Measurement results

Our algorithm an be implemented on top of any ongestion ontrol algorithm that measures

RTT samples(mostloss-basedTCPs). Wehoseto implement SpeedDetetontop ofNewReno

and BIC TCP, the two versions most widespread today. We refer to these modied version as

RenoSD and BICSD. Our Linux-based, kernel-level implementation ontains about 40 lines of

C ode. The algorithm impliessmall omputation overhead: three extra words of memory per

TCP ow, and up to three extra basi operations (omparison/assignment) per RTT samples

reeived.

Oursimpletestbedhasthreenodes. ThedatapathonsistsoftheTCPsender,theemulation

node,and thereeiver, witha singlebottlenekqueueat the emulation node. We arryouttwo

dierent types ofexperiments.

First,weemulateinter-systemhandovers,wherethenewlinkhassigniantlylargerapaity

handovertime,apaityofthelinkfollowingthe emulationnodeishangedto zeroforsome

pe-riod(interruptiontime)toaount forphysial linkoutage. Duringthisperiodnotransmissions

are madebythe emulation node,but end hosts arenot notied ofthe event. In partiular, the

TCP sender is ontinuing to ll up the bottlenek queue with more data pakets (aslong as it

keepsreeivingupstreamACKsalreadyontheirwayfromtheemulation node). After the

inter-ruption time haspassed, new linkharateristis are set, thus resuming transmissionalong the

whole path. Certainly, this modeldoesnot apture every aspetof a handover, but we believe

itmodelsrealityadequatelyfromthe apaity point of view.

Seond,westart anumberofparallel TCPowsalongthesame pathandstopmostofthem

simultaneouslyafter aertaintime. Thisowdynamis reatesaapaityinreaseaspereived

by the remaining ows in the system. As it an be notied, this senario does not assume a

wirelessenvironment,resultsfromthesetypeofexperimentsareharateristitowirednetworks

aswell.

Note that serious apaity hanges an our due to dierent reasons, as well. Wireless

tehnologies thatimplement the shared hannel onept (WLAN, HSPA, 3GPP-LTE, WiMAX

et.) areprimary andidatesforsuhevents. Capaityhangesaremainlyausedbypopulation

hangesinaellbeauseofmobility,andradiolinkqualityhange duetointerfereneandfading

eets. OneouldimagineanextremesituationofauserwithanativeTCPsessionmovingfrom

a denselypopulated ell into anemptyone. In thisase this user'sTCP session willexperiene

a seriousapaityup-swith. Although we donot investigate radiolinkqualityhange anduser

mobility, we believe that results presented below arevalid in ase of sudden apaity inreases

of any kind.

Eetiveness

First, we investigate the time needed to reah full linkutilization by RenoSDand BICSD

om-paredto their respetive baseversions afteraninter-system handover. Link apaitybeforethe

handover was set to 10 Mbit/s and propagation RTT was set to 70 ms. Post-handover, link

apaityis inreasedto 100 Mbit/s,while thepropagationRTT remainsunhanged.

Results an be seen in Figure 4.10. RenoSD shows a 20 seond improvement ompared to

basi NewReno measured when their respetive ongestion window reahes its rst maximum

afterthe handoverevent. Itisalsoimportantto seethatRenoSDreahes50%linkutilizationin

lessthan 2seonds; thisisthepoint wherethe trigger isdeativatedasRenoSDrevertsbakto

normalongestionavoidane. NotethatFigure4.10(a) depitsasingleexperiment,nevertheless

all experiment runs exhibited almost the same harateristis with slight variations due to the

exat state of therespetive TCP owwhen the triggerwasativated.

BICSD also outperforms its ounterpart by approximately 3 seonds in reahing full link

(a)NewRenovs.RenoSD (b)BICvs.BICSD

Figure4.10: Congestion windows during handover

Wehoseto present the worst-asegain: inthisasethe BICowwasaught inaluky state

bythehandover,whenitwaslosetoswithingtoitsmostaggressivewindowupdatemode[83℄.

Detetion time

Next,wemeasurethedetetiontimeofSpeedDetetomparedtoPNCD[88℄inourowdynamis

senario. We usea self-developed Linuximplementation of PNCD (inludingthe orresponding

bandwidth estimators BEand RE) asthe authors'versionis not available. Theapaityof the

bottleneklink is100 Mbit/s,thepropagation RTT is 70ms andthe bottlenek queueis setto

1.2 BDP.

We start

N

^{parallel TCP owsover abottlenek link, andat time}

t 0

^{we stop}

N − K

^ows.

Thisauses aapaityinrease proportionalto

N

K

^{,thatthealgorithmseither detetor don't. If}

theydo,wereordthetimeelapsed sine

t ₀

^{. Speiallyfor} SpeedDetet,themeasuredreation time issimply thetime whenthe triggeris ativated. PNCD on theother hand detetsan

up-swith ifits internal

ncc

^{variable exeedstheurrent}

cwnd

^{,thereforewe monitor thedierene}

of thetwo variables and reordthetimestamp when

ncc > cwnd

^beomestrue.

First, we show the senario where

N = 20

^and

K = 2

^{. As it an be seen in Figure 4.11,}

SpeedDetetreatssigniantlyquikerthanPNCDwiththedierenebeingapproximately370

ms for ow 2 and more than 700 ms for ow 1. Note that SD triggers at about the same time

for both ows.

Thedistributionof detetiontimesisdemonstrated onFigure4.12. Using logsale onthey

axis reveals two regimes for both algorithms. The rst is the range where both of them detet

up-swithes and we all it the reliable range. This range is

K = 1..12

^{in this ase, whih}

orrespondstoapaityinreases70%andup. Belowthislimit, for

K = 13..19

^{,bothalgorithms}

Figure4.11: SD versus PNCD:

N = 20, K = 2

of apaity inrease is lessthan some threshold, and this threshold an be eetively tunedfor

SpeedDetet byhangingsensitivityparameter

p ₁

^.

We observe thatin thereliable regime, reation times of our algorithm arealways lessthan

those of PNCD's. Thedierene inthis partiular setup isaround

100..150

^{ms,andit isdue to}

thefat thatPNCD needs

2 · RT T ≈ 140

^{ms timeto establisha new, aurate estimatefor the}

hanged bandwidth.

Our main observation is thatSD operates onsistently when parallel ows are present (i.e.,

whethertotriggerornot),whilePNCDtypiallytriggersforasingleowrightaftertheup-swith

andmissestotriggerforallotherows. Furthermore,SDdetetssuddenapaityinreasesfaster

than PNCD inawide range ofsettings.

In document B daeUiveiyfTe h (Pldal 120-123)