Lecture on Wirel e ss N e tw o rk Mea s urements & Mode l ing

1. LectureonWirelessNetwork Measurements&Modeling Prof.Maria Papadopouli UniversityofCrete ICS-FORTH http://www.ics.forth.gr/mobile 1

2. Empirical measurements • Canbe beneficial inrevealing • deficienciesof awirelesstechnology • different phenomenaofthe wirelessaccess& workload • Provide data formodellingeffortsaiming toproducemorerealistic models&synthetic traces • Enablemeaningful performanceanalysis studiesusingsuch • empirical, synthetic tracesandmodels

3. Network workload can be characterized based on: • Amount of traffic • Number of packets or bytes • Downloading vs. uploading • Packet or flow arrival process • Interactivity model • Application type • Usage pattern • Network topologies can be described based on • Their connectivity • Link charactertsics • Distribution & density of peers • Degree of clustering • Co-residency time • Inter-contact time • Duration of disconnection for the Internet • Interaction patterns • Network performance benchmarks include: • Jitter • Latency/delay/response delay • One way, round-trip delay • Packet losses • % packet loss, burstiness of packet loss, relative position of the bursts of packet losses within a flow • Subjective metrics (e.g., opinion score)

4. Modeling TrafficSizes of files stored on web servers, data files transferred through the Internet, and files in general purpose Unix file-systems modeled using heavy-tailed distributionsInternet traffic (amount of bytes) exhibits a self-similar nature.Web traffic also exhibits self-similarity. The majority of web traffic in wired networks is below 10KB, while a small percentage of vey large flow account for 90% of the total traffic. Power laws can describe web flow sizes.So Poisson processes cannot accurately model the traffic load. However, they can be used to model the arrival of user sessions (e.g., telnet connections or arrivals at a wireless AP).

5. Poisson process • Stochastic process which counts the number of events and the time that these events occur in a given time interval. • The time between each pair of consecutive events has an exponential distribution with parameter λ and each of these inter-arrival times is assumed to be independent of other inter-arrival times • Exponential distribution: PDF X: Prob(X=x)=λ e-λx, x  [0, ), E[X] = 1/λ, Var[X] = 1/λ2 Power law, Zipf distribution, Pareto distribution are “heavy-tail” distributions Pareto:

6. a self-similar object is exactly or approximately similar to a part of itself (i.e. the whole has the same shape as one or more of the parts). A self-similar object “looks” roughly the same on any scale

7. PropagationModels • Oneofthemost difficultpart of theradiochannel design • Doneinstatisticalfashionbasedon measurementsmadespecifically foranintendedcommunicationsystemor spectrumallocation • Predictingthe averagesignalstrengthat agivendistance fromthe • transmitter 3

8. SignalPower Decaywith Distance • Asignaltravelingfromone nodeto anotherexperiences fast (multipath)fading,shadowing& pathloss • Ideally,averagingRSS oversufficientlylong timeinterval • excludes theeffectsofmultipathfading&shadowing  • generalpath-lossmodel: • _ • P(d) =P0–10n log10 (d/do) • n:pathlossexponent • P(d):theaveragereceivedpower in dBat distanced P0 is thereceivedpower in dBat a shortdistanced0 _

9. Signal Power Decaywith Distance • Inpractice, the observationinterval is notlong enough tomitigate • theeffects of shadowing • Thereceived power is commonlymodeledtoincludebothpath-loss • &shadowingeffects, the latterofwhicharemodeledasazero-mean Gaussianrandomvariable withstandard deviation σshin the logarithmicscale,P(d),in dB canb_e expressed: • P(d) ~ N(P(d), σsh2) • Thismodelcanbe usedin bothline-of-sight(LOS)&NLOSscenarios withappropriatechoiceof channelparameters

10. Important aspects of monitoring • Identification of the dominant parameters • based on the aspects you need to measure, the parameters that need to be monitored are decided … • Strategic placement of monitors • e.g., at routers, APs, clients, and other devices, … • Automation of the monitoring process to reduce human intervention in managing the monitors and collecting data • Aggregation of data collected from distributed monitors to improve the accuracy, while maintaining a low communication and energy overhead • (cross-)validation study to verify that the collected traces correspond to representative conditions

11. Monitoring • Dependingontypeofconditionsthatneed to bemeasured, • monitoringneedsto beperformedat • Certainlayers • Spatio-temporalgranularities • Monitoringtools • Arenotwithoutflaws • Severalissuesarise whentheyareusedinparallelfor • thousandsdevicesofdifferenttypes& manufacturers: • Fine-graindatasampling • Timesynchronization • Incompleteinformation(missing values, incorrect values) • Dataconsistency • Vendor-specific information & dependencies (often not publicly available)

12. Monitoring tools • Fine-graindatasampling • What is its spatio-temporal granularity? • Timesynchronization • Clocks have different drifts… • Incompleteinformation(missing values, incorrect values) • How do you handle missing values? Explain what caused them! • Various techniques to “fill” the gaps, if necessary, such as interpolation, model-prediction, matrix completion and other sophisticated statistical analysis methods … • Explain the outliers! Are there due to misconfigured devices? Or network anomalies? Or due to the occurrence of “extreme” phenomena? • Dataconsistency • Vendor-specific information & dependencies (often not publicly available) • e.g., RSSI differs depending on the manufacture

13. Monitoring &DataCollection • Finespatio-temporaldetail monitoringcan • Improvetheaccuracyof the performanceestimatesbutalso • Increasetheenergyspendingsanddetectiondelay • Network interfacesmayneedto • Monitorthechannel infiner &longertimescales • Exchangethisinformationwith other devices

14. Challengesin Monitoring (1/2) • Identificationofthedominantparametersthrough • sensitivity analysisstudies • Strategicplacementofmonitorsat • Routers • APs • clientsandother devices ofusers • Automationofthe monitoringprocess toreduce humanintervention • inmanagingthe • Monitors • Collectingdata

15. Challengesin Monitoring (2/2) • Aggregationofdatacollectedfromdistributedmonitorstoimprove the accuracywhilemaintaining low overheadintermsof • Communication • Energy • Cross-layer measurements,collecteddataspanningfromthe physical layeruptotheapplicationlayer,are required

16. WirelessNetworks • Areextremely complex • Havebeen usedformanydifferentpurposes • Havetheirowndistinctcharacteristicsduetoradiopropagation characteristics&mobility • e.g., wirelesschannels can behighlyasymmetricandtimevarying • Note: • Interactionofdifferentlayers& technologiescreates many situations that cannotbe foreseenduringdesign& testing stages oftechnology development

17. Empirically-based Measurements • Real-lifemeasurementstudiescanbe particularlybeneficialin • revealing • deficienciesofa wirelesstechnology • different phenomena ofthe wirelessaccessandthe workload • Richsetsofdatacan • Impelmodelingeffortstoproducemore realisticmodels • Enablemoremeaningful performanceanalysisstudies

18. It is important to assess whether typical assumptionsare realistic! Typical assumptions in performance analysis of wireless networks • Models&analysis ofwired networksare valid for wirelessnetworks • Wirelesslinksaresymmetric • Linkconditionsarestatic • Thedensityofdevicesin anareais uniform • Thecommunicationpairsare fixed • Usermobilityis basedonrandom-walk models

19. Wireless Access Parameters • Traffic workload • Indifferenttime-scales • Indifferentspatial scales(e.g., AP,client,infrastructure) • Inbytes,number ofpackets, number offlows,application-mix • Delays • Jitter anddelay per flow • Statisticsat anAP and/orchannel • Packet losses • %packet loss & burstiness of packet losses • Usermobilitypatterns • Linkconditions & channel quality • Networktopology

20. TrafficLoadAnalysis • Asthe wirelessuserpopulationincreases,the • characterizationoftrafficworkloadcan facilitate • Moreefficient network management • Betterutilizationofusers’scarceresources • Application-basedtrafficcharacterization

21. HourlySessionarrival rates 15

22. Traffic Load at APs • Widerangeofworkloadsthatlog-normalityisprevalent • Ingeneral,traffic loadislight,despitethe longtails • No clear dependencywithtype ofbuildingtheAPislocatedexists • Althoughsomestochastic orderingis presentin • Tailofthedistributions • Dichotomy amongAPsisprominentin bothinfrastructures: • APs dominatedbyuploaders • APs dominatedbydownloaders • As the totalreceivedtraffic atanAP ↑ • There is also ↑ inits totaltraffic sent • ↓ inthesent-to-receivedratio

23. Trafficload at APs • Substantialnumberofnon-unicastpackets • Numberofunicastreceivedpackets strongly correlatedwith • numberofunicastsentpackets(in log-logscale) • MostofAPssend& receivepacketsofrelativelysmallsize • SignificantnumberofAPsshow ratherasymmetricpacket sizes • APswith largesent&smallreceive packets • APswithsmallsent& largereceivepackets • Distributionsofthenumberofassociations& roaming operationsareheavy-tailed • Correlationbetweenthetraffic load& numberof associations • inlog-logscale

24. Ingeneral,the traffic loadis light longtails RAREEVENTSOF LARGE SIZE

25. Wide-rangeof workloads Astotalreceivedtraffic at an AP ↑ its totaltraffic sent↑

26. Light trafficload Wide ranges & dichotomies APs with uploaders APs with downloaders Rare events of large size

28. Application-basedTraffic Characterization • Usingportnumbers to classify flowsmayleadtosignificantamounts • ofmisclassifiedtrafficdueto: • Dynamicportusage • Overlappingportranges • Trafficmasquerading • Oftenpeer-to-peer&streamingapplications: • Usedynamicportstocommunicate • Portranges ofdifferentapplicationsmayoverlap • Maytrytomasqueradetheir trafficunderwell-known “non- • suspicious”ports, suchasport80

29. Desirable Properties forModels • Accuracy • Tractability • Scalability • Reusability • “Easy”interpretation

30. Relatedwork • Richliteraturein trafficcharacterizationin wirednetworks • Willinger,Taqqu,Leland,Parkonself-similarityof EthernetLANtraffic • Crovela,BarfordonWeb traffic • Feldmann,PaxsononTCP • Paxson,Floyd onWAN • Jeffay,Hernandez-Campos,SmithonHTTP •  •  •  • Trafficgeneratorsfor wired traffic • Hernandez-Campos,Vahdat, Barford,Ammar, Pescape,… • P2Ptraffic • Saroiu, Sen,Gummadi,He,Leibowitz, … • On-linegames • Pescape, Zander, Lang,Chen,… • Modellingofwirelesstraffic • Menget al.

31. DimensionsinModelingWirelessAccess • Intended userdemand • Usermobility patterns • Arrivalat APs • RoamingacrossAPs • Channel conditions • Networktopology

32. Mobilitymodels • Groupor individualmobility • Spontaneousorcontrolled • Pedestrianor vehicular • Knowna priori or dynamic • Random-walkbasedmodels • –Randway model in ns-2 • Markov-basedmodel

33. A VerySimple ChannelModel Gilbertmodel Computestationaryprobabilities Pib Pbb Idle Busy • Pii • Pbi • Achannelcanbeinthe idleorbusystate • Markov-based modelallowsusto determine: • Howmuch timethesystem spendsineach state • Probability ofbeingin a particularstate • Inrealrife, there is non-stationarityduedynamic phenomena

34. Markov chain model: (definition)random process usually characterized as memoryless: the next state depends only on the current state and not on the sequence of events that preceded it. Example: • Thestatesrepresentwhethertheeconomyisin a bullmarket, a bearmarket, or a recession, during a givenweek: • a bullweekisfollowedbyanotherbullweek 90% ofthetime, a bearmarket 7.5% ofthetime, and a recessiontheother 2.5%. Labelthestatespace {1=bull, 2=bear, 3=recession} Define the transition matrix:

35. (Continuing from the previous slide) From this figure, it is possible to calculate the long-term fraction of time during which the economy is in a recession, or on average, how long it will take to go from a recession to a bull market • Thedistributionoverstatescanbewrittenas a stochastic row vector xwiththerelationx(n + 1) = x(n)P. Soifattime n thesystemisinthe state 2 (“bear”) then, attime n + 3thedistributionis

36. UNC/FORTHweb archive Onlinerepositoryofmodels, tools,andtraces – Packetheader,SNMP,SYSLOG, synthetic traces, … http://netserver.ics.forth.gr/datatraces/ Free login/passwordtoaccessit Simulation&emulationtestbedsthat replaysynthetictraces forvarioustraffic conditions Mobile ComputingGroup@UniversityofCrete/FORTH http://www.ics.forth.gr/mobile/ maria@csd.uoc.gr