4.2 Measuring high-speed TCP performane during mobile handovers
4.2.4 Buer overow probability at handovers
solving for
α
,we obtain the following hain inequalities.α =
1 − T 0 T 0 ′ + Q ′
W ∗ >
1 − T 0 T 0 ′ + Q ′
µ ′ T 0 ′ >
1 − T 0
T 0 ′
µ ′ T 0 ′ = (T 0 ′ − T 0 )µ ′
(4.3)Theseinequalitiesinturnyieldalowerbound for
α
usinglinkharateristisonstantsonly.α > (T 0 ′ − T 0 )µ ′
Itmeansthatforasingleowif
α
issetbelowthislevel,theprotoolwillbeunabletoutilizethenewlink. Forexample,FASTinourdown-swithsenariowith1500bytepaketswillneverreah
full link utilization in the equilibrium state for
α < (250
ms− 120
ms) · 10
Mb/s≈ 110
pakets.Thisfat isreeted inFigure4.2(o); for
α = 200
FAST utilizesthefull apaity oftheType-1linkbut for
α = 40
itdoesnot.Measurements summary
We observed that loss-based TCP variants adapt slowly to apaity up-swithes. Although
urrentTCPvariantsdoprobeinreasedapaity(e.g.,duringongestionavoidane inNewReno
or during max inrease in BIC), the time these probes take limit the time resolution of the
detetion. A possible solution ould be to expliitly detet inreased link apaity and initiate
an aggressive window-growing phaseon suh events.
FASTTCP handlesapaityup-swithesimmediately,but mayfailto utilizethelinkaftera
down-swith. ThereasonliesinFAST'swaytoalulatetheamountofpaketsinthebottlenek
buerif
α
is too low for a give down-swith senario, FAST will never utilize the new link.Inreasingthe default value of
α
wouldunneessarilyllthebuer too,therefore webelieve the dynami setting ofα
based on available bandwidth measurements would be the real solution here.S
C
∆
RTT base
link outage
Figure 4.4: A simplied TCP ontrol loop as seen by the sender. Data pakets leaving the
sender S rstgetqueued (andpossiblydropped)atthebottleneklink,thenarriveat thelient
C. There, ACK pakets are triggered that travel bak to the sender. Base propagation delay
(
RT T
base) is aumulated at some point during theyle.
In Figure 4.5 we illustrate the eet of short linkoutages enountered e.g., during wireless
handovers. Assuming that the wireless aess link is the bottlenek of the system we would
see wireless buer oupany growing up to the maximum buer size then dropping to some
non-zerovalueandgrowing tomaximumbuersizeagain, periodially. Whenalinkoutagehits,
link bandwidth past the buer drops to zero for a short timeausing an abrupt peak inbuer
oupany. If at this point there is no free buer spae to aommodate the peak, an overrun
ours.
Themoment whenthebuerusagepeakhits,isindependent oftheTCPsender'sstate,thus
the probability of buer overruns equals the ratio of the buer growth period when a peak of
given size ould not be aommodated. In partiular, if
T p
denotes the buer growth yle'speriod timeand
T o (B p )
isthetimeduringthis periodwhenausagepeakofsizeB p
wouldausean overrun, theprobability ofthe overrun is
P(B p ) = T o (B p )
T p .
(4.4)Wewillalulate
T p
andT o
forasingleAIMDTCP owwiththefollowingassumptionsand notation.•
Thepropagationround-tripdelayT
,themaximumbuersizeB
maxandthelinkbandwidth
µ
pastthequeueareonstants. (Note: TheBDP produtofthelinkisthus notedasµT
.)•
LetW
minand
W
maxdenote the extremes of the ongestion window
W (t)
orresponding to minimumand maximumbuer oupany,respetively,whenthere isnolinkoutageinA
B
C
PSfragreplaements
B(t) B
maxtime
T p T o
Figure 4.5: Buer oupany dynamis withpeaks due to link outages (AIMD ongestion
on-trol). Dottedline indiates the maximumbuer size above whih pakets aredropped fromthe
buer. (Note: Therelativewidthofthepeaks isillustration;inatypialsenarioitisnegligible
ompared to
T p
.)•
LetB
max> µT
andW
max> µT + B
max, otherwise the buer would empty during the
dereasephaseofthe well-knownsawtooth patternof
W (t)
,andTCP ouldnot utilizethefullavailablebandwidth. Thisalsoimpliesthatthebuerisneverempty,thatis
B(t) > 0
for all
t
. As the buer is neverempty,the TCP sender's windowW (t)
grows and shrinksperiodially with
B (t)
.•
We onsider Appropriate ByteCounting to beimplemented or delayed ACK-s turnedo,thus windowgrowth isreally 1paket/
T
.•
As we onsider only a single ow, buer oupany at any timet
is thedierene of theatualongestionwindowofthatowandthedatanotinthebuerat
t
,B(t) = W (t) − µT
.Letthen
W ∗ (B p )
denotetheritial window size thatsolvesB
max= W ∗ (B p ) + B p − µT
thatis at whihwindowsize a buerusage peak
B p
ausesa buer overrun.As
µT
isonstant,T o (B p )
equalsthetimeneededtogrowthewindowfromW ∗ (B p )
toW
max.In general, the time
t(W 1 , W 2 )
to grow the window fromW 1
toW 2
an be alulated as givenin[116℄,provided theTCP isinongestionavoidanephaseandthequeueisnon-empty,thatis
both
W 1
andW 2
are> µT
(formore detailsplease refer to theoriginalpubliation [116℄):t(W 1 , W 2 ) = W 2 2 − W 1 2 2µ .
From here
T o (B p ) = t(W ∗ (B p ), W
max)
with
W
max= B
max+ µT W
min= B
max+ µT
2
W ∗ (B p ) = B
max− B p + µT
beomes
T o (B p ) = W
max2 − W ∗ (B p ) 2
2µ = B p (2(B
max+ µT ) − B p )
2µ .
(4.5)T p
on theother handis the timeittakes togrow thewindowfromW
minto
W
max,thatis
T p = t(W
min, W
max) = W
max2 − W 2
min
2µ =
3
4 (B
max+ µT ) 2
2µ .
(4.6)Substituting (4.5) and(4.6) into (4.4) we obtainthe probability ofbuer overrunsas
P (B p ) = T o (B p )
T p = 4B p (2(B
max+ µT ) − B p )
3(B
max+ µT ) 2 .
(4.7)Weontinue withestimationof thebuer usagepeak
B p
. Suh peaks areausedbythefatthatduringlinkoutagethebuerisstillbeinglleduntileitheroutageendsorthesendersstops
sending, due to lak of returning ACK-s. The latter happens when the ACK triggered by the
last paket served before the outage reahes the sender, that is exatly a base round-trip time
(
T
) after thestart ofthe outage. Clearly,B p
dependsontheduration (width)of thepeak,andtherefore weexamine
B (t)
aroundthe peak inmoredetail.We note theduration of link outageby
τ
,and in Figure 4.6we show thetwo ases ofB(t)
evolution depending on the relation of
τ
andT
. In both ases prior to the outage the buergrows at the dierene of the TCP sending rate
λ
and the queue's servie rateµ
. During theoutage theservierate iszero, and thebuer lls at therate of
λ
. Thisinreased llingperiodlasts for the shorter of
τ
andT
, but the sender does not ease sending until afterT
(the rstnon-reeption ofan ACK). Whenlinkoutageends,thebuerresumesto growwithrate
λ − µ
if pakets are still oming from the sender. Otherwise the buer starts to deplete at rate
µ
.Depletion ends when the sender reeives the rst ACK after the outage, and this duration is
min(τ, T )
.In both ases in Figure 4.6,
B p
is themaximum growth ofB(t)
ompared to the referenease. For thetwo ases thengrowth is
B p =
λτ
ifτ < T ,
λT − (λ − µ)T = µT
ifτ > T .
(4.8)
Letus also assumehere thatthe periodtimeof TCP windowgrowth is
T p ≫ max(τ, T )
PSfragreplaements
B (t)
− µ λ − µ
λ − µ
λ B p
τ τ
T
t
(a)
τ < T
PSfragreplaements
B (t)
− µ
λ − µ
λ B p
τ
T T
t
(b)
τ > T
Figure4.6: Bueroupationgrowtharoundthepeaks. Dashedlinesmarkbuergrowthinase
of nolinkoutages. Non-zerogrowthrates arealso noted.
from whih itfollows that
λ − µ ≪ λ
during the time-frame of an outage. It means thatwe an approximate
λ ≈ µ
in(4.8) , thus weexpress buer oupany growth as
B p ≈ µ min(τ, T ).
(4.9)We show two examples of our ndings. First, we plot the probability
P (B p )
in (4.7) using(4.9) in Figure 4.7 with various interruption times and maximum buer sizes (
µ = 10
Mbit/s,T = 200
ms). It an be seenthat theprobabilityof bueroverruns an be reduedifthebuer is sized appropriately for handover outages. Seondly, in Figure 4.8 we show an exampleB(t)
peak as measured in our emulation testbed. A NewReno TCP in ongestion avoidane phase
was measured with parameters
µ = 10
Mbit/s,B
max= 250
kbyte,T = 200
ms andτ = 50
ms.The emulated handover started at
t ≈ 101.45
s, and as expeted, thebuer grew suddenlyandthe sender stopped due to ACK shortage after
T
. The alulation in(4.9) gives an estimatedbuer growth of
B p = 62.5
kbyte, whih mathesthe testbed results. Using theanalyti modeltheprobabilityof bueroverrun inthisase is
P (B p ) ≈ 0.25
.Theabove ndings were derived for AIMDTCPs, but basedon these we an formulate the
following about non-AIMDvariantsaswell. First,FAST behaves dierently from AIMDTCPs
beause of its buer lling behavior. If there is only a single FAST ow in the equilibrium
state, it would use
α
pakets from the buer,sine in equilibriumthe window isW = µT + α
for FAST (no sawtooth). It means that if
B p > B
max− α
, a buer overrun will our with aprobabilityof 1,otherwise
P (B p ) = 0
. Seondly,other non-AIMDTCP variantslike BIC result0.01 0.1 1
100 50 200 150
300 250
250 500 750 1000 0.01
0.1 1 probability
outage [ms]
buffer size [kbyte]
probability
Figure 4.7: Buer overow probability of NewReno. (Note: Lowest value for
B
max isµT = 2
Mbit= 250
kbyte,belowwhihP (B p ) = 1
.)infundamentallydierent,moreaggressivequeuedynamis. Asaresulttheprobabilityofbuer
overruns isexpetedto belargerthan inase ofAIMDprotools.
Inreasing buer sizes to overome the problem presented in this setion is not feasible for
many reasons, therefore we propose a dierent solution, provided that the bottlenek buer is
in the mobile network. Our basi onept is to temporarily inrease the maximum buer size
until the transient eets of the handover wear o. The buer size inrease an be given and
revoked per user for those that need it, and this way we do not impose large requirements on
total system buer apaity. We propose that on handover the buer is inreased by
B p
as in(4.9) , as all parameters of this expression are known to the mobile system. It isexpeted that
TCPs in steady state will derease their buer usage aording to Figure 4.6, thus the extra
buer spaean be laimedbak. Thiswaythe probability ofbuer overruns ouldberedued
or eliminated.