Conlusion

Inthishapterwehaveinvestigatedtheperformaneofongestionontrolmehanismsinmobile

environments.

First,we have presenteda measurement study on theperformane of existingstandard and

high-speed TCP variants in a futuristi mobile environment (4.2). We measured throughput

adaptationusingouremulationtestbed,enteredaroundmobilenetworkissues: linkinterruption

and large sale apaity hange.Thekey takeawaysare:

•

^{All examined protools exept for FAST adapt slowly to sudden large sale apaity}

in-reases. Suh inreases are expeted to happen frequently in a mobile environment, and

protoolsneed toimplementmehanismstodetetandutilizethefreeavailablebandwidth

muhfaster asthey doit urrently.

•

^{We found that FAST TCP an perform poorly after an} inter-system down-swith. Suh handovers may onfuse FAST, and depending on stati parameter settings this may lead

to drasti link underutilization. The problem is the onsequene of FAST relying on its

bogusassumption on the base propagation delay. To solve these problems, thebaseRTT

estimation needsto bexedin FAST.

•

^{We evaluatedthe}probabilityofbueroverruns atthebottleneklinkduring linkoutages, and suggesteda novel solutionfor the problem.

Seond, motivated by the results above, we have proposed SpeedDetet, a simple TCP

ex-tension for handling sudden apaity inrease (4.3). We have shown how this an be ahieved

whileaomplishing our ve design goals: minimality,easy deployability, nosignaling overhead,

fairness andeieny. Our mainndings arethefollowing:

•

^{We have proposedanovelalgorithm,} SpeedDetet,for detetingandreatingpromptly to suddenly available unused apaity. The basi idea of SpeedDetet is omparing a

long-term RTT estimate to the srtt ommonly found in standard TCP implementation; if the

dierene issigniant asudden apaity up-swithis likely to have happened.

•

^{We presented how optimal} sensitivity ofSpeedDetet an be ahieved throughparameter tuning.

•

^{We showed by testbed} measurements and extensive simulations that our proposed ap-proah:

(i)

^{signiantly improves TCP's reation time to apaity inreases;}

(ii)

outper-forms PNCD (part of TCP Westwood [88℄) interms of reation time and onsisteny for

a wide-range of settings;

(iii)

^{retains the} intra-protool fairness of the underlying TCP variant, while being onsiderably fair to standard TCP; and

(iv)

^{workswell both inase}

of mobilehandoversand general owdynamis.

We argue that the SpeedDetet extension an be implemented on top of virtually any

loss-based TCP version thatkeeps RTT estimates. A possible diretion of future work ould be to

implement our mehanismon top of other establishedongestion ontrol methods and evaluate

their performane. Furthermore, a performane omparison against delay-based TCP versions

ouldbebeneial. Additionaly,SpeedDetetouldbeanalyzedatthesub-owlevel,buildingon

topofamulti-pathtransportprotool. Finally,aeldstudy inanoperationalmobilebroadband

provide us with invaluable measurement dataand helpus evaluate SpeedDetet ina real-world

environment.

Summary

Wesummarize our ontributions anddisuss their pratial appliation possibilitiesbelow.

5.1 Priing Internet aess under user inuene

InChapter2,westudiedmultiplefaetsoftheimpatofuserinueneonInternetaesspriing.

5.1.1 Contribution

In Setion 2.2, rst, we showed how loyalty introdued a new equilibrium in a repeated game

setting resultingin theooperationof ISPs. Seond,we investigated theaseof adierentiated

ustomer population by introduing dual reservation values, and show how it led to new, pure

strategyNashequilibriaindiatingthatISPsouldmakethemost protbymilkingtheir

respe-tive loyaluser base. Third, we onstruted two novel models forustomer loyaltyinorporating

two important aspets of the users' purhasing deisions: prie sensitivity and inherent human

unertainty. Weevaluated the impatofuser loyaltythrough thesemodels byextensive

simula-tions ina numberof relevant senarios. A key nding was that a higher level of loyalty in the

userpopulationledtolargerprotsfor theISPs. Last, wedisussedhowtheinterplayofservie

bundling and userloyaltyan be modeledand analyzedvia simulationand extendedgames.

In Setion 2.2, we presented results on a novel priing mehanism for Internet Servie

Providers alled user-inuened priing, where users an vote for the exat priing sheme

im-plemented inthe next billing yle. Our assumptions were that ISPs want plannable revenues,

while users want to keep their osts low. We showed that under user-inuened priing, users

of dierent tra volumes (heavy,medium, light)an ooperate toahieve lower ostsutilizing

side-payments. Wemodeledthisproessasaweightedooperativevotinggame, andderivedthe

equilibrium solutions and payos on the individual user and group level. We showed how the

ratio ofdierent users andmaximumreasonableside-payments aettheoutome ofthevoting

game. Wealsoderivedthedistributionofpowerinvariousregimes ofthegame. Resultsindiate

that mediumusers arepivotal inthedeision-making proess.

5.1.2 Appliation possibilities

Results presented in Setion 2.2 proves that user loyalty is an important fator when priing

Internet aess. This should open up theeyesof servieproviders and make them reward loyal

ustomers with premium servie instead of getting into prie wars for new ustomers. The

aompanying simulationstudypresentedin[J2℄an provide an inputtothepriingstrategy of

Internet Servie Providers. Onthe other hand, the user-inuened priing sheme (see Setion

2.3) is designed for atual use by ISPs. There are a number of pratial questions onerning

theimplementation of the mehanism; please refer to 2.3.5for a detaileddisussion.

In document B daeUiveiyfTe h (Pldal 126-130)