffershould be aggregated into a burst and released to thenetwork. The controller takes care of scheduling burstgenerations based on several criteria, of which some areimposed by the network, e.g. maximum and minimum burstlength, and others are imposed by QoS requirements ofincoming packet traffic, e.g. maximum delay that a packetcan tolerate in an assembly buffer. Accordingly, variousproposals have been investigated for this purpose, see e.g.References [10, 15]. The controller may apply differentalgorithms to different assembly buffers.

In general, burst assembly algorithms may be classifiedinto three major categories, namely time-based, volume-based and hybrid algorithms. In the time-based algorithm,the controller is equipped with a timer. Once a packet arrivesto an empty assembly buffer, the corresponding timer isset to a time threshold TTh. This threshold is determinedconsidering maximum delay that a packet can tolerate in theingress node. Then, as soon as the timer expires, all packetsin the buffer are aggregated into a burst and sent out. Thetimer is deactivated when the buffer is emptied. If length ofthe burst generated in this way is below a given level Lmin,padding has to be used to fulfil the minimum burst lengthrequirements. The value of Lmin is dictated by the networkarchitecture, and depends on the ratio between number ofdata and control channels of WDM links in the network.Specifically, Lmin has to be selected large enough so as toavoid possible conflicts between reservation messages ofdifferent data bursts over the control channel [15].

In the volume-based algorithm, the controller checks theaggregate length of packets in the assembly buffer eachtime a new packet arrives. As soon an the aggregate lengthexceeds a predefined threshold LTh all packets in the bufferare assembled into a new burst. This threshold must belarger than the minimum burst length requirement of thenetwork and is usually selected with respect to maximi-sation of utilisation of resources over the network [15].Algorithms of this category are usually not recommendedfor delay-sensitive traffic, because there is no guarantee onthe upper-bound delay that a packet may experience in thebuffer.

Alternatively, in the hybrid assembly algorithm, the con-trol unit keeps track of both aggregate volume of packetsin the buffer and the time elapsed since the first packet has

arrived. That is, the timer is set to TTh once a packet arrivesand finds the buffer empty, and length of the queue is com-pared against a length threshold LTh upon each new arrival.Then, a new burst will be generated when either the timergoes off or the volume threshold is exceeded. In either case,the timer will be deactivated after the buffer is emptied. In analgorithm of this category load intensity determines whichcriterion, between time and volume, will be used to gener-ate a new burst at a given time. That is, if the load intensityis below a specific level, a new burst will be generated TThunits of time after the first packet has arrived, however, ifthe load intensity is heavy enough, bursts of length LTh willbe generated back to back so that no packet will encountermaximum assembly delay of TTh. In the former case, it islikely that the timer expires while the aggregate length isless than the minimum burst length requirement. In thatcase, padding has to be used. The hybrid algorithm is infact the most general burst assembly algorithm and also themost reasonable one since it makes it possible to have con-trol over both the burst size as well as the burst assemblydelay. Accordingly in this paper we put our focus on thehybrid algorithm.

As discussed earlier, tuning the assembly parameters isthe key to the successful operation of the OBS network.This is, in fact, because profile of the traffic flowing in thenetwork—i.e. statistics of burst size, burst interarrival time,burstiness of the burst arrival process and burst assemblydelay—is directly shaped by the profile of traffic at theinput of the assembly unit as well as the applied assemblyparameters. Now let us briefly discuss the role of differentelements of the burst-level traffic on the operation of theOBS network. Burst size statistics acts as one of the maincriteria for designing the control plane of the OBS network,see Reference [15]. In general, the burst size variation hasa direct impact on increasing the complexity of designingthe control plane. Also, the principles of the teletraffictheory suggest that statistics of the burst interarrival timeas well as the burstiness of the burst arrival process exertinfluence on the burst loss probability, which is by far oneof the main criteria of the OBS network performance. Thisis further detailed in Reference [16]. Additionally, the timeinput packets spend in the assembly buffer, i.e. the burstassembly delay plays an important role both in terms ofthe overall design and performance of the network and thequality of service provided to the connection at the end toend level. That is, on one hand increasing the allowableburst assembly delay—through increasing TTh—resultsin a burst trace which is smoother, both in terms of burstsize and burst arrival process, and on the other hand itdeteriorates operation of the higher layer protocols, e.g.



Table 1. Input traces used in our analysis.

Trace No. of packets Data rate (Mb/s) Mean packet size (B) Mean packetinterarrival time (�s)

TR1 17 672 402 24.94 645.47 206.87TR2 12 368 694 17.48 645.73 295.57TR3 8 836 042 12.47 645.18 413.74

by triggering the TCP time out, see Reference [10]. In thiswork we evaluate statistics of all these metrics in detail.

5. INPUT PACKET TRACES

We focus on the hybrid burst assembly algorithm and asits input we use packet level traces collected on 18 July2007 from the Munich scientific network (Muenchener Wis-senschaftnetz, MWN) in Germany. This network connectssome 55 000 hosts at two major universities and severalresearch institutes to the Internet via a 10 Gb/s link. Thedaily volume of traffic transferred over this link is around3–6 TB. The trace of packets travelling on this link has beenrecorded for 24 h, though for our analysis we select a partof the trace which is 1 h long (10a.m.–11 a.m.). Over this1-h period the link is highly utilised.

As explained in the last section, in the first step, weneed to classify packets according to their destined egressedge node. The exact implementation of the classificationnecessitates knowledge about the exact topology as wellas routing in the considered network. Alternatively, one canclassify packets based on the autonomous system (AS) theyare destined to. This is one of the closest approximations tothe edge-based classification that is feasible without mucheffort. In fact, AS-based classification might be somewhatfiner than the classification based on the egress edge node,therefore it is considered as a simple method mimicking theedge based classification in a conservative manner. Accord-ingly we opt for the AS-based classification in this work. Asa result of the classification, we come up with several sub-traces, each with a different destination AS. For our analysiswe rely on the subtrace with the highest data rate.

Since we would like to analyse the impact of trafficintensity on the burst assembly process, we decided todownsample the selected subtrace at different rates. Forthis purpose, we apply the connection-level downsampling,where packets belonging to different end-to-end connec-tions are identified based on their source and destination IPaddresses and port numbers. After having identified the con-nections in the considered subtrace, we randomly remove

a certain fraction of the connections from the trace so thatthe data rate of the resulting subtrace is equal to the desireddata rate. In this way, we create two additional subtraceswhose data rates are 70 and 50% of the original subtrace.Table 1 depicts the information of the three traces that areused in our analysis.

5.1. Statistical characteristics of packet traces

One of the major distinctions between the work presentedhere and the other related works discussed earlier is usingdifferent type of input to the system. Therefore, it is impor-tant to first explore the statistical characteristics of the threetraffic traces, which are denoted as TR1, TR2 and TR3 inTable 1. Figure 2(a) shows the volume of the traffic asso-ciated with each of the three traces averaged over 10 msintervals. Also, Figure 2(b) shows the empirical comple-mentary cumulative distribution function (CCDF) of packetinterarrival times for the traces. We observe that in all threecases, the plotted CCDF is a straight line, except for thevery small interarrival times, which is an indication thatthe marginal distribution of packet interarrival times can bewell approximated by the exponential distribution. In fact,the minor deviation between the CCDFs and the exponentialdistribution at the very small interarrival times is associatedto the back to back arrival of packets on the link.

Depicted in Figure 2(c) is the empirical cumulative dis-tribution function (CDF) of packet sizes for the three traces.The packet size varies between 40 and 1500 B and the largestspikes are located at 40, 1492, 53 and 1500 B, respectively.

To investigate the correlation structure of the packettraces, we estimate the autocorrelation of packet interarrivaltimes for the lags up to 30. As depicted in Figure 3(a), thepacket arrival is a correlated process and the three examinedtraces exhibit more or less similar correlation structure. Tocheck for the scaling property in the packet traces, we con-duct a wavelet-based multi resolution analysis (MRA) [17].Figure 3(b) depicts that there is a strong scaling behaviourwith the estimated Hurst parameters 0.947, 0.972 and 0.947for TR1, TR2 and TR3, respectively. Also, it is seen that theonset of LRD for all three cases is at scale j = 5 correspond-ing to 300–500 ms.



0 0.5 1 1.5 2 2.5 3 3.5

x 106

0

0.2

0.4

0.6

0.8

1MB

1.2

1.4

1.6

1.8

2

(a) (b) (c)

0 2 4 6 8 1010?7

10?6

10?5

10?4

10?3

10?2

10?1

100

0 500 1000 15000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1TR1TR2TR3

TR3TR2TR1

TR1TR2TR3

T ime [msec] P acket interarrival time [msec] P acket Size [B ytes]

Co

mp

lem

en

tary

cu

mu

lati

ve d

istr

ibu

tio

n f

un

cti

on

Cu

mu

lati

ve

dis

trib

uti

on

fu

nc

tio

n

Figure 2. Input packet traces: (a) volume of arriving traffic averaged over 10 ms intervals, (b) CCDF of packet interarrival times and(c) CDF of packet sizes.

2 4 6 8 10 12 14 1524

26

28

30

32

34

36

Octave j (2j x 10 [msec])

Log 2 (

Ene

rgy)

T R1T R2T R3

0 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48500.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

T ime interval [msec]

Coe

ffici

ent o

f var

iatio

ns (

CoV

)

T R3T R2T R1

5 10 15 20 25 300

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

Lag

Aut

ocor

rela

tion

T R 1T R 2T R 3

(b) (c)(a)

Figure 3. Input packet traces: (a) autocorrelation of packet interarrival times, (b) wavelet-based global scaling plot using waveletDaubechies 4 (bytewise) and (c) coefficient of variation of number of arriving bytes as a function of time scale.

As we explain in the next section, the short range bursti-ness of the input traces is a factor that can largely influencethe burst-level statistics. Therefore, in Figure 3(c) weestimate the coefficient of variation (CoV) of the volume oftraffic collected over small (up to 50 ms) time intervals. Weobserve that in the short time scales the traffic gets smootherwith time and also the CoV decreases with input data rate.

In summary, at the input to the assembler we have acorrelated long-range-dependent process, which is quitedifferent from that assumed in analytical works including

Reference [3]. More precisely, although the marginaldistribution of the packet interarrival time follows theexponential distribution and is the same as the assumptionwe make in Reference [3], the packet size distribution andthe correlation structure of the arrival process are clearlydifferent from our assumptions in Reference [3]†.

† In Reference [3] we assumed the arrival process to be Poisson and thepacket sizes to be exponentially distributed.



6. BURST-LEVEL ANALYSIS

In this section we analyse the characteristics of traffic at out-put of the burst assembly unit. For this purpose we develop asimulation model of the hybrid burst assembly algorithm, asthe most general algorithm, in Perl [18]. The model allowsus to apply the burst assembly to the recorded packet tracesthat were discussed in the previous section. The output ofthe simulation model are traces of bursts that are recordedfor further statistical evaluations. In order to characterisethe burst-level traffic, we should conduct several simula-tion experiments varying different parameters of the system.There are three important such parameters namely, inputtraffic rate, the time threshold TTh and the length thresh-old LTh. Therefore, we must design our experiments in away that we can observe the impact of all these parameterson the burst-level characteristics. On the other hand, fromthe discussion in Section 4, we know that depending on thecombination of these three parameters there might be twotypes of bursts at the output of the burst assembler. Namely,bursts that are generated due to expiration of the assemblytimer (hereinafter type T bursts) and bursts that are gener-ated before the timer expires because the length thresholdis met (hereinafter type Z bursts). Consequently, the statis-tical characteristics of generated burst trace will be largelyrelated to the relative fraction of type T (Z) bursts, i.e. theprobability that the assembly timer goes off, PTout . In orderto facilitate the process of characterising the burst traffic, wefirst characterise PTout and investigate the impact of inputparameters on this function and then in the next step designour experiments based on different values of PTout .

6.1. Analysis of PTout

Let us begin our analysis with an illustrative example andassume that the input packet-level traffic is of type constantbit rate (CBR) with rate RCBR. For given values of inputtraffic rate and the length threshold, PTout as a function ofTTh has a deterministic behaviour as depicted in Figure 4.That is, there is a certain value of t∗ that if the time thresholdis set to be smaller than this value, all generated bursts willbe of type T, otherwise all of them will be of type Z. Now,let us replace the CBR packet trace at the input with a burstytrace like the ones we consider in our experiments. In thiscase, PTout will no longer be a deterministic function ofPTout , however, we still can argue that the assembly processis a renewal process. That is, the assembly buffer forgetsits history every time a new burst is generated and sent out.On account of this, the deviation of the PTout from the onedepicted in Figure 4 will be only influenced by the variation

0

t = LTh

CBRR* TTh

PTo

ut

1

Figure 4. Probability of the assembly timer expiry as a functionof TTh at given LTh and input traffic rate for the CBR traffic.

of the input traffic at the time scales not greater than thetimer threshold of the assembly process. As a measure ofthe variation of the input traffic we rely on the CoV of theinput traces as illustrated in Figure 3(c).

To characterise PTout , we use the simulation model toestimate the fraction of T bursts at different values of timeand length threshold for all the three packet traces. Thesimulation results are shown in Figure 5. Furthermore andfor the sake of comparison we calculate values of PTout as afunction of TTh based on the analytical models developed inReference [3] for the TR1. The results from the analyticalmodel are denoted as Model in Figure 5(c) and comparedto the results of our simulation, denoted as Sim. Notethat for calculating PTout using the analytical model, weonly need the mean input traffic rate, average packetsize—which are both easily extracted from TR1—andthe assembly parameters. We observe that there is a cleardistinction between results of the analytical model and thesimulation, which stems from the difference in the inputtraffic characteristics of the two cases. Although the modelaccurately predicts the value of TTh at which PTout = 0.5,the range of variation of TTh when PTout varies from 0 to 1is much larger for the simulation than the analytical case.

Now let us further explore the results of our experiments.From Figure 5 we observe that for given input traffic rateR and length threshold LTh, the value of time threshold atwhich PTout = 0.5 can be very well approximated by LTh

R.

The labels of the x-axis on plots of Figure 5 at (3.21, 4.48,6.41, 9.15, 12.82, 16.04, 32.07) depict TTh = LTh

Rfor the

corresponding settings. To evaluate the accuracy of thisapproximation, we calculate the relative error between thevalue of TTh that results in PTout = 0.5 in our simulation andits corresponding approximation in Table 2. The accuracyof this approximation improves with the input traffic rate



0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1(a) (b) (c)

PTo

ut

0

3.21

4.48

6.41 10

15

20

25

30

TR3TR2TR1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

PTo

ut

0

3.2

6.4

1

10

16.0

4

25

32.0

7

40

50

60

70

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

PTo

ut

0

6.4

1 9

.15

12.8

2

20

25

30

35

40

45

50

TR1TR2TR3

SimModel

Time threshold [msec] Time threshold [msec] Time threshold [msec]

LTh

=10kB

LTh

=20kB

LTh

=100kB

LTh

=50kB

Figure 5. Probability of the assembly timer expiry with the packet traces as input as a function of TTh at (a) LTh = 10 kB, (b)LTh = 20 kB and (c) TR1.

Table 2. Relative error between the estimated TTh at PTout = 0.5from simulation and the approximation using LTh

R(the values are

in per cent).

TR1 TR2 TR3

LTh = 10 KB 2.54 2.64 3.02LTh = 20 KB 0.45 0.68 0.82LTh = 50 KB 0.48 — —LTh = 100 KB 0.17 — —

and the selected length threshold so that the relative erroris less than 1% for settings with LTh � 20 KB at all threeinput data rates. To explain this behaviour we notice that fora given input rate, the corresponding time threshold lead-ing to PTout = 0.5 increases with the length threshold that,in turn, leads to a smaller CoV according to Figure 3(c).When the CoV decreases, we expect the traffic rate to becloser to the average value and hence a better approxima-tion. We further note that increasing LTh from 10 to 20 kBhas a very large impact on the relative error, though thisis not the case when we further increase LTh to 50 and100 kB. In fact, although the CoV curve is a monotonicallydecreasing function of the time interval, the tilt of the curvedecreases with time interval as depicted in Figure 3(c).

Another important observation is that for given values ofLTh and PTout , the input traffic rate and the the inverse of thecorresponding TTh are linearly related. That is, at a givenLTh if the input rate increases by a factor of 2, then TThshould be halved in order that PTout remains unchanged.

This again is explained with the help of CoV curves inFigure 3(c). That is, the value of CoV remains almost fixedif we scale the traffic rate and the 1

TThby the same factor.

Nevertheless, it is obvious that CoV does change at a giveninput rate, when we simultaneously scaling TTh and LTh,meaning that there is no linear relation between LTh andTTh at a given input rate.

To further elaborate on the linear relationship betweenthe input rate and the inverse of the time threshold, we takethe curves of PTout versus TTh for the case of TR1 at LTh =10, 20 and 50 kB and using these three curves estimate thecorresponding PTout versus TTh curves for TR2 and TR3by appropriately varying the values of TTh. Then for eachcase, we compare the results and quantify the accuracy ofthe estimation by calculating the root mean square error(RMSE) between the curve achieved from the simulationand the approximated one. The results shown in Table 3indicate the high accuracy of the estimation.

In the next sections we focus on the characteristics of theburst traces such as burst length and burst interdeparture

Table 3. Root mean square error of estimating PTout at TR2 andTR3 based on PTout at TR1.

TR2 TR3

LTh = 10 KB 1.8 × 10−3 7.4 × 10−3

LTh = 20 KB 2.9 × 10−3 7.0 × 10−3

LTh = 50 KB 1.5 × 10−3 3.7 × 10−3



Table 4. Values of TTh for the 24 scenarios considered for simulation experiments (ms).

LTh = 10 KB LTh = 20 KB

PTout = 0 PTout = 0.25 PTout = 0.5 PTout = 0.75 PTout = 0 PTout = 0.25 PTout = 0.5 PTout = 0.75

TR1 ∞ 4.668 3.126 1.907 ∞ 8.712 6.386 4.420TR2 ∞ 6.632 4.456 2.726 ∞ 12.310 9.088 6.318TR3 ∞ 9.176 6.219 3.925 ∞ 17.200 12.720 9.024

time distributions. For this purpose, we design severalscenarios based on the three input traces, two values ofLTh = 10, 20 kB and different values of TTh. The selectionof the values of TTh is done based on the PTout curves ofFigure 5. That is, for each combination of input traffic rateand LTh, we select four different values for TTh in a waythat they lead to PTout = 0, 0.25, 0.5, 0.75. In this way,we end up with 24 different scenarios. The correspondingvalues of TTh for each of the scenarios are shown inTable 4.

6.2. Burst length

In this section we characterise the marginal distribution ofburst size. For this purpose we simulate the burst assemblyfor each of the 24 scenarios listed in Table 4. The cor-responding CDFs of burst size are depicted in Figure 6.Additionally, the average and variance of burst size for eachscenario is shown in Table 5.

we observe that CDF of the burst size is composed oftwo regions; a fraction equal to PTout of bursts are smaller

than LTh in size (type T) and the rest are those with sizelarger than LTh (type Z). The distribution of type Z bursts islimited to the range (LTh, LTh + 1500) B and is influencedmuch by the distribution of a single packet size, which is atmost 1500 B long.

For given length threshold and input data rate, the averageburst size decreases and its variance increases with PTout .There are two reasons for that. First, the average size ofT bursts are smaller than those of Z bursts. In addition tothat, even when we compare the average burst size only forT bursts we notice the same trend, which stems from thefact that according to Table 4 a decrease in PTout—causedby an increase in TTh—allows the burst assembler collectmore packets before the timer goes off. For the same reason,we observe that the average size of the type T bursts inFigure 6(b) is larger than two times the average size of typeT bursts in Figure 6(a) although the length threshold onlydoubles.

To further study the impact of input traffic rate we con-sider the difference between burst sizes at fixed values ofLTh and PTout and at different input rates. In this case, the

0 2 4 6 8 10 11.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Burst size [kB]

Cum

ulat

ive

dist

ribut

ion

func

tion

TR1TR2TR3

0 2 4 6 8 10 12 14 16 18 20 21.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Burst size [kB]

Cum

ulat

ive

dist

ribut

ion

func

tion

TR1TR2TR3

(a) (b)

PTout

= 0.75

PTout

= 0.5

PTout

= 0.25

PTout

= 0

PTout

= 0.75

PTout

= 0.5

PTout

= 0.25

PTout

= 0

Figure 6. CDF of burst size: (a) LTh = 10 kB and (b) LTh = 20 kB.



Table 5. Estimated mean and variance of burst size for different scenarios.

PTout = 0 PTout = 0.25 PTout = 0.5 PTout = 0.75

Mean (B) Var (B2) Mean (B) Var (B2) Mean (B) Var (B2) Mean (B) Var (B2)

LTh = 10 kBTR1 10 625 1.78 × 105 9668 4.1 × 106 8254 8.79 × 106 6147 1.19 × 107

TR2 10 635 1.78 × 105 9684 4.03 × 106 8269 8.74 × 106 6176 1.19 × 107

TR3 10 631 1.81 × 105 9704 3.85 × 106 8336 8.34 × 106 6328 1.14 × 107

LTh = 20 kBTR1 20 653 1.88 × 105 19 219 9.66 × 106 16 979 2.25 × 107 13 447 3.37 × 107

TR2 20 660 1.89 × 105 19 232 9.51 × 106 17 014 2.21 × 107 13 533 3.32 × 107

TR3 20 648 1.9 × 105 19 232 9.46 × 106 17 101 2.11 × 107 13 772 3.12 × 107

10−1

100

101

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

(a) (b)1

Time [msec]

Cum

ulat

ive

dist

ribut

ion

func

tion

PTout

= 0.75

PTout

= 0.5

PTout

= 0.25

PTout

= 0

10−1

100

101

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time [msec]

Cum

ulat

ive

dist

ribut

ion

func

tion

PTout

= 0.25

PTout

= 0

PTout

= 0.5

PTout

= 0.75

LTh

= 20 KBL

Th = 10 kBL

Th = 10 kB

LTh

= 20 KB

Figure 7. Marginal CDF of burst interdeparture times: (a) TR1 and (b) TR2.

shape of the CDF function remains almost unchanged, how-ever, the average size of bursts (particularly of T bursts)increases as the input rate decreases, though the differenceis not much. The reason for this difference is attributed tothe small difference in the variation of traffic at differentrates over small time scales. Accordingly, this differencediminishes as the time scale increases.

6.3. Burst interdeparture time

Now let us consider the marginal distribution of burst inter-departure times. We use the same 24 scenarios describedin Section 6.1. The CDF of burst interdeparture times arepresented in Figures 7 and 8. Also, the mean and varianceof the interdeparture times are listed in Table 6.

Similar to the burst size CDF, here again the CDF func-tions consists of two distinctive regions, namely, region Iassociated with the bursts whose iterdeparture time is

10−1

100

101

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time [msec]

Cum

ulat

ive

dist

ribut

ion

func

tion

PTout

= 0.75

PTout

=0.5

PTout

= 0.25

PTout

= 0

LTh

= 10 kB

LTh

= 20 kB

Figure 8. Marginal CDF of burst interdeparture times for TR3.



Table 6. Estimated mean and variance of burst interdeparture times for different scenarios.


Mean (ms) Var (ms2) Mean (ms) Var (ms2) Mean (ms) Var (ms2) Mean (ms) Var (ms2)

LTh = 10 kBTR1 3.41 5.3 × 10−6 3.1 2.43 × 10−6 2.65 1.05 × 10−6 1.97 3.25 × 10−7

TR2 4.87 1.05 × 10−5 4.43 4.84 × 10−6 3.78 2.09 × 10−6 2.83 6.67 × 10−7

TR3 6.82 1.92 × 10−5 6.22 8.88 × 10−6 5.34 3.84 × 10−6 4.06 1.28 × 10−6

LTh = 20 kBTR1 6.62 1.22 × 10−5 6.16 6.04 × 10−6 5.45 2.74 × 10−6 4.31 8.28 × 10−7

TR2 9.46 2.39 × 10−5 8.8 1.17 × 10−5 7.79 5.36 × 10−6 6.19 1.64 × 10−6

TR3 1.32 4.64 × 10−5 1.23 2.24 × 10−5 1.1 1.03 × 10−5 8.83 3.36 × 10−6

0 5 10 15 20 25 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

T ime [msec]

Cum

ulat

ive

dist

ribut

ion

func

tion

0 5 10 15 20 25 30 35 40 45 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Cum

ulat

ive

dist

ribut

ion

func

tion

T ime [msec]

Estimated C DfFitted Gamma

Fitted GammaEstimated C DF

T R1

T R2

T R3T R3

T R2

T R1

(a) (b)

Figure 9. Fitting the Gamma distribution to the estimated burst interdeparture times at PTout = 0: (a) LTh = 10 kB and (b) LTh = 20 kB.

smaller than TTh and region II associated with the rest ofthe bursts. Note that for the case PTout = 0 we only observeregion I since TTh = ∞.

Recalling from Section 4, the burst interdeparture time iscomposed of two parts. The first part, tp, is the differencebetween the time that last burst has been created and the timethat a new packet arrives to the assembly buffer and triggersformation of a new burst. This part is characterised by theinterarrival time of individual packets, and as demonstratedbefore it is well approximated by the exponential distribu-tion. The second part of a burst interdeparture time, tb, ischaracterised by the difference between the time the assem-bly timer has been set and the time that either the assemblytimer expires or the volume criterion is met. Therefore, thispart always assumes values not larger than TTh.

Consequently, the region I in the CDFs of Figures 7 and 8is associated to the fraction of type Z bursts whose total for-mation time (tp + tb) is smaller than TTh. In other words, theburst interdeparture time in this region is characterised bythe conditional distribution of sum of n consecutive packet

interarrival times given that this sum is smaller than TThand that aggregate sum of corresponding packets’ sizes islarger than LTh. In Reference [3], we demonstrate that forthe simplified case of Poisson packet arrivals and exponen-tial packet size distribution, this conditional distribution ischaracterised by the Erlang-k distribution. Similarly, fromFigures 7 and 8 we observe that the distribution in region Ipictorially resembles the Gamma distribution,‡ which is thegeneral case of the Erlang-k distribution. To validate thishypothesis, we apply the nonlinear regression to the esti-mated CDFs of Figures 7 and 8 for the case PTout = 0. In thismethod, the Gamma distribution is fitted to the estimatedCDFs using the least-square fit method [19]. The results ofregressions are shown in Figure 9, where estimated CDFs

‡ The Gamma CDF is defined as follows.

F (x; a, b) = 1ba�(a)

∫ x

0ta−1e

−tb dt

where, �(z) =∫ ∞

0tz−1e−t dt.



Table 7. Estimated parameters of the best Gamma fit and corre-sponding regression errors for PTout = 0.

(a, b) R2 RMSE

LTh = 10 kBTR1 (2.05,1.69) 0.9989 8.58 × 10−3

TR2 (2.13,2.33) 0.9988 7.5 × 10−3

TR3 (2.39,2.90) 0.9991 7.64 × 10−3

LTh = 20 kBTR1 (3.53,1.89) 0.9994 7.11 × 10−3

TR2 (3.74,2.55) 0.9996 5.89 × 10−3

TR3 (3.88,3.44) 0.9997 5.6 × 10−3

and the corresponding best Gamma fits are plotted together.Also, Table 7 depicts the shape and scale parameters ofthe corresponding fitted Gamma distribution, denoted as aand b respectively, together with the corresponding RMSEand R2, which serve as measures of the accuracy of theregression [19]. We observe, both pictorially and quantita-tively, that the distribution of burst interdeparture times inregion I is consistent with the Gamma distribution and theaccuracy of the fit increases as the total number of packetscontributing to the interarrival times increases.

Now we consider the region II of CDFs shown inFigures 7 and 8. Unlike the region I, the CDF in this regionis associated to both types of bursts. In fact, there might besome Z bursts whose total formation time exceeds TTh ifthe associated tp is large enough. Nonetheless, the fractionof these kind of bursts in region II is small in comparison tobursts of type T. In fact, for each scenario the fraction of typeZ bursts that fit in region II is calculated as 1 − PTout − P0,where P0 is the fraction of bursts whose interdeparturetimes fit in region I. From the figures it is seen thatthis fraction becomes smaller as the input rate increases.Accordingly, we argue that the CDFs in region II are mainlycharacterised by TTh + tp, where tp is consistent with theexponential distribution as characterised in Figure 2(b).

6.4. Burst assembly delay

In this section we look into the delay that individual packetsexperience in the assembly buffer. The mean and varianceof delay values for the considered scenarios are listed inTable 8. The distribution functions corresponding to TR1and TR3 are shown in Figure 10. The distribution for TR2is not shown because of space limit.

It is seen that, as expected, the average delay increaseswith the length threshold and the inverse of the input traf-fic rate. Also, for given length threshold and input trafficrate the average delay is in direct relation with the timethreshold TTh. In the considered scenarios, the average valueof delay is always smaller than 10 ms. Nevertheless, whenwe set TTh = ∞ the maximum per packet delay increasesto 73.74 ms for the scenario with TR2 and LTh = 20 kB,though not shown here.

From the CDFs we observe that the per packet assemblydelay is distributed over three regions, namely, very smalldelay values (from 0 up to tens of �s), delay values in therange of tens of �s up to TTh and finally there is a fractionof packets with delay equal to TTh (except for the case ofTTh = ∞). The very small delay values in the first regionare associated to the packets that are very close to the tailof the associated bursts, i.e. the packets that arrive shortlybefore the timer expires in T bursts or the length thresholdis reached in Z bursts. It is seen that only a small frac-tion of packets belong to this category, in particular, whenthe length threshold increases. The mass of the distribu-tion is concentrated in the second region so that this regionaccounts for at least 80% of the assembled packets. The dis-tribution of the packet delay in this region is also influencedby PTout so that when PTout approaches 1, the distributiontends towards the uniform distribution and on the other handwhen PTout approaches 0 this distribution tends towards theexponential distribution. Finally, for the cases with a limitedtime threshold a small fraction of packets experience a delay

Table 8. Estimated mean and variance of packet delay in assembly buffer for different scenarios.


Mean (ms) Var (ms2) Mean (ms) Var (ms2) Mean (ms) Var (ms2) Mean (ms) Var (ms2)

LTh = 10 kBTR1 1.97 3.85 × 10−6 1.63 1.83 × 10−6 1.32 9.69 × 10−7 0.936 3.91 × 10−7

TR2 2.80 7.59 × 10−6 2.33 3.70 × 10−6 1.88 1.99 × 10−6 1.34 8.12 × 10−7

TR3 3.88 1.43 × 10−5 3.25 6.94 × 10−6 2.65 3.78 × 10−6 1.94 1.65 × 10−6

LTh = 20 kBTR1 3.66 9.94 × 10−6 3.20 5.81 × 10−6 2.73 3.62 × 10−6 2.11 1.87 × 10−6

TR2 5.19 1.98 × 10−5 4.55 1.16 × 10−5 3.90 7.31 × 10−6 3.03 3.81 × 10−6

TR3 7.25 3.84 × 10−5 6.37 2.21 × 10−5 5.48 1.41 × 10−5 4.35 7.65 × 10−6



0.01.001 .1 1 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time [msec]

Cum

ulat

ive

dist

ribut

ion

func

tion

PTout

=0.75

PTout

=0.50

PTout

=0.25

PTout

=0

.001 .01 .1 1 10 1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time [msec]

Cum

ulat

ive

dist

ribut

ion

func

tion

PTout

=0

PTout

=0.25

PTout

=0.5

PTout

=0.75

LTh

= 10 kB

(a) (b)

LTh

= 10 kB

LTh

= 20 kBL

Th = 20 kB

Figure 10. CDF of individual packet delay in the assembly buffer: (a) TR1 and (b)TR3.

Table 9. Estimated mean and variance of number of packets per burst for different scenarios.


Mean Var Mean Var Mean Var Mean Var

LTh = 10 kBTR1 16.47 51.90 14.99 30.52 12.80 21.33 9.53 15.72TR2 16.47 53.73 14.99 33.26 12.81 23.57 9.56 17.08TR3 16.48 54.12 15.04 34.78 12.92 25.66 9.81 18.67

LTh = 20 kBTR1 32.02 123.22 29.80 77.01 26.32 57.41 20.85 46.94TR2 31.99 129.44 29.78 85.35 26.35 65.03 20.96 52.62TR3 32.00 139.30 29.81 94.70 26.51 74.32 21.35 59.05

equal to TTh. This fraction is actually related to the packetsat the head of the burst. Consequently, at fixed input rate andLTh the fraction of this kind of packets increases with PTout .

6.5. Number of packets per burst

Now let us turn to the statistics of number of packets perburst. Table 9 shows the main statistical measures of numberof packets per burst as estimated for the scenarios introducedearlier. Also, Figure 11 depicts the frequency plots of thenumber of packets per burst for TR1 and LTh = 10 kB. Dueto the space limit we decided not show the frequency plotsfor other scenarios, whose characteristics are consistentwith those of TR1.

First, we look into the impact of increasing TTh whilekeeping the input traffic rate and the length threshold fixed.In this case, we expect that the average value of numberof packets per burst increases, which is indeed the caseas shown in Table 9. Nevertheless, this relation is not alinear one as the burst size is always limited by the lengththreshold.

Furthermore, the impact of changing the input traffic ratewhile keeping PTout fixed on the average number of pack-ets per burst is almost negligible. Nevertheless, decreasingthe traffic rate slightly increases the variance of number ofpackets, which is attributed to the fact that the traces withsmaller rates exhibit higher variations.

We also observe that the mass of the frequency plot isconcentrated around the average value so that the 98th per-centile of the distribution is located at or smaller than 2 × n̄,n̄ being the average number of packets per burst.

6.6. Smoothing impact of assembly

Finally, we investigate the impact of burst assembly pro-cess on the burstiness of the traffic. For this purpose, wefirst apply the MRA method to the burst traces at the outputof the burst assembler. The results for TR1 at LTh = 10 kBare depicted in energy plots of Figure 12. To compare withthe input traces, we further show the associated energy plotsof the input packet trace in the same figure. It is seen that the



0 50 100 1500

0.02

0.04

0.06

0.08

Number of packets

Rel

ativ

e fr

eque

ncy

0 50 100 1500

0.02

0.04

0.06

0.08

0.1

Number of packets

Rel

ativ

e fr

eque

ncy

0 20 40 60 80 100 120 1400

0.04

0.08

0.12

Number of packets

Rel

ativ

e fr

eque

ncy

0 20 40 60 80 100 1200

0.04

0.08

0.12

Number of packets

Rel

ativ

e fr

eque

ncy

PT out

=0 PT out

=0.25

PT out

=0.5 PT out

=0.75

Figure 11. Frequency plots of number of packets per burst at TR1 and LTh = 10 kB.

0 2 4 6 8 10 12 14 1527

28

29

30

31

32

33

34

35

36

Octave j (2j x 10 [msec] )

Log 2 (e

nerg

y)

PTout

=0.5

PTout

=0.25

PTout

=0

PTout

=0.75

Packet trace

Figure 12. Wavelet-based global scaling plot using waveletDaubechies 4 (bytewise) for burst trcaces at TR1 and LTh = 10 kB.

LRD property of the input traces remains untouched by theassembly process. This, however, comes at no surprise sincethe impact of the assembly process on the input traffic is onlylimited to the time scales in the range of TTh (or the maxi-mum burst assembly time for the case of PTout = 0), whichis always much smaller than the time scale associated withthe onset of LRD. Nevertheless, Figure 12 suggests that the

burst assembly does reduce the burstiness of traffic at smalltime scales. To further investigate this, we estimate CoV ofnumber of burst departures over different time intervals forTR1 as depicted in Figure 13. For the sake of comparison,for each scenario we also plot the CoV of a fictitious burstarrival process that follows the Poisson process and has thesame rate as the real burst trace of the considered scenario,denoted in the figures as Poisson.

It is seen that the burst assembly can greatly smooth outthe traffic at the small time scales such that the resultingburst trace is even smoother than the Poisson process. Infact, this smoothing impact improves with PTout , which isattributed to the fact that the variability of the burst interde-parture time decreases as the fraction of T bursts increases.We further observe that scaling the burst length thresholdinfluences the range of time scales over which the burst traceis smoother than Poisson. In fact, at a fixed input rate whenthe length threshold increases, the time threshold shouldbe scaled accordingly, and this results in an increase in thetime scale over which the input traffic is affected by theassembler.

7. PERFORMANCE ANALYSISAND DISCUSSION

In this section, we discuss the results presented in the lastsection with respect to their importance in design and per-formance analysis of the OBS network.



0 10 20 30 40 50 60 700

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6


Coe

ffici

ent o

f var

iatio

n

Assebled, PT out

=0.75

Assembled, PT out

=0.5

Assembled, PT out

=0.25

Assembled, PT out

=0

P oisson, PT out

=0.75

P oisson, PT out

=0.5

P oisson, PT out

=0.25

P osson, PT out

=0

0 10 20 30 40 50 60 700

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2


Coe

ffici

ent o

f var

iatio

n

Assembled, PT out

=0

Assembled, PT out

=0.25

Assembled, PT out

=0.5

Assembled, PT out

=0.75

P oisson, PT out

=0

P oisson, PT out

=0.25

P oisson, PT out

=.5

P oisson, PT out

=0.75

(a) (b)

Figure 13. Coefficient of variation of number of burst departures as a function of time scale at TR1: (a) LTh = 10 kB and (b)LTh = 20 kB.

First and foremost, we observe that the output of theassembler in this work and that presented in Reference [3]share some similar statistical characteristics. This is in spiteof the fact that for the analytical modelling in Reference [3]we used very simple models—i.e. Poisson—for the inputpacket-level traffic. The similarities are in the marginal dis-tribution of the burst interdeparture time as well as the shortrange burstiness of the burst-level traffic. More specifically,it is demonstrated that the burst interdeparture times areconsistent with a piecewise combination of the Gammaand negative-exponential distributions, which is in agree-ment with the results presented in Reference [3]. Also, itis revealed that the assembly process smooth out the burst-level traffic at the short time scales—in the range of theburst assembly time—so that it becomes even smoother thanthe Poisson process. Nevertheless, the scaling behaviour ofthe input traces at the large time scales remains untouched,which is also in accordance with the arguments made in Ref-erence [10]. As demonstrated in Reference [16], smoothingeffect of the assembly process can have significant impacton the design and performance optimisation of the OBS net-works, which contain either no or limited buffering capacityin forms of optical buffers. In the following we carry out aperformance study to further demonstrate this effect.

Our performance analysis experiment is carried out ona single output port of a core OBS switching node havingW wavelength channels with the total capacity of Dr Mb/s.The switch supports full wavelength conversion capabilityand as for the optical buffering we consider two differentscenarios. In the first scenario, the switch does not have anybuffering capacity, whereas in the second one the outputport of the switch is equipped with a single WDM fibre

delay line (FDL) buffer, which can delay a data burst ford s over any of its wavelength channels.§ We evaluate theloss rate—bytewise—of the considered link while loadingit with the assembled burst traffic and compare the resultswith the case when the burst arrival process is modelled byPoisson. Since in this experiment we are only interested inthe impact of the smoothing effect of the assembly on theburst arrival process, the burst sizes in case of Poisson burstarrivals are taken from the assembled burst traces. That isthe difference between the two cases is only in the burstarrival processes.

To feed the considered link with the assembled bursttraffic, we assume that the link multiplexes 10 traffic flows‖each being a burst-level flow resulting from aggregatingone of the three packet flows listed in Table 1. By varyingthe number of aggregated TR1, TR2 and TR3 flows—from0 to 10—present in the 10-flows combination, we achievedifferent utilisation levels at the output link. To aggregateeach packet trace into a burst trace, we select LTh = 10 kBand randomly set PTout in the range [0, 0.9]. Further, theresulting burst trace is circularly shifted in time so asto avoid possible synchronisation among the burst flowsgenerated from the same packet trace.

For this experiment we develop a simulation model in thediscrete event simulator OMNeT++ [20]. In our simulationwe assume Dr = 300 Mb/s and d is set to be equal to thetransmission time of an average-size burst (7500 B) over

§ The readers are referred to Reference [16] for the detailed description ofthe system model.‖ To justify this number we note that in the NSFNET with 14 nodes, 42undirectional links and 182 source–destination flows, applying the shortestpath routing we end up with 9.2 flows per link in average.



0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.750.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

(a) (b)0.45

Offered load

Dro

p ra

te

TracePoisson

0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.750

0.05

0.1

0.15

0.2

0.25

0.3

0.35

Offered load

Dro

p ra

te

TracePoisson

W=2

W=4

W=1

W=1

W=2

W=4

Figure 14. Lossrate over the core link: (a) no FDL buffer and (b) single FDL buffer.

the link. Depicted in Figure 14 are the data loss rates asfunction of the link utilisation—with the associated 90%confidence intervals—at different number of wavelengthchannels W = {1, 2, 4} and for both buffering scenarios.The results for the assembled burst traffic is denoted asTrace and those associated with the Poisson burst arrival isdenoted as Poisson.

We observe that the loss rate for the case with the assem-bled burst trace is always smaller than that with the Poissonburst arrival process for the considered scenarios. The dif-ference is noticeable for the link with and without FDLbuffers, though the difference in the former case is larger.Specifically, for the link with single buffer we observe upto more than 100% increase in the loss rate when the trace-based burst arrival is replaced with the Poisson process.

8. CONCLUSION

We presented the results of a case study on the burst assem-bly process in an OBS network using real traces collectedfrom the Internet. The results throw light on statistical char-acteristics of the traffic at the output of the burst assemblyunit. Also, through simulation experiments we demon-strated that the burst assembly process positively influencesthe performance of the network.

From these results we can benefit in two ways. First, theydemonstrate that a target performance of the network canbe achieved with lesser resources than are usually assumed.Also, the results emphasise that design and performanceevaluation of new architectures and protocols can be accu-rately carried out with much simpler traffic models, which

means less efforts. These are in line with our findings inReference [16] about the impacts of burst assembly andalso in accordance with the conclusion made in Reference[7] that the influence of the self-similarity present in theinput packet traces is negligible with regards to the lossprobability in core of OBS networks.

Moreover, we demonstrated the importance of the role oftime out probability PTout in characterising the burst-leveltraffic. Consequently, we delved into the relation betweeninput parameters and PTout and derived a rule of thumbfor approximating the value of the time threshold at whichPTout = 0.5 only based on knowing the average input trafficrate and the applied length threshold. The characterisationof PTout together with the derived rule of thumb simplifythe design and set-up of the burst assembly unit in practicalsituations.

ACKNOWLEDGEMENT

The authors thank Joerg Wallerich (T-Systems, Germany) for hiscontribution to the early phase of this work. This work was sup-ported in part by the European Commission (7th ICT FrameworkProgramme) within the Network of Excellence project BONE(Building the Future Optical Network in Europe) and within theCollaborative project 4WARD.

REFERENCES

1. Qiao C, Yoo M. Optical burst switching (OBS)—a new paradigm foran optical internet. Journal of High Speed Networks 1999; 8(1):69–84.

2. Turner JS. Terabit burst switching. Journal of High Speed Networks1999; 8(1):3–16.



3. Rostami A, Wolisz A. Modeling and synthesis of traffic in opti-cal burst-switched networks. Journal of Lightwave Technology 2007;25(10):2942–2952.

4. Laevens K. Traffic characteristics inside optical burst switched net-works. In Proceedings of Optical Networking and Communications(Opticomm), 2002; 137–148.

5. Rodrigo MV, Goetz J. An analytical study of optical burst switchingaggregation strategies. In Proceedings of Broadnets, Workshop OBS,2004.

6. Yu X, Chen Y, Qiao C. A study of traffic statistics of assembled bursttraffic in optical burst switched networks. In Proceedings of OpticalNetworking and Communications (Opticomm), 2002; 149–159.

7. Izal M, Aracil J. On the influence of self-similarity on optical burstswitching traffic. In Proceedings of Globecom, Vol. 3, 2002; 2308–2312.

8. Mountrouidou X, Perros H. Characterization of the burst aggrega-tion process in optical burst switching. In Proceedings of Networking,2006; 752–764.

9. Ge A, Callegati F, Tamil LS. On optical burst switching and self-similar traffic. IEEE Communications Letters 2000; 4(3):98–100.

10. Yu X, Li J, Cao X, Chen Y, Qiao C. Traffic statistics and performanceevaluation in optical burst switched networks. Journal of LightwaveTechnology 2004; 22(12):2722–2738.

11. Izal M, Aracil J, Morato D, Magana E. Delay-throughput curves fortimer-based obs burstifiers with light load. Journal of Lightwave Tech-nology 2006; 24(1):277–285.

12. Figueiredo GB, da Fonseca NLS, Melo CAV, Salvador MR. On thetransformation of multifractal traffic at ingress optical burst switches.In Proceedings of IEEE ICC, 2006.

13. Mountrouidou X, Perros H. On traffic modeling of burst aggregationalgorithms using video and data traces. In Proceedings of GridNets,2007.

14. Garcia NM, Monteiro PP, Freire MM. Burst assembly with real ipv4data—performance assessment of three assembly algorithms. In Pro-ceedings of NEW2AN, LNCS 4003, 2006; 223–234.

15. Xiong Y, Vandenhoute M, Cankaya H. Control architecture in opticalburst-switched wdm networks. IEEE Journal on Selected Areas inCommunications 2000; 18(10):2062–2071.

16. Rostami A, Wolisz A. Impact of edge traffic aggregation on the perfor-mance of FDL-assisted optical core switching nodes. In Proceedingsof IEEE ICC, 2007.

17. Abry P, Veitch D. Wavelet analysis of long-range-dependent traffic.IEEE Transactions on Information Theory 1998; 44(1):2–15.

18. The Perl Programming Language. Available online at: https://www.perl.org

19. Draper NR, Smith H. Applied Regression Analysis (3rd edn). JohnWiley & Sons: New York, 1998.

20. Discrete event simulation system OMNeT++. Available online at:http://www.omnetpp.org

AUTHOR’S BIOGRAPHY

Ahmad Rostami received the B.Sc. degree (with honors) in electrical engineering from Shahed University, Tehran, Iran, in 2000 andthe M.Sc. degree in electrical engineering from Amir-Kabir University of Technology, Tehran, in 2003. Since May 2004, he has been aResearch Assistant with the Telecommunication Networks Group (TKN), Technical University of Berlin. From 2001 to 2003, he wasan R&D Engineer at the Iran Telecommunication Research Center, Tehran. His research interests are in optical WDM networks, witha focus on traffic modeling and performance-related issues of optical transport networks.

Adam Wolisz received the B.S., Ph.D., and Habilitation degrees from the Silesian University of Technology, Gliwice, Poland, in1972, 1976, and 1983, respectively. After a period with Polish Academy of Sciences (until 1990) and GMD-Fokus, Berlin, Germany(1990–1993), he joined Technische Universität Berlin, in 1993, where he is the Chaired Professor for Telecommunication Networksand the Executive Director of the Institute for Telecommunication Systems. He is also an Adjunct Professor with the Departmentof Electrical Engineering and Computer Science, University of California, Berkeley. His research interests are in architectures andprotocols of communication networks. He is currently focusing mainly on wireless/mobile networking and sensor networks.

Anja Feldmann is a full professor at Deutsche Telekom Laboratories a unit of Deutsche Telekom and an An-Institut of the TechnischeUniversitaet Berlin, Germany. From 2000 to 2006 she headed the network architectures group first at Saarland University and then at TUMuenchen. Before that (1995 to 1999) she was a member of the Networking and Distributed Systems Center at AT&T Labs – Researchin Florham Park, New Jersey. She has published more than 50 papers and has served on more than 40 program committees, includingas Co-Chair of Sigcomm 2003 and as Co-PC-Chair of Sigcomm 2006 and IMC 2009. She received a M.S. degree in Computer Sciencefrom the University of Paderborn, Paderborn, Germany, in 1990 and M.S. and Ph.D. degrees in Computer Science from CarnegieMellon University in Pittsburgh, USA, in 1991 and 1995, respectively. Her current research interests include new network architecturesas well as understanding the current Internet and its application for the purpose of performance debugging and intrusion prevention.


Date post:	04-Apr-2020
Category:	Documents
Upload:	others
View:	2 times
Download:	0 times

Traffic analysis in optical burst switching networks: a trace-based … · 2009-11-10 · TRAFFIC...

Documents