small but steady insider population is present. Naturally, a lower
c i
generally results in moreinsiders (although, therearesome utuations).
Again,ifwelookatthedown-saledplot,aninterestingtrendemerges: evenatomparatively
high entry osts, there is a steady-rate growth in tehnology penetration. This implies that
wireless ommunities networks are in an expanding phase today. This shows the possibilityof
theemergene ofa global wireless village.
high-relevane users drive thetehnology diusion. Moreover, theurrent real-world
pen-etration level and struture of global ommunity networksould imply steady growth for
the future.
While relaxing assumptions and using more omplex analyti tehniques might be hard in
this senario, there areplentyof pratial researhhallenges to takle inthefuture.
Mediator payo and business models. While thebusiness model we used inthis work
is based on the FON onept, we annot yet determine if it is optimal in any (soial welfare,
maximummediatorprot,et.) aspet. Further researhapplyingmehanismdesigntehniques
to theproblemofreating andevaluatingeientbusiness modelsforuser-providednetworking
is highly relevant. Additionally, more sophistiated user ategories an be introdued to rene
revenueow. Ontheother hand,the existeneofmultiplemediators indueompetition among
them,whihanbealsotakenintoonsideration. Last,theimpatofaglobalwirelessommunity
network on 3G/4Gmobile operators shouldalso be studied (see [64℄for guidelines).
Mobility patterns. Forour simulationswe onstrutedmobilitygraphsasrealisti as
pos-sibleintermsof usersloations, mobilitypatternsand roamingintensities. Howeverwe believe,
thatsuh anetworkevolvesaordingto theurrent stateoftheuser, ISPandmediator games,
whih eventuates a dynami struture, withtemporally hanging harateristis. Although the
understanding and modeling of the dynami properties of networks and ommunities is still in
heavydevelopment[76℄[77℄[78℄,itwouldbeinterestingtoseehowthedenedgamerulesinterat
withtheunderlying roaming struture.
Trameasurements. ModelingISPostsproportionaltotra isfar fromtrivialinthis
ontext. Foranadvanedostmodeloneshouldtakeintoaounttraandusageharateristis
as measured in the very network. This would inlude dierent tra proles for users (e.g.,
heavy hitters vs. e-mail and browsing). Also,the utility ofusers is greatly dependent on tra
prioritizationtehniquesandqualityofexperiene. Whiletherearesomeexistingsurveysdealing
withtheFONsystem[79℄,we still laka large-salemeasurement study to build upon.
Regulatoryaspets. Aountability(tyingthetratoagivenuser)andthefeasibilityof
lawful intereption are hallenging issues: these features may be neededif/when a truly global
wireless villageemerges. Boththelegalframeworkanditsenabling implementationarefar from
trivial: these arepossible diretionsfor a dierent branh ofhighlyinteresting futurework.
Congestion Control in Dynami Mobile
Environments
Limitations of urrent TCP protools and espeially their performane in high-speed
environ-ments are demonstrated in numerous studies [95℄ [96℄ [97℄ [98℄. As one of themain dierenes
to today's networks, it is expeted that future protools will have to operate on muh higher
speeds. For this end, high-speed variants of TCP are being developed and even deployed to
performbetterandbemoreadaptive tourrentandfuturetehnologies withhighbandwidthor
high bandwidth-delay produt (BDP). Suh links are beoming widespread as both wired and
wireless networks oer higher aess speeds, and the number of users and bandwidth-hungry
appliations inreases (e.g.,peer-to-peer lesharing networks).
Onthe other hand, wireless networks have always been hallenging for TCP, espeially for
ongestion ontrol algorithms. However, the evaluation of the topi is ambiguous; some
re-searhers argue that wireless links do not impose serious problems to TCP [85℄ [99℄ [86℄ [101℄,
while others deal with TCP's limitations in wireless environments in detail [102℄ [103℄. Sine
wireless tehnology approahestransmissionrates previously unheardof,futureongestion
on-trolprotoolshavetodealwithbothwirelessand high-speednetworkharateristisatthesame
time.
Further ompliating the issue, heterogeneity is emerging as a key harateristi of urrent
and future wireless networks. With the rih diversity of wireless networks already deployed
(WLAN, GPRS,EDGE,UMTS, HSPA,CDMA2000)or likely tobedeployedinthenearfuture
(3GPP-LTE/SAE, WiMAX),thedemandforommuniationmehanismswhihan operate
ef-iently overthe physial satteredness ofaesstehnologies issubstantial. Even a singledata
owanexperieneawidevarietyofunderlyingnetworktehnologies, thereforeswiftadaptation
to hanged onditions is essential to provide a high-quality servieto end-users. An espeially
hallengingtaskforTCPistoopewithhandovers: theseareinherentlypresentinmobileellular
Figure4.1: Measuredround-trip timesduring a handover
systems. Sine thedeployment of new wireless tehnologies are done inan inremental manner
thusseveraltehnologieso-exist,inter-tehnologyhandoversour. Theseinter-tehnology
han-dovers an result in a sudden inrease inlink apaity(referred to asan up-swith) as seenby
TCP. E.g.,ifausermovesfroma HSPA ellto an LTEell theresulting up-swith anbe tens
of Mb/s.
Theontribution of thishapter istwofold. InSetion 4.2, we investigate whether proposed
high-speed TCPvariantsperformwellinurrentandfuturemobilenetworkswithspeialregard
to handovers. Sine motivation for high-speed TCPs was rst neessitated by wired systems,
peuliaritiesofmobilenetworkshavenotalwaysbeen onsideredduringprotooldesign. Onthe
other hand,existing studies fous on today's networks, and do not try to predit what impat
would future hanges have on the same protools. Our goal is to provide measurement results
of high-speed TCP proposals in emulated future mobile environments and also to suggest the
inlusion of mobile-spei issues in the IRTF TMRG draft [104℄, and other douments. In
Setion 4.3 we propose SpeedDetet, a TCP extension for handling apaity up-swithes. The
keyidea ofour proposalis toimprove apaityutilization atup-swithes byexpliitly deteting
freeapaitythentriggeringtheunderlyingongestionontrolalgorithmtoanaggressiveprobing
regime. We argue that this an be done by monitoring measured round-trip times: a sudden
derease of RTT means that the bottlenek buer is empty as a result of inreased serviing
apaity(seeFigure4.1). Notethataseond,thoughmuhlesslikelyreasonforRTTdereaseis
thatthepaketsarere-routedthroughanewpathwithsmallerpropagationRTT.Theourrene
of this event isvery rareinthelifetimeof agiven ow, hene we omitits analysis.
4.1 Bakground
4.1.1 TCP variants for high-speed networks
In our work we examined four high-speed TCP variants and NewReno. The latter is inluded
for referene asthe most widely usedTCP avor today. Inthefollowing we shortlysummarize
themain featuresof the ongestionontrol algorithms inthese protools.
TCP's ongestion ontrol is window-based; that is, the amount of outstanding data in the
network is diretly ontrolled by the protool. The instantaneous sending rate at time
t
anbe estimated as
λ(t) = W (t)/T (t)
whereW (t)
is the ongestion window size andT (t)
is theround-trip time.
NewRenoTCP.ThisisthelatestversionofthelassialTCP[105℄. Theongestionontrol
isomposedoftwo phases,probing andderease. Probing inturnismadeup ofan exponential
inrease and a linearphase, slow start and ongestion avoidane, respetively. Whena timeout
or three dupliate ACK-s noties the sender of a loss event, the probing phase ends, and the
windowis eitherhalvedor dereased to 1. Formally,
W ← W + a(W )
on ACKreeption, and
W ← W · (1 − b(W ))
after thethirddupliate ACK,with
a(W ) = 1/W
andb(W ) = 1/2.
Thisalgorithm is oftenreferredto asAIMD(Additive Inrease, Multipliative Derease).
HighSpeed TCP. ThisalgorithmmodiesNewReno's inreaseanddereaseparameters to
ahievelargerongestionwindows[81℄. Funtions
a(W )
andb(W )
arederivedfromtheresponsefuntion
W
HS(p)
thatdesribestherelationshipofW
andthesteady-statepaketlossprobabilityp
. The response funtion for HighSpeed is a modied version of NewReno's response funtionW
NR(p) = 1.2/ √ p
. Inpartiular,W
HS(p) =
W
NR(p)
ifW 6 W
low, 0.12/p 0.835
ifW > W
low,
thatleads to
b(W ) = f (log W )
anda(W ) = f
W 2 b(W ) 2 − b(W )
.
Salable TCP. SalableTCP hanges theoriginal algorithm toupdatetheongestion
win-dow independently of
W
, suh thata(W ) = 0.01
andb(W ) = 0.125
[82℄. This strategy isessentiallyMIMD (Multipliative Inrease,Multipliative Derease).
BIC TCP. BIC stands for Binary Inrease Congestion, as ongestion window setting is
based on a binarysearh [83℄. On apaket loss, BIC rst reords
W
max= W
,then reduesW
aordingtoNewReno'salgorithmwith
b(W ) = 0.2
,andnallyreordsthenewwindowasW
min(eetively
W
min= (1 − b)W
max). Then, using thetwo extremes
W
minand
W
maxitperformsa
binary searh by jumping to their mean value, and swithing to a linear inrease phase. When
W = W
maxis reahed BIC starts an aggressive searh for thenew
W
max(max probing) until a
new lossours.
FAST TCP. FAST TCP omplements loss-based ongestion ontrol with queuing delay
estimation[84℄. Inpratie,
W
isupdatedperiodiallyusingT 0
astheminimumRTT(onsideredaspropagation-only omponent),and
T
asthelast measured RTT.W ← (1 − γ )W + γ T 0
T W + α
(4.1)
The aggressiveness ofthealgorithm is determined byonstants
α
,whihrepresents thenumberof thepakets to be heldinthebuer ofthebottleneklink at theequilibriumstate,and
γ
,anexponential smoothing parameter(
0 < γ 6 1
).4.1.2 TCP performane in wireless mobile networks
There isasigniant bodyofresearhrelatedto theinteration ofmobilewirelessenvironments
and TCP ongestion ontrol. A number of papersdeal with theperformane of standard TCP
in wireless networks [86℄ [99℄ [101℄. Others proposemodiations to standard TCP [85℄ [89℄ or
totally new ongestionontrol protools [87℄ [88℄[100℄ to dealwithproblems induedby mobile
networks.
The behaviorof TCP andother TCP-friendly ongestion ontrol shemes during handovers
arepresentedin[106℄ and[107℄. Although thesepapersdealwithhandovers theyassumeonlya
moderatelyhangingenvironment. AuthorsofTCP-Ho[108℄dealwithTCPperformaneduring
link outages, whih an be onsidered as a simple intra-tehnology handoverfrom TCP's point
of view. Furthermore, there is a paper assessing the performane of high speed TCP versions
during mobilehandovers[C7℄, thatwe buildon. Oneof themainresults presentedthere isthat
urrenthigh-speedTCP versionshave troubledealingwithsuddenapaityinreasesanddrops.
Withregard to non-ongestion detetion, an existing solution is Persistent Non-Congestion
Detetion (PNCD)part of TCP Westwood (TCP-W)whih is based on ontinuous rate
es-timation [88℄. PNCD raises a notiation whenever the ahieved sending rate is signiantly
lower than a possible estimated rate. The reation to the notiation generated by PNCD in
ssthresh values,thatontinuesuntila paketlossevent. A shortomingofAgileProbing isthat
oneit starts, it lls upthe buerompletely ausing unneessarypaket drops. Also, TCP-W
bases its rate estimation on minimum RTT measurements, that an be misleading in ertain
situations [109℄.
TCP-Hando is proposed in [92℄ to expliitly deal with handovers in mobile networks. A
shortoming ofthis solution is theuseof dediatedhandoversignaling between TCP-HOlient
andserver. SinethehandovermaynotbeseenbythehostOSatall,adediatedontrolmethod
is neessary to inform the host TCP about a handover. This solution introdues additional
overheadand omplexity to ongestionontrol.
In[102℄ and[104℄authors introduesguidelines foradequateevaluationofTCP performane
inwirelessandwiredenvironments. Weadheretotheseguidelines inourperformaneevaluation
presentedinSetions4.2.3, 4.3.3 and4.3.4.