P. 1
51260

51260

|Views: 4|Likes:
Published by Rajendra Bhosale

More info:

Published by: Rajendra Bhosale on Jul 20, 2012
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

12/18/2013

pdf

text

original

1

Performance Metrics for IEEE 802.21 Media
Independent Handover (MIH) Signaling
David Griffith, Richard Rouil, and Nada Golmie
Abstract—The IEEE 802.21 Media Independent Handover
(MIH) working group is developing a set of mechanisms to
facilitate migration of mobile users between access networks
that use different link-layer technologies. Among these are
mobility managers that create and process signaling messages
to facilitate handovers. The MIH signaling architecture being
developed in the Internet Engineering Task Force (IETF) allows
any transport layer protocol to carry MIH messages. The IETF
has considered using the unreliable but lightweight transport
available with the User Datagram Protocol (UDP) as well as
the reliable stream-oriented transport with congestion control
offered by the Transmission Control Protocol (TCP). In this
paper we develop mathematical models that result in expressions
for the characteristic function of the time required to complete
exchanges of an arbitrary number of MIH signaling messages
between a mobile node (MN) and a remote mobility manager
(MM). Our models also provide expressions for the average
amount of overhead associated with MIH message exchanges
due to retransmissions either by the MIH signaling entities or by
the transport-layer protocol. In addition, we provide simulation
results that confirm the results from the mathematical model and
illustrate the effect of varying transport parameters such as the
TCP maximum retransmission timeout.
Index Terms—Media Independent Handover (MIH), TCP,
UDP, handover latency, signaling overhead
I. INTRODUCTION
Handovers in mobile wireless networks have been thor-
oughly studied for the case where the handover is hori-
zontal (i.e. a mobile node’s movement causes it to end a
connection with one access point and begin another with a
different access point that uses the same Layer 2 technology
as the previous one)[1],[2]. With the proliferation of Layer
2 wireless technologies (e.g. IEEE 802.11 (WiFi) and IEEE
802.16 (WiMAX)), manufacturers of mobile telecommuni-
cations devices have recognized the benefit of developing
wireless mobiles with multiple antennas that are capable of
performing so-called vertical handovers, in which the mobile
switches active connections from a network access point that
uses one Layer 2 technology to another network access point
that uses a different Layer 2 technology.
Various standards bodies such as the IETF and IEEE 802.21
have been developing new protocols to support fast handovers
at different layers. The IETF’s Mobility for IP: Performance,
Signaling and Handoff Optimization (MIPSHOP) and IP
Mobility Optimizations (MOBOPTS) working groups in the
IETF have developed enhanced versions of IPv4 and IPv6
that allow a mobile user to receive packets by maintaining
The authors are with the National Institute of Standards and Technology,
Gaithersburg, MD 20899 USA.
This work was supported in part by the Office of Law Enforcement
Standards (OLES).
information on the user’s current location in the network.
These enhancements include Mobile IPv6 [3], Fast Mobile
IPv6 (FMIPv6) [4], Hierarchical Mobile IPv6 (HMIPv6) [5].
FMIPv6 and HMIPv6 in particular were designed to solve the
problem of excessive handover latency at Layer 3. The key
to reducing latency lies in beginning the handover process
early enough that the mobile node’s connection to the target
access network is fully prepared before the old connection fails
(make before break). This requires giving higher layers access
to link-layer information such as triggers related to received
signal quality. The IEEE 802.21 Media Independent Handover
(MIH) working group is developing mechanisms to use Layer
2 triggers associated with events to provide early warning of
impending handovers and thereby decrease handover latency
[6].
The IEEE 802.21 architecture consists of MIH Users located
above Layer 4 that use MIH Service Access Points (SAPs) to
communicate with MIH services at lower layers. MIH Users
are the initiation and termination points for MIH signaling
sessions. The MIH Function (MIHF), located between Layer
2 and Layer 3, provides handover services (including the
Event Service (ES), Information Service (IS), and Command
Service (CS)) through SAPs that are defined by the IEEE
802.21 working group. The IETF working group identified
low latency and reliable delivery as two key requirements
that will be imposed by some MIH Users [7]. In addition,
congestion control may be required depending on the amount
of data an MIH User generates. Thus some MIH signaling
messages should be carried over a reliable transport protocol
like TCP, which offers reliable delivery and congestion control
at the expense of low latency. Other messages may be better
delivered by a less reliable but lightweight transport protocol
like UDP. The MIPSHOP working group is considering a
number of possible transport solutions, discussed in [8].
The draft [8] allows any transport protocol to carry MIH
signaling messages, and identifies several factors that can
affect the choice of transport protocol that should be consid-
ered when designing an MIH implementation. These include
message size, message generation rate, and the effect of
retransmission on performance (e.g. latency). The draft notes
that short time-sensitive messages such as those associated
with CS and ES should travel over UDP but use retransmis-
sions managed by the MIH User itself, while longer messages
such as the database updates that are part of the IS need TCP’s
congestion control. These recommendations do not include a
quantitative analysis to support them; it is our goal in this
paper to supply a general model that can be used to get
performance metrics for MIH messages over either UDP or
TCP while varying network parameters such as data rate, MIH
2
τ1
τ1 + 2(1 + TMIH(1))
τ1 + 1 + TMIH(1)
τ2
τ2 + Tproc(2)
τ3
Mobile Node (MN) Mobility Manager (MM)
t t
D1
D2
MSG1,1
MSG1,2
MSG1,3
ACK1,2
ACK1,3
MSG2,1
ACK2,1
τ1 + 1
τ2 + Tproc(2) + 2
Fig. 1. Sample signal flow between MN and MM, in which packet loss
occurs.
message size, and the network transit time delay distribution.
In addition, we show results obtained using simulation tools
that we developed that allow network planners to consider
various of deployment scenarios.
The rest of this paper is organized as follows. In Section II
we consider the UDP transport where MIH uses ACKs and
retransmission timers. In Section III, we consider TCP; in the
course of this analysis, we examine the effect of introducing
MIH retransmission timers in addition to TCP’s recovery
mechanisms. We use simulation results to verify our models
in Section IV, and examine some of the design trade-offs
associated with the transport layer protocols. We summarize
our results in Section V.
II. MIH SIGNALING OVER UDP
We first consider the case where the MIH User uses UDP to
carry signaling messages. In this case, MIH ensures reliability
by requesting that each outgoing message be acknowledged
by its recipient. In addition, MIH maintains a retransmission
timer for each sent message. We denote the timer duration
for the ith MIH message by T
MIH
(i). If the ACK for the ith
message is not received by the time the retransmission timer
expires, MIH will send a new copy of the message, up to a
limit of R
i
retransmissions.
A. MIH over UDP Latency
We can derive the distribution of the time required to
complete an exchange of M MIH messages. We assume that
an MIH User that receives an MIH message immediately
generates an MIH ACK followed by an MIH response message
that constitutes the next part of the information exchange,
such as a handover signaling sequence. We show an example
message exchange between a mobile node (MN) and mobility
manager (MM) in Fig. 1. The time between the generation of
the ith ACK and the transmission of the subsequent (i +1)th
MIH message is a deterministic quantity that we denote as
T
proc
(i). This reflects the amount of time used by the receiving
node to perform any tasks associated with the arriving MIH
message. We define τ
i
to be the time when the (i − 1)th
message in the handover message exchange is received. For
the case of the first message in the handover sequence, τ
1
is
the time when the message is generated. The amount of time
required to move the ith message out of the transmission buffer
is
i
= L
i
/B s, where L
i
is the length of the first message in
bits and B is the bandwidth of the channel in bits/s. The ith
message traverses the network and, if it is not lost, it arrives
at its destination after some random delay that we denote as
D
i
.
The node that receives the first MIH message immediately
transmits an MIH ACK and, at time τ
2
= τ
1
+
1
+ D
1
+
T
proc
(i), sends the second MIH message back to the node
that initiated the message exchange. In the figure, we show
the MIH ACKs as requiring no time to transmit. We use this
simplification because the total time to complete the exchange,
H, is not affected by the amount of time spent sending
MIH ACKs, and because the exchange’s state advances only
when each of the two endpoints receives MIH messages.
Each message in the exchange sequence is generated at time
τ
i
= τ
i−1
+
i−1
+D
i−1
+T
proc
(i), and the exchange finishes
when the MIH ACK for the Mth MIH message is received.
The total time required to complete the exchange is
H = τ
M
+
M
+D
M
+D
M+1
−τ
1
= D
M+1
+
M

i=1
_
D
i
+
i
+T
proc
(i)
_
, (1)
where D
M+1
is the time for the Mth MIH message’s ACK
to traverse the network.
We assume that the delays {D
i
}
M+1
i=1
are independent iden-
tically distributed random variables with cumulative distrib-
ution function F
D
(t) and characteristic function Φ
D
(ω) =
_

0
f
D
(t) exp(jωt)dt, where f
D
(t) = dF
D
(t)/dt is the
density of D. In practice, the delays will be correlated, since
the forward and reverse paths pass through the same network
elements; we assume independence to simplify the analysis.
In the absence of packet loss, the characteristic function of the
handover delay, Φ
H
(ω), is
_
Φ
D
(ω)
_
M+1
exp
_

M
i=1
[
i
+
T
proc
(i)]
_
.
To model the case where packet losses occur, let p be the
probability that an MIH message or MIH ACK is lost while
transiting the network. The exchange of a message and ACK is
successful with probability P
s
= (1−p)
2
. For the exchange of
M messages to be successful, each message must be received
successfully. This requires that at least one of R
i
+1 attempts
to transmit the ith message be successful, for i = 1, 2, . . . , M.
The probability that a series of message transmission attempts
does not fail is 1−p
Ri+1
. Thus the probability that a message
exchange fails is
P
fail
= 1 −
M

i=1
(1 −p
Ri+1
). (2)
The product is taken over M messages, not M + 1, because
we assume that the loss of the final MIH ACK message will
not prevent completion of the message exchange. A sender
will also re-transmit a message if the sum of the message’s
transit time and the transit time of its corresponding ACK
is greater than the timeout, T
MIH
(i). We assume that each
3
retransmitted message arrives after its preceding copy, so that
it is not possible for a message retransmission to arrive at the
destination before a retransmission that was sent earlier.
Consider the first message in the exchange, which was first
transmitted at time τ
1
. If the message is lost, as shown in
Fig. 1, the sender will not receive an MIH ACK and its MIH
event timer will expire at time τ
1
+
1
+T
MIH
(1). At that time,
the sender will generate a second copy of the first message.
Each additional message that is lost will cause a timeout at
the sender’s MIH User and a retransmission of the original
message, up to a limit of R
1
+ 1 attempts in total. Fig. 1
shows an MIH ACK being lost during the the first message’s
second transmission attempt. This does not add to the handover
delay because the MM is already responding to the first
MIH message. Lost MIH ACKs also trigger retransmissions.
The probability that k transmission attempts fail to reach the
destination prior to a final, successful attempt is p
k
(1 − p).
Each failure to send the ith message adds an additional
δ
i

i
+T
MIH
(i)
units of time to the total time required to complete the
signaling for the handover. We condition on the event that the
message is successfully delivered during one of its allowed
R
i
+1 transmission attempts; thus the probability that the ith
message transmission is preceded by a delay of k · T
MIH
(i) is
p
k
(1 −p)/(1 −p
Ri+1
) for k = 0, 1, . . . , R
i
.
For each of the M messages in the handover, we assume
that the loss performance is independent of that of all the other
messages. The characteristic function of the additional delay

D
i
associated with timeouts caused by losses of the ith MIH
message is therefore
Φ

D
i
(ω) =
1 −p
1 −p
Ri+1
Ri

k=0
p
k
e
jkωδi
=
1 −p
1 −p
Ri+1
1 −(pe
jωδi
)
Ri+1
1 −pe
jωδi
, (3)
where the D superscript indicates that MIH is in datagram
mode (i.e. using UDP transport).
Since the total exchange time H accounting for dropped
messages is H =

M+1
i=1
D
i
+

M
i=1
_

i
+T
proc
(i) +∆
D
i
_
, the
characteristic function for H is
Φ
H
(ω) = (Φ
D
(ω))
M+1
e
jω(T
P
+T

)
M

i=0
Φ

D
i
(ω), (4)
where T
P
=

M
i=1
T
proc
(i) is the total message processing
time spent by both the connection endpoints, and T

=

M
i=1

i
is the total time spent transmitting the messages that
successfully reach their destinations. If the MIH User allows
the same number of retransmissions for each message, so that
R
i
= R for all i, then Equation (4) becomes
Φ
H
(ω) =
(1 −p)
M
1 −P
fail

D
(ω))
M+1
e
jω(T
P
+T

)
×
M

i=0
(1 −(pe
jωδi
)
R+1
)
(1 −pe
jωδi
)
. (5)
Because the characteristic function of H is the Fourier
transform of the density f
H
(t), we can use it to get any
desired statistic for H. For instance, we may want to know
the probability that the exchange time exceeds some threshold
t, which may be associated with a desired level of Quality
of Service (QoS) for the MN. We compute this probability
by getting F
H
(t) = Pr{H ≤ t}, the cumulative distribution
function of H, and then computing 1−F
H
(t). We can get the
distribution function by inverting the characteristic function
using Equation (3.6) from [9], which is
F
H
(t) =
2
π
_

0
Re[Φ
H
(ω)]
sin(ωt)
ω
dω. (6)
Using the trapezoidal rule, we can approximate this integral
as follows, which is Equation (4.4) from [9]:
F
H
(t) ≈
ht
π
+
2
π
N
h

k=1
Re[Φ
H
(kh)]
sin(kht)
k
, (7)
where h is the step size for the summation on the ω-axis, and
N
h
is the number of points taken in the summation. N
h
must
be large enough to produce an accurate approximation, and h
must be small enough to guarantee accuracy while not being so
small that it results in rounding errors. For the computations
that we performed for this paper we found that h ≈ 5 ×
10
−3
radians/s and N
h
= 2 ×10
4
produced good results.
We consider an example to illustrate our results. In Fig. 2,
we consider the case described in Section 10.4 of [10],
in which a mobile node (MN) signals a remote mobility
manager (MM) to perform a handover from an IEEE 802.11
WLAN network to a cellular network. We plot 1 − F
H
(t),
the probability that the handover time exceeds a given time,
t, for three cases in which the probability of packet loss
and the MIH timeout duration vary. The number of messages
required to complete the handover is M = 3. The number of
retransmissions that are allowed for each message is R = 2.
The one-way transit delay is assumed to be exponential with
an average value of 1 normalized unit of time. The processing
time associated with each message is assumed to be zero. The
MIH timeout values for all three messages are the same for
each case plotted in the figure, and are given as multiples of
µ
D
, the expected value of D, which is unity.
The effect of message retransmissions is most clearly visible
in the black curve with diamonds that is associated with
p = 0.01 and T
MIH
= 15µ
D
. The performance curve has a
staircase shape with plateaus that correspond to the packet loss
probability and whose location on the time axis is determined
by the MIH timeout. Decreasing the timeout duration to
T
MIH
= 5µ
D
produces the behavior shown in the red curve
with squares, which is much steeper and which does not
exhibit the same plateau effect that can be seen in the curve
associated with p = 0.01 and T
MIH
= 15µ
D
. If we decrease
the packet loss probability to p = 10
−4
while keeping the
MIH timeout at T
MIH
= 15µ
D
, we obtain behavior shown
by the curve with blue circles. Some plateauing is visible in
this curve; the first plateau is located at the same position
on the time access as the first plateau the same packet loss
probability and the larger MIH timeout. The value of 1−F
H
(t)
at this plateau is two orders of magnitude below the plateau
4
0 5 10 15 20 25 30 35 40
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
t, time units
P
r
{
H

>

t

}


p = 10
−2
, T = 15µ
D
p = 10
−2
, T = 5µ
D
p = 10
−4
, T = 15µ
D
Fig. 2. Probability that the handover time, H, exceeds t seconds for three
sets of values of packet loss probability p and MIH timeout, T.
associated with the packet loss probability value of 0.01.
The three curves in Fig. 2 demonstrate that the performance
of MIH over UDP does not depend on the values of the
retransmission timeout if the timeout period is on the order of
the round-trip time between the MIH connection endpoints. A
small packet loss probability will also eliminate the effect of
retransmissions.
We can get the first and second order moments of H from
Φ
H
(ω). The expected value of H is
µ
H
=
1
j

H
(ω)

¸
¸
¸
¸
ω=0
, (8)
which we can evaluate in the case of Equation (4) as
µ
H
= (M + 1)µ
D
+T
P
+T

+
M

i=1
µ

D
i
, (9)
where µ
D
is the expected one-way transit time, and µ

D
i
is
µ

D
i
=
1
j


D
i
(ω)

¸
¸
¸
¸
ω=0
=
_
(1 −p)R
i
+ 1
1 −p

R
i
+ 1
1 −p
Ri+1
_
δ
i
, (10)
where lim
p→1
µ

D
i
= R
i
δ
i
/2.
Recall that the transit times {D
i
}
M
i=1
are mutually indepen-
dent, and are independent of the additional delays {∆
D
i
}
M
i=1
,
which are themselves mutually independent. The processing
times and packet transmit times are deterministic. Thus the
variance of H is
σ
2
H
= (M + 1)σ
2
D
+
M

i=1
σ
2

D
i
, (11)
10
−4
10
−3
10
−2
10
−1
10
0
10
−2
10
−1
10
0
10
1
10
2
Packet loss probability, p
T
i
m
e
,

s
e
c


µ
H
, T
MIH
= 10µ
D
µ
H
, T
MIH
= µ
D
σ
H
2
, T
MIH
= 10µ
D
σ
H
2
, T
MIH
= µ
D
Fig. 3. Mean and variance of the handover time H versus packet loss
probability, p, for short and long MIH timeout values (T
MIH
= µ
D
and
T
MIH
= 10µ
D
, respectively).
Each of the variances of {∆
D
i
}
M
i=1
are:
σ
2

D
i
= E{(∆
D
i
)
2
} −µ
2

D
i
= (−1)
d
2
Φ

D
i
(ω)

2
¸
¸
¸
¸
ω=0
−µ
2

D
i
=
_
p
(1 −p)
2

(R
i
+ 1)
2
p
Ri+1
(1 −p
Ri+1
)
2
_
δ
2
i
, (12)
where lim
p→1
σ
2

D
i
= (R
i
+ 2)R
i
δ
2
i
/12.
We can use Eq. (9)-(12) to plot the mean and variance of H
for MIH signaling over UDP versus the p, the probability of
message loss, in Fig. 3. For the plots in the figure, we used the
network parameters given in Table I, but we let µ
D
= 100 ms.
We let T
MIH
take values of µ
D
and 10µ
D
to show the effect of
varying the MIH timeout. For this set of network parameters,
the value of T
MIH
does not have an impact on the statistics of
H for p < 10
−3
, and the effect on µ
H
is not noticeable for
p < 10
−2
. Thus a proactive approach to handovers, including
initiating the handover before the message loss probability
becomes too large, allows consistent handover performance
over a wide range of MIH timeout values.
B. MIH over UDP Overhead
As indicated by Fig. 2, setting the timeout interval to be
on the order of the network delay will minimize the effect of
packet loss on the handover delay. Indeed, as T
MIH
(i) → 0,
the additional time required to complete a handover vanishes.
However, setting the timeout interval to be too small will result
in unnecessary generation of duplicate copies of messages,
adding to the overhead associated with the handover. We can
quantify the overhead penalty associated with a particular
set of values for the timeouts {T
MIH
(i)}
M
i=1
by computing
the expected number of messages generated per attempted
exchange.
Assuming that there are retransmission opportunities re-
maining, the sending node will transmit an additional copy of
a message if the previous copy is lost. If the previous copy is
5
not lost, a retransmission will take place if the corresponding
MIH ACK is lost. If neither the message nor its ACK are
lost, a retransmission will occur if the time from the message
transmission to the ACK reception is greater than T
MIH
(i).
Thus the probability that a given message transmission attempt
results in another copy of the message being sent is
P
rt
(i) = p + (1 −p)[p + (1 −p)(1 −F
RTT
(T
MIH
(i)))]
= (2 −p)p + (1 −p)
2
[1 −F
RTT
(T
MIH
(i))]
= 1 −(1 −p)
2
F
RTT
(T
MIH
(i)), (13)
where F
RTT
(t) is the cumulative distribution of the message
round-trip time, RTT. For ease of computation, we assume
that the RTT is the sum of two independent one-way transit
times that each have the cumulative distribution F
D
(t). Thus
F
RTT
(t) =
_

0
f
D
(u)F
D
(t −u)du where f
D
(t) is the density
of D. Equivalently, the characteristic function of RTT is
Φ
RTT
(ω) = (Φ
D
(ω))
2
. For example, if D is exponential with
mean µ
D
, the cumulative distribution of RTT is
F
RTT
(t) =
_
0, t < 0
1 −(1 +t/µ
D
)e
−t/µ
D
, t ≥ 0.
The probability that the sending node generates m copies
of the ith message, given that R
i
retransmissions are allowed,
is
π
m
=
_
(1 −P
rt
(i))P
m−1
rt
(i), m = 1, 2, . . . , R
i
P
Ri
rt
(i), m = R
i
+ 1.
(14)
From this expression we can get the expected number of
messages sent,
n
i
=
Ri+1

m=1

m
= (1 −P
rt
(i))
Ri

m=1
mP
m−1
rt
(i) + (R
i
+ 1)P
Ri
rt
(i)
=
1 −P
Ri+1
rt
(i)
1 −P
rt
(i)
, (15)
and the average total number of messages required for the
exchange is
N
MSG
=
M

i=1
n
i
=
M

i=1
1 −P
Ri+1
rt
(i)
1 −P
rt
(i)
. (16)
This expression does not account for the MIH ACKs that are
generated each time a node receives an MIH message. We can
count those as well; an ACK is generated by a node each time
its MIH User receives an MIH message. For the ith message
in a M-message sequence, the sending node will produce m
copies with probability π
m
as defined in Equation (14). The
probability that any one of these m copies is lost is p; the
probability that k copies out of m are successfully received
by the destination node is
m
C
k
p
m−k
(1 − p)
k
. Therefore
the expected number of MIH ACKs that are generated in
connection with the ith message is
a
i
=
Ri+1

m=1
π
m
m

k=0
_
m
k
_
kp
m−k
(1 −p)
k
= (1 −p)
Ri+1

m=1

m
= (1 −p)n
i
. (17)
To get the expected number of messages of both types sent
during the exchange, we sum a
i
over all messages {i} and
combine the resulting expression with the expected number of
MIH messages from Equation (16), giving
N = (2 −p)
M

i=1
1 −P
Ri+1
rt
(i)
1 −P
rt
(i)
. (18)
From Equation (15), n
i
→ R
i
+ 1 as P
rt
(i) → 1, which
happens when p → 1 or when T
MIH
(i) → 0. We ignore the
first case because the probability of the exchange failure is
1. In the second case, N → 2

M
i=1
(R
i
+ 1), which is the
maximum possible number of messages that can be generated;
R
i
+1 copies of the ith message are generated for all i because
the timer immediately expires after each message transmission,
and each copy results in an MIH ACK’s being sent back.
We consider the example from Section II-A in which M = 3
and R = 2, and µ
D
= 1, so that all times are normalized to
the average transit time. We plot N for three values of the
packet error probability, p: 0.5, 0.1, and 10
−6
in Fig. 4. The
resulting curves show that the average number of messages
per exchange is insensitive to changes in p over a wide range
of values for the MIH timeout, particularly when the timeout
is on the order of the average packet transit time across the
network. As the packet loss probability increases to near unity,
we see a significant deviation in N; if we let p →1 we would
obtain a horizontal line corresponding to N = M · R = 9.
The curves in Fig. 4 suggest that we can prevent the number
of messages generated per exchange from being excessively
large if we let the MIH timeout be at least two times larger
than the measured packet transit time.
So far, our treatment of the message exchange overhead has
considered only those cases where the exchange completes
successfully. In these cases, at least one of the R
i
+ 1
transmission opportunities succeeds at each of the M stages of
the exchange. In general, however, exchanges fail if all of the
allowed number of transmission attempts at one of the stages
result in the MIH message’s not reaching its destination. In
order to assess the load that the MIH Users collectively offer
to the network, we must consider the effect of transmissions
associated with incomplete exchanges. The probability that all
copies of the ith MIH message are lost is p
Ri+1
. Thus the
probability that the mth message is the last message in a given
exchange is
χ
m
=
_
_
_
p
R1+1
, m = 1
p
Rm+1

m−1
i=1
(1 −p
Ri+1
), 1 < m < M

M−1
i=1
(1 −p
Ri+1
), m = M
The expected number of MIH messages that are generated
during an exchange, accounting for the loss of all R
i
+1 copies
6
0 2 4 6 8 10
0
2
4
6
8
10
12
14
16
18
T
MIH
, multiples of mean transit time
M
e
a
n

n
u
m
b
e
r

o
f

m
e
s
s
a
g
e
s

p
e
r

h
a
n
d
o
v
e
r
p = 5 × 10
−1
p = 10
−1
p = 10
−6
Fig. 4. Mean number of messages generated by MIH signaling endpoints
during an exchange where M = 3 and R
i
= 2 for i = 1, 2, 3.
of the ith message at any point in the exchange, is
N
MSG
=
M

m=1
χ
m
m

i=1
n
i
.
We can obtain a simplified version of this expression if we as-
sume that R
i
= R and T
MIH
(i) = T
MIH
for i ∈ {1, 2, . . . , M}.
Then, P
rt
(i) = P
rt
for i ∈ {1, 2, . . . , M} as well. Defining
q 1 −p
R+1
and n
1−P
R+1
rt
1−Prt
, we get
N
MSG
= (1 −q)n
M−1

m=1
mq
m−1
+Mq
M−1
n
=
(1 −q
M
)n
1 −q
=
(1 −P
R+1
rt
)(1 −(1 −p
R+1
)
M
)
(1 −P
rt
)p
R+1
. (19)
Since a
i
= (1 − p)n
i
, the total number of MIH messages in
the exchange is N = (2 −p)N
MSG
.
III. MIH SIGNALING OVER TCP
MIH provides reliable transport over UDP by using internal
timers to trigger retransmission of lost or excessively delayed
messages. We consider the case of MIH over TCP, which
is a connection-based, stream-oriented protocol that provides
reliable transport of application data. Here, MIH allows TCP to
be solely responsible for guaranteeing message reliability and
does not use its own timeouts. We compute the performance
metrics associated with a message exchange between MIH
Users by using the techniques from Section II.
There has been voluminous work on TCP behavior. In
particular, [11] and the work that builds on it, [12], have
respectively characterized the loss and delay behavior of
TCP during a steady-state bulk data transfer and during the
transmission of an arbitrarily small amount of data. A TCP
model for short-lived flows, such as those associated with MIH
signaling, was developed in [13] but does not incorporate a
maximum retransmission timeout (RTO). Our latency analysis
considers the degenerate case in the model [12] in which the
number of segments to be transferred at any time is one. In
addition, we assume that the SYN/SYN-ACK exchange has
already taken place during the start up of the MIH User.
For this discussion, we assume for the sake of tractability
in the analysis that the size of each MIH message is equal to
the size of the data payload of a TCP segment. Thus, each
segment contains exactly one MIH message, which will be
sent as soon as it is put into TCP’s transmit buffer. The time
to send each message is therefore s for all i ∈ {1, 2, . . . , M}.
In practice, an MIH message may not fit exactly within a TCP
segment. This mismatch will present problems with regard to
completing a message exchange in a timely manner. If the
MIH message is smaller than a TCP segment, TCP will not
send the segment until enough bits have been put into the
transmission buffer by the MIH User. This is because TCP is
a stream-oriented protocol rather than a packet-oriented one.
If MIH timeouts, described in Section II, are in use, the MIH
message may not get sent until enough duplicate copies are
put into the TCP transmit buffer to fill a segment and trigger
its transmission by TCP. If MIH is not using timeouts, the
MIH message may not get sent at all. Conversely, if the MIH
message is larger than a segment, the first portion of the MIH
message, up to an integer multiple of a segment length, will
get sent while the remainder of the message will remain in
the TCP’s buffer until enough additional bits are generated by
the MIH User to cause the segment to be transmitted.
A. MIH over TCP Latency
In the absence of MIH timeouts, TCP will automatically
retransmit messages if its own retransmission timer expires
or if it receives three duplicate acknowledgments from the
destination node at the other endpoint of the connection. We
assume that there are no unacknowledged segments when
the exchange begins. Because the TCP transmission window
will be at least one segment long, the node initiating the
message exchange will be able to transmit the first message
immediately. If there is no packet loss, the destination node
will generate an acknowledgment for the source node as soon
as its TCP layer receives the source’s segment. When the
source node receives the acknowledgment, it will advance its
own transmission window and possibly expand its congestion
window, depending on whether it is in Slow Start mode or
Congestion Avoidance mode.
Because we assume that the source node does not have
any outstanding unacknowledged segments when it begins the
exchange, it will not receive any duplicate acknowledgments
from the destination if the first message is lost. Therefore,
a source node will not respond to a packet loss by going
into Fast Retransmit/Fast Recovery mode; instead, it will wait
until its retransmission timer expires. When this happens,
the source node will shrink its congestion window to one
segment, retransmit the message, and double the RTO value.
If subsequent retransmitted messages are also lost, the source
node will continue to double its RTO value until it reaches
a maximum, RTO
max
= 2
K
RTO, where K is the maximum
7
τi
t
τi + + TRTO
t
τi + 2 + 3TRTO
τi + 3 + 7TRTO
τi+1
τi+1 + Tproc(i + 1)
τi+2
MSGi,1
MSGi,2
MSGi,3
Mobile Node (MN) Mobility Manager (MM)
ACKi,2
MSGi+1,1 + ACK
ACKi+1,1
Di
Fig. 5. Sample MIH signal flow between MN and MM when TCP transport
is used and when some messages and TCP ACKs are lost in transit.
number of times that TCP doubles the RTO value. We assume
that TCP will continue to send copies of the lost segment until
it receives an acknowledgment from the destination endpoint.
We show an example of an MIH message transmission
over TCP in Fig. 5. In the figure, the ith message in the
message sequence, which is transmitted at time τ
i
, is lost in
transit. The lack of a TCP ACK causes the sending node’s
TCP to retransmit the segment and double the RTO. The
first retransmission attempt succeeds in the sense that the
MM receives the message and begins preparing a response
message. However the TCP ACK is lost on the way back
to the MN, so the MN’s TCP sends a second copy of the
segment at time τ
i
+2 +3T
RTO
and doubles the RTO interval
again. This segment is also lost, and the MN’s TCP will send
another copy if no ACK arrives by the time τ
i
+ 3 + 7T
RTO
.
Because the destination node receives the first transmission of
the MIH message at time τ
i+1
= τ
i
+ + D
i
, it generates
a response MIH message at time τ
i+1
+ T
proc
(i + 1). This
message is contained within a TCP segment whose header’s
acknowledgment field contains the byte number that follows
the last byte of the ith message. When the MN receives this
message at time τ
i+2
, its TCP layer notes the acknowledgment
and stops retransmission of the ith message.
Our latency analysis is similar to our treatment of MIH
over UDP from Section II. The principal differences lie in
the exponential expansion of TCP’s timeout window, up to
a maximum amount, with each successive timeout and the
lack of a limit on the number of retransmission attempts.
Note that the use of TCP does not affect the time penalty
associated with each successful transit, nor does it affect the
processing times for each message at the MIH layer. Thus the
only change is a new expression for the additional time penalty
imposed on the ith message in the message sequence. Because
the RTO doubles with each retransmission, the time when the
rth retransmission occurs (given the first transmission attempt
occurred at time τ
i
) is τ
i
+r+T
RTO
+2T
RTO
+· · ·+2
r−1
T
RTO
=
τ
i
+ r + (2
r
− 1)T
RTO
for r = 1, 2, . . . , K. Immediately
after the rth retransmission attempt, TCP expands the RTO to
0 5 10 15 20 25 30 35 40
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
t, time units
P
r
{
H

>

t

}
p = 10
−4
, R = 2, K = 2
p = 10
−4
, R = 2, K = 6
p = 10
−2
, R = 2, K = 6
p = 10
−2
, R = 4, K = 6
p = 10
−4
, R = 4, K = 6
p = 10
−4
, R = 1, K = 6
Fig. 6. Probability that the MIH message exchange time, H, exceeds t
seconds for six sets of values of packet loss probability p, RTO, and number
of RTO stages, K.
2
r
T
RTO
. At the Kth retransmission attempt, the retransmission
occurs at time τ
i
+ K + 2
K
T
RTO
− T
RTO
= τ
i
+ K +
RTO
max
− T
RTO
, after which the RTO expands to RTO
max
.
For subsequent retransmissions (r > K), the RTO remains
at its maximum value and the rth retransmission occurs at
time τ
i
+r +(r −K +1)RTO
max
−T
RTO
. The characteristic
function of the extra delay ∆
C
i
caused by packet loss when
MIH is using TCP transport (i.e. in connection mode) is thus
Φ

C
i
(ω)
= (1 −p)
K

r=0
p
r
e
jω[r+(2
r
−1)TRTO]
+ (1 −p)

r=K+1
p
r
e
jω[r+(r−K+1)RTOmax−TRTO]
= (1 −p)
K

r=0
p
r
e
jω[r+(2
r
−1)TRTO]
+
(1 −p)p
K+1
e
jω[(K+1)+2RTOmax−TRTO]
1 −pe
jω(+RTOmax)
. (20)
Inserting this expression into Equation (4) in place of Φ

D
i
(ω)
gives us the characteristic function for H. By inverting Φ
H
(ω)
to obtain F
H
(t) using Equation (6), we can again get Pr{H >
t} = 1 −F
H
(t) for various values of t. We plot performance
curves for the case where M = 3 in Fig. 6, which shows that,
similar to the UDP case, latency is minimized by using small
timeouts and fewer backoff stages. In addition, we observe the
same plateau effects that we saw in the UDP case; they are
most noticeable when the packet loss rate is high.
We can use the characteristic function to obtain the expected
value and the variance of the exchange time, H. The expres-
sion for µ
H
is nearly identical to Equation (9); the only change
is replacing µ

D
i
with the expected value of ∆
C
i
, which we
obtain from Equation (20) as follows:
µ

C
i
=
p
1 −p
+
(1 −p −2
K
p
K+1
)p
(1 −p)(1 −2p)
T
RTO
. (21)
8
The second fraction assumes the indeterminate form 0/0 when
p = 1/2. We can use L’Hˆ opital’s Rule to get
lim
p→1/2
µ

C
i
= +
K+ 2
2
T
RTO
.
Equation (21) is closely related to Equation (18) in [12], which
itself comes from [11]. We can manipulate our Equation (21)
to produce Equation (18), with an additional factor of p,
by letting → 0 and observing that the general form of
Equation (19) from [12], G(p) = 1 +

K
k=1
2
k−1
p
k
, can be
written as
G(p) =
1 −p −2
K
p
K+1
1 −2p
.
The cause of the additional factor of p is that Equation (18)
gives the average duration of Z
TO
, the TCP timeout period,
given that a timeout occurred; thus Z
TO
must be at least T
RTO
time units long. Our Equation (21) gives the average time to
transmit a segment in addition to the network transit delay; the
occurrence of a timeout is not a given condition, and so ∆
C
i
has a minimum value of zero. We can also relate Equation (21)
to the expression for C
1
1
(the mean time to send one segment
with an initial congestion window size of one segment) in [13]
by letting K →∞, since their model assumes no upper bound
on the RTO value. The expression in [13] has an additional
factor of p as well, because we are computing the additional
delay due to message losses.
The variance of H when MIH messages use TCP is given
by Eq. (11) with σ
2

C
i
replacing σ
2

D
i
. We get the variance of

C
i
in the same way that we got Equation (12), which gives
us a more complex but still closed-form expression:
σ
2

C
i
= 2
(1−p)
2
−(2p)
K
p(2+K−3(K+1)p+2Kp
2
)
(1−p)
2
(1−2p)
2
pT
RTO
+
_
1−p
(1−2p)
2
(1−4p)
+
(2p)
K+1
(1−6p−(2
K+4
−9)p
2
−4p
3
)
(1−p)
2
(1−2p)
2
(1−4p)

(4p)
K
p(7−29p−4p
3
+p
K+2
−4p
K+3
)
(1−p)
2
(1−2p)
2
(1−4p)
_
pT
2
RTO
+
p
2
(1−p)
2
.
(22)
This quantity becomes undefined when p = 1/2 or p = 1/4.
As in the case of µ

C
i
, we can take limits and get
lim
p→1/4
σ
2

C
i
=
4
9

2
+
2(−5 + 9 · 2
K
) −3K
9 · 2
K
T
RTO
+
4
K
(27K −1) + 9 · 2
K+1
−1
9 · 4
K+1
T
2
RTO
(23)
and
lim
p→1/2
σ
2

C
i
= 2
2
+
8 + 5K +K
2
2
T
RTO
+
2(−9 + 13 · 2
K
) −8K −K
2
4
T
2
RTO
. (24)
We show plots of µ
H
and σ
2
H
in Fig. 7, using the same
parameters that we used to generate Fig. 3. Here we consider
two values of the TCP RTO value, µ
D
and 10µ
D
. Note that
the performance of UDP and TCP is identical for small values
of p, and that the value of p for which there is a noticeable
increase in the value of µ
H
and σ
2
H
is around the same for
both the UDP and TCP transport cases. When TCP is used,
10
−4
10
−3
10
−2
10
−1
10
0
10
−2
10
−1
10
0
10
1
10
2
Packet loss probability, p
T
i
m
e
,

s
e
c


µ
H
, T
RTO
= 10µ
D
µ
H
, T
RTO
= µ
D
σ
H
2
, T
RTO
= 10µ
D
σ
H
2
, T
RTO
= µ
D
Fig. 7. Mean and variance of the handover time H versus packet loss
probability, p, for short and long TCP RTO values (T
RTO
= µ
D
and T
RTO
=
10µ
D
, respectively).
the mean and variance of H increase without bound as p →1,
because TCP attempts to deliver an unacknowledged segment
until it receives an ACK. The gain from doing this is the
eventual success of the transmission as long as p = 1, but
this benefit is less meaningful in the case of signaling for
handovers, because the amount of time available to complete
message delivery is short. Again, this figure demonstrates the
importance of beginning critical handover signaling exchanges
early, so that degrading link conditions do not expand the
amount of time required to successfully deliver signaling
messages.
B. MIH over TCP overhead
We also can compute the expected number of messages
generated during a message exchange as a function of the
RTO value T
RTO
and the packet loss probability, p. Similar to
the UDP case, retransmission of an MIH message occurs if
the message is lost, its TCP ACK is lost or the RTO timer
expires before the sender receives the ACK. However, we
have to account for the exponential dilation of the timeout
interval until it reaches its final value, RTO
max
. Thus, similar
to Equation (13), the probability that the kth transmission
attempt fails and requires another attempt is
φ
k
=
_
1 −(1 −p)
2
F
RTT
(2
k−1
T
RTO
), k ≤ K
1 −(1 −p)
2
F
RTT
(RTO
max
), k > K.
(25)
where we assume that T
RTO
is the same for all M messages
in the message sequence. We assume that there is no limit on
the number of times TCP attempts to retransmit a segment.
The probability that m copies of the ith message gets sent
is the probability that the TCP ACK for the mth transmitted
copy of the MIH message arrives within the time limit after
m− 1 failures. Defining psi 1 − (1 − p)
2
F
RTT
(RTO
max
),
9
(i.e. φ
k
= ψ for k > K) we have
π
m
=
_
_
_
1 −φ
1
, m = 1
(1 −φ
m
)

m−1
k=1
φ
k
, 1 < m ≤ K
(1 −ψ)(ψ)
m−K−1

K
k=1
φ
k
, m > K
(26)
for m ∈ Z
+
, the set of positive integers. Using this expression
we can get the mean number of copies sent for a given
message:
n
i
=

m=1

m
= 1 −φ
1
+
K

m=2
m(1 −φ
m
)
m−1

k=1
φ
k
+

m=K+1
m(1 −ψ)(ψ)
m−K−1
K

k=1
φ
k
= 1 −φ
1
+
K

m=2
m(1 −φ
m
)
m−1

k=1
φ
k
+
1 +K(1 −ψ)
1 −ψ
K

k=1
φ
k
. (27)
As was the case with MIH transport over UDP, k out of
m segments generated by a source will arrive at their des-
tination (and cause ACKs to be sent back) with probability
m
C
k
p
m−k
(1−p)
k
. Thus the expected number of ACKs is (1−
p)n
i
, and the total number of messages created in connection
with the ith phase of the exchange is (2 − p)n
i
. Because
the retransmission parameters do not change from message
to message, n
i
is a constant, and the mean total number of
messages sent during an exchange is N = (2 −p)Mn, where
n is given in Equation (27).
In Fig. 8, we plot the average number of messages in
a handover versus RTO, where the RTO is expressed as
multiples of the average one-way network transit time, µ
D
.
The handover exchange is composed of M = 3 MIH messages
where TCP is used for transport and there are no timeouts
and retransmissions at the MIH User. The figure bears a
strong resemblance to Fig. 4, which plots N versus the MIH
timeout for the MIH/UDP case. An important difference is
the asymptote at RTO = 0, which is due to our assumption
that TCP never stops trying to send an un-ACKd segment. In
contrast, MIH is limited to a finite number of retransmission
attempts when it uses UDP. The MIH/TCP case is similar to
the MIH/UDP case in that reducing the timeout to be on the
order of the network round-trip time significantly increases
N in both cases. Thus a conservative RTO estimate reduces
overhead, but this is a TCP function that is not controllable
by the MIH User.
C. The Effect of Combining MIH Timeouts with TCP
In this subsection, we present an analytical model that quan-
tifies some of the negative effects that result from combining
the TCP layer and MIH User retransmission mechanisms. The
MIH User generates an MIH ACK message for each MIH
message that it receives. Our model examines the case of an
0 2 4 6 8 10
0
3
6
9
12
15
18
21
24
27
30
RTO, multiples of mean transit time
M
e
a
n

n
u
m
b
e
r

o
f

m
e
s
s
a
g
e
s

p
e
r

h
a
n
d
o
v
e
r
p = 5 × 10
−1
p = 10
−1
p = 10
−6
Fig. 8. Mean number of messages generated by MIH signaling endpoints
using TCP transport during a handover where M = 3 and K = 6, for various
RTO values.
MIH Indication message transmission from a MN to the MM.
This exchange involves one MIH message and one MIH ACK,
for a total of two messages generated jointly by both nodes’
MIH Users. In addition to the MIH messages, the connection
endpoints generate TCP ACKs every time they receive a TCP
segment.
There are two cases involving the relative values of the MIH
timeout T
MIH
and the TCP timeout, T
RTO
. We consider only the
case where the MIH timeout occurs later than the initial value
of the TCP RTO (T
MIH
≥ T
RTO
). Letting the MIH timeout
occur before the RTO would introduce additional copies of
the message into the TCP queue before TCP has a chance to
time out and retransmit, resulting in unnecessary extra traffic.
In this situation, the MIH User’s timeout occurs after the TCP
RTO, so the MIH layer will not begin generating duplicate
messages until TCP has been in Timeout mode for some
period of time. The number of messages that the MIH User
generates depends on the amount of time that TCP spends
in Timeout retransmitting the original copy of the message.
We first examine the Request message from the MN to the
MM. The MIH layer will produce m additional copies of the
1st message, 0 ≤ m ≤ R
1
, if the time from the first message
transmission attempt until the TCP layer receives an ACK, U
1
,
lies in the half-open interval
_
mT
MIH
(1), (m+1)T
MIH
(1)
_
. The
probability of this event is
Pr{U
1

_
mT
MIH
(1), (m+ 1)T
MIH
(1)
_
}
=
_
(m+1)TMIH(1)
mTMIH(1)
f
U1
(u) du
= F
U1
((m+ 1)T
MIH
(1)) −F
U1
(mT
MIH
(1)). (28)
Because MIH is limited to a finite number of retransmissions,
R
1
, it follows that the probability of exactly R
1
retransmis-
sions is the probability that U
1
is greater than R
1
T
MIH
(1),
which is 1 −F
T
M
(R
1
T
MIH
(1)).
From our analysis in Section III, we know that τ
2
− τ
1
=
D
1
+ ∆
C
1
, so we have already characterized the random
10
time from the first transmission attempt to the segment’s
successful reception at the MM; its characteristic function is
Φ
D
(ω)Φ

C
1
(ω). We next get the characteristic function of U
1
,
the time from the 1st message’s reception at the MM to the
arrival of a TCP ACK at the MN. There are two ways the
MN can get a TCP ACK from the MM. The MN can receive
a TCP ACK in the header of the MIH ACK that the MM’s
MIH User generates when it receives the MN’s MIH message;
this assumes that the MM’s TCP uses delayed ACKs. The
MN can also receive a TCP ACK in the header of the next
MIH message in the handover sequence, as long as the MM’s
transmission window is large enough to send both the MIH
ACK and the second MIH message. When any one of these
packets first arrives at the MN, the local TCP will note that
the transmitted segment was correctly received. If the ACK
is received before T
RTO
units of time have expired, TCP will
continue normal operations. If the MN receives the ACK after
TCP has gone into timeout, TCP will end the timeout and
begin Slow Start.
We compute the characteristic function for ∆
C
1,ACK
, the
additional time required to send the MIH ACK for MSG
1
from
the MM to the MN. There are two possibilities, depending
on whether T
proc
(2) is larger than T
RTO
. If T
proc
(2) ≥ T
RTO
and the MIH ACK gets lost in transit, the MM’s TCP will
time out before the 2nd MIH message is ready to be sent; the
message will go into the transmission queue until the MM’s
TCP receives an ACK. The MM’s TCP will behave exactly
like the MN’s TCP, retransmitting the MIH ACK’s segment
and expanding the RTO at each attempt until it reaches its
maximum value, RTO
max
= 2
K
T
RTO
.
In Fig. 9, we show the exchange of messages between
the MN and MM when MIH reliability mechanisms (ACKs
and retransmissions) are used in addition to TCP, and when
T
proc
(2) > T
RTO
. The 1st MIH message travels from the MN
to the MM and is delayed by D
1
units of time. There is
an additional time penalty of ∆
C
1
units associated with failed
attempts by the MN’s TCP to transmit MSG
1
. Immediately
upon receiving MSG
1
, the MM’s MIH layer sends an MIH
ACK to the MN whose TCP header contains the sequence
number of the last octet in MSG
1
. We assume that all MIH
messages and MIH ACKs have the same transmission duration
of units of time. The delay experienced by the MIH ACK is
D
1,ACK
, which has the same characteristic function, Φ
D
(ω),
as D
1
. Failed transmission attempts by the MM’s TCP cause
an additional delay of ∆
C
1,ACK
. When the MIH ACK finally
arrives at the MN, the MN’s TCP immediately responds with
a 40-byte TCP ACK, which reaches the MM after a delay
of D

1,ACK
time units, where D

1,ACK
also has characteristic
function Φ
D
(ω). Like [12], we assume that TCP ACKs are
never lost in transit. From Fig. (9),
Φ
U1
(ω) =
_
Φ
D
(ω)
_
2
Φ

C
1
(ω)Φ

C
1,ACK
(ω)e
j2ω
. (29)
where Φ

C
1,ACK
(ω) is given by Equation (20).
If T
proc
(2) ≤ T
RTO
, and the MM’s TCP transmission window
is at least two segments long, the MM will send the 2nd MIH
message before the RTO expires, as shown in Fig. 10. In the
figure, we show the case where both the MM’s MIH ACK
τ
1
τ
2
Mobile Node (MN) Mobility Manager (MM)
t t
D
1
D
1,ACK
copies
MSG
1
ACK
TCP ACK

MSG
1
lost
ACKs

MSG
1
lost

C
1

C
1,ACK

D

1,ACK

MSG
1
U
1
Fig. 9. Signal flow for an exchange of an MIH message and MIH ACK
message between MN and MM over TCP, when Tproc(2) > T
RTO
.
τ
1
τ
2
Mobile Node (MN) Mobility Manager (MM)
t t
copies
MSG
1
ACK
TCP ACK

MSG
1
lost
ACKs

MSG
1
add’l lost
MSG
1
U
1
τ
2
+ T
proc
(2)
τ
2
+ T
RTO
MSG
1
ACK
MSG
2
Fig. 10. Signal flow for an exchange of an MIH message and MIH ACK
message between MN and MM over TCP, when Tproc(2) ≤ T
RTO
.
and MSG
2
transmissions fail. If no ACK comes back from
the MN by the time t = τ
2
+ T
RTO
, the MM’s TCP will
enter timeout, shrink its congestion window to one segment,
and begin retransmitting the MIH ACK while the 2nd MIH
message waits in the transmission queue. So the MM will have
one attempt to send the 2nd MIH message before its TCP goes
into timeout. The MN has two chances to receive a segment
containing an ACK for the 1st MIH message and send an ACK
back to the MM. Because of this, the characteristic function
for ∆
C
1,ACK
changes slightly from its form in Equation (20),
as we now show.
In this case, ∆
C
1,ACK
takes the values 0 and T
proc
(2) with
probabilities (1−p) and (1−p)p, respectively. The first retrans-
mission of the MIH ACK happens at time t = τ
2
++T
RTO
if
no ACK is received for either the MIH ACK or the 2nd MIH
message, so ∆
C
1,ACK
= + T
RTO
with probability (1 − p)p
2
.
Continuing this pattern and using the results of our analysis
in Section III, we see that ∆
C
1,ACK
= r + (2
r
−1)T
RTO
with
probability (1−p)p
r+1
, for r = 1, 2, . . . , K. Recalling that the
maximum RTO is RTO
max
= 2
K
T
RTO
, we see that ∆
C
1,ACK
=
11
r +(r −K+1)RTO
max
−T
RTO
with probability (1−p)p
r+1
for r ≥ K. Thus we obtain the following expression for the
characteristic function when T
proc
(2) < T
RTO
:
Φ

C
1,ACK
(ω)
= (1 −p) + (1 −p)pe
jωTproc(2)
+(1 −p)
K

r=1
p
r+1
e
jω[r+(2
r
−1)TRTO]
+ (1 −p)

r=K+1
p
r+1
e
jω[r+(r−K+1)RTOmax−TRTO]
= (1 −p)[1 +pe
jωTproc(2)
]
+(1 −p)
K

r=1
p
r+1
e
jω[r+(2
r
−1)TRTO]
+
(1 −p)p
K+2
e
jω[(K+1)+2RTOmax−TRTO]
1 −pe
jω(+RTOmax)
. (30)
From Fig. (10), we can work out the characteristic function of
U
1
by conditioning on the whether MSG
2
is lost in transit, as
follows:
Φ
U1
(ω) = (1 −p)
_
Φ
D
(ω)
_
2
Φ

C
1
(ω)Φ

C
1,ACK
(ω)e
j2ω
+ p
_
Φ
D
(ω)
_
2
Φ

C
1
(ω)e
jω(2+Tproc(2))
=
_
(1 −p)Φ

C
1,ACK
(ω) +pe
jωTproc(2)
_
×
_
Φ
D
(ω)
_
2
Φ

C
1
(ω)e
j2ω
,
where Φ

C
1,ACK
(ω) is given by Equation (30).
With the characteristic function Φ
U1
(ω) in hand, we can get
an estimate of the MIH overhead by computing the size of the
backlog of copies of the MIH Indication message at the MN
at the time the MN receives a TCP ACK for the MIH Request
message. Using Equation (28) and the inversion formula in
Equation (6), we get β(m) Pr{m messages in backlog}:
β(m) = Pr{mT
MIH
(1) ≤ U
1
< (m+ 1)T
MIH
(1)}
= F
U1
_
(m+ 1)T
MIH
(1)
_
−F
U1
_
mT
MIH
(1)
_
=
2
π
_

0
Re[Φ
U1
(ω)]
_
sin
_
(m+ 1)ωT
MIH
(1)
_
−sin
_
mωT
MIH
(1)
_
_

ω
(31)
for m = 0, 1, 2, . . . , R
1
−1 and
β(R
1
) = Pr{U
1
> R
1
T
MIH
(1)}
= 1 −F
U1
_
R
1
T
MIH
(1)
_
= 1 −
2
π
_

0
Re[Φ
U1
(ω)] sin
_
ωR
1
T
MIH
(1)
_

ω
.
(32)
The expected number of messages in the backlog is B =

R1
m=0
mβ(m). Thus the average number of MIH messages
put into the TCP transmission queue for each MIH message
that the MIH User needs to send is the original message plus
the ones in the backlog, or 1 +B.
In Fig. 11, we we plot the average backlog size versus p,
the probability of MIH message loss for three different values
10
−2
10
−1
10
0
0
0.5
1
1.5
2
2.5
3
p
B


T
MIH
=150 msec, T
RTO
= 200 ms
T
MIH
=200 msec, T
RTO
= 200 ms
T
MIH
=400 msec, T
RTO
= 200 ms
T
MIH
=150 msec, T
RTO
= 2 ms
T
MIH
=200 msec, T
RTO
= 2 ms
T
MIH
=400 msec, T
RTO
= 2 ms
Fig. 11. Plot of B, the average size of the message backlog generated
by MIH retransmissions due to expiration of the MIH timer, versus p, the
probability of message loss for different values of the MIH timeout, T
MIH
.
of the MIH timeout, T
MIH
: 150 ms, 200 ms, and 400 ms.
We used the set of parameters shown in Table I, including
T
RTO
= 200 ms; we also used T
RTO
= 2 ms to show the
effect of increasing the TCP retransmission rate. For all three
values of T
MIH
and for T
RTO
= 200 ms, B is insensitive to
p for p < 0.01. The figure also shows that increasing the
MIH timeout reduces the size of the backlog. Computing B
using different values for T
proc
does not result in a significant
change in the curves. Note that using a smaller value of T
RTO
produce curves that have the same value for small values of
p as the curves where T
RTO
= 200 ms; the small T
RTO
curves
are flatter and begin increasing toward R at a larger value of
p. This is because decreasing T
RTO
= 200 increases the TCP
retransmission rate and it becomes more likely that a TCP
ACK will arrive at the MN before the MIH timeout expires,
thus reducing the backlog, although it will increase the TCP
segment transmission rate.
IV. SIMULATION RESULTS
To further quantify the performance of the MIH User over
different transport layers, we performed a set of simulations
using the ns-2 tool, which we enhanced by extending the tool’s
mobility framework [14]. To generate our results, we used a
simple network topology shown in Fig. 12 that consists of a
mobility manager located at a remote Point of Service (PoS)
connected to an access point via a backbone network that we
have abstracted as a single router and a lossy link. Within
the coverage area of the access point is a single mobile node
that communicates with the access point over a IEEE 802.11
wireless LAN link.
In each scenario that we considered, the MN first connects
to the AP and registers with the PoS. Then, depending on
whether we are simulating Indications or Request/Response
exchanges, the MN generates Indications every 0.5 second
or the PoS generates requests every 0.5 second, respectively.
We take performance measurements by taking traces of the
relevant output parameters and averaging those traces over
12
Fig. 12. Network topology used in simulations of Indication and Re-
quest/Response MIH message exchanges.
4000 seconds of simulation time, between the 5 s and 4005 s
marks. The typical standard deviation of the data in each of
the graphs in this section is on the order of 1% of the mean
data value plotted; thus error bars would not be visible and so
are not plotted.
The parameters that we used in our simulations are shown in
Table I. The simulations examined two types of MIH message
exchanges. The first type consists of a Indication message
sent by the mobile node to the mobility manager. The second
type of exchange involves a Request message generated by
the mobility manager which produces a Response message
from the mobile node. Both types of MIH message exchanges
occur according to a Poisson process with an average exchange
arrival rate of two events per second. Each simulation run
covered a time interval of 6005 seconds. The packet loss rate
on the link connecting the router to the access point was
allowed to vary between 0 and 0.5. Two different transport
layers were used in the simulations: UDP and TCP. The
parameters for both transport layers are given in Table I. We
used a variety of values for the TCP maximum RTO, as shown
in the table, and varied the link loss rate for each maximum
RTO value that we used.
We plot values for the probability that MIH transaction
completes successfully in Fig. 13. We compare the results from
ns-2 with theoretical results that we obtained by computing
1 − P
fail
from Equation (2). In all the scenarios that we con-
sidered, we obtained excellent agreement between the values
predicted by the model and the results that we obtained from
the simulations. The graph further shows that the probability
that a message exchange is unable to complete successfully de-
creases as the number of the packets in the exchange increases,
as we would expect. In addition, using MIH acknowledgments
significantly increased the success rate.
A. MIH Delay
In this subsection, we examine the delay performance of
MIH over both UDP and TCP transport layers. In all of the
TABLE I
SIMULATION PARAMETERS
IEEE 802.11
Data rate (Mb/s) 11Mb/s
Coverage area radius (m) 50
Links
Speed (Mb/s) 100
Delay (s) 0.01
UDP
Max packet size (byte) 1000
Header size (bytes) 8
TCP
Max Segment Size (bytes) 1280
Min RTO (s) 0.2
Max retransmission Unlimited
Queue size Unlimited
Header size (bytes) 20
IP header
IPv6 header (bytes) 40
MIH Function
Transaction timeout (s) none
R
i
2
Tproc(2) (s) 0.2
Simulation configuration
Duration (s) 6005
loss probability, p variable [0, 50%]
RTOmax (s) 0.2, 0.3, 0.5, 0.75, 1
Indications/s 2
Requests/s 2
MIH Packet size (bytes) 200
Fig. 13. Theoretical and simulated values for MIH transaction completion
probability.
figures in this subsection, the sets of parameters that we used
are indicated in the legend appearing in the figure.
In Fig. 14 we plot the average time required to com-
plete the simple Indication transmission as well as the Re-
quest/Response message exchange. The bars in the figure do
not show the variance of the simulation data; rather they
show the value of the standard deviation of the completion
time, i.e. the upper terminus of a bar is located at a value
one standard deviation greater than the corresponding mean.
We examine the delay with MIH acknowledgment messages
and retransmissions (for which T
MIH
= 0.5 s for Indications
and T
MIH
= 0.3 s for Request/Response) and without. In
the figure, there are two sets of horizontal marks that corre-
13
Fig. 14. Theoretical and simulation values of mean MIH handover time for
MIH handovers where M = 1 (Indication) and M = 2 (Request/Response)
over UDP.
spond to the scenarios in which MIH Indications and MIH
Request/Response exchanges take place without the use of
MIH ACKs. In both cases, the message exchange will fail
unless the initial attempt to transmit each MIH message is
successful. Thus, the only variance in the exchange completion
time comes from random delays in the network which, in this
case, are very small. We have an average delay of 20 ms and
240 ms for the Indication and Request/Response exchanges,
respectively. The delay curves also show strong agreement
between the simulation results and the mathematical model’s
prediction from Equation (9) and Equation (10), especially for
small values of p.
In Fig. 15 we show simulation results showing the mean
message exchange time for Indications when TCP is used in
addition to MIH ACKs and timeouts. The MIH timeout T
MIH
is set to 3RTO
max
. The figure shows much greater sensitivity
to the packet loss probability as the transaction delay increases
exponentially with respect to the value of T
RTO
. In addition,
using MIH ACKs over TCP further increases the average
delay. When MIH ACKs are used and the average TCP round-
trip delay is greater than the MIH retransmission timeout
interval, MIH will create duplicate packets that go into the
TCP queue. These duplicate packets cause additional delays
because all of them must be transmitted and acknowledged
before the next MIH Indication message can be sent.
We also show simulation results showing the mean transac-
tion completion time for Request/Response exchanges when
TCP is used in addition to MIH ACKs and timeouts in Fig. 16.
We observe that the exchange completion time decreases when
the packet loss reaches 45% for larger values of RTO
max
. This
is because the delays at this level of packet loss are so large and
most Request/Response transactions do not complete within
the ACK timeout interval. Thus the delay shown in the figure
is associated only with the minority of successful exchanges.
To further show the negative effect of combining MIH
retransmissions with TCP’s reliable delivery, we plot the
average Indication completion time for the case where we use
Fig. 15. Mean MIH Indication transaction completion time over TCP
transport with MIH timeouts and acknowledgments.
Fig. 16. Mean MIH Request/Response transaction completion time over TCP
transport with MIH timeouts and acknowledgments.
TCP transport without MIH reliability features in Fig. 17. If
no MIH ACK is used, the MIH User sends a Request and
waits for a Response. In our scenario, the Responses received
are always considered valid although that might not be the
case in real implementations. The figure does not exhibit the
type of severe performance degradation that we observed in
Fig. 15, largely because the TCP transmission queue is no
longer being filled with extra copies of messages that were
lost during transmission attempts. If we compare the average
delay to what we observed in connection with UDP, we see
that the average completion time is much higher in the case of
TCP. However, using TCP results in a much lower probability
that a given exchange will fail.
In Fig. 18 we show the average Request/Response comple-
tion time when MIH retransmission are not being used over
TCP. As we expect, the transaction delay for a given maximum
RTO value is approximately 2.25 times greater than it is in the
case when only a Indication message is being transmitted, as
14
Fig. 17. Mean MIH Indication transaction completion time over TCP
transport without MIH timeouts and acknowledgments.
Fig. 18. Mean MIH Request/Response transaction completion time over TCP
transport without MIH timeouts and acknowledgments.
shown in Fig. 17. Note that in all four figures we do not begin
to see serious increases in the delay for either type of message
exchange until the packet loss probability is on the order of
0.1. Also, the large average delay values associated with packet
loss rates of near 0.5 are an indication that a greater number of
transactions are able to successfully complete within the time
limit than we observed when we used MIH reliability features
in conjunction with TCP.
B. MIH Overhead
In this subsection, we examine the amount of overhead
associated with the various transport layer options. In Fig. 19
we plot theoretical and simulation results for transport over
UDP for both the Indication and Request/Response exchanges.
We observed very close agreement between the theoretical
and simulation results for the Indication exchange, and a
maximum error of 7% for the Request/Response exchange
when p = 0.25. If MIH reliability is not in use, there is no
Fig. 19. Theoretical and simulation values of MIH overhead for Indication
and Request/Response exchanges over UDP.
overhead penalty although this will result in a lower success
rate for both types of exchanges. With MIH acknowledgments
and retransmissions, we observed a slightly higher overhead
penalty in connection with Indication exchanges relative to
the Request/Response case. This follows from Equation (19),
which shows that increasing M, the number of messages in
an exchange, decreases N
MSG
, with N
MSG
→n/(1−q) as M
becomes large.
Next we examine overhead associated with using TCP
transport. In Fig. 20 and Fig. 21, we plot the expected
number of MIH packets sent to the transport layer per MIH
message generated by the MIH User for Indications and
Request/Response exchanges, respectively. If retransmissions
by the MIH layer are turned off, this number is 1. When
there is no packet loss (p = 0), an MIH ACK message
gets sent for each indication/request/response that a node
receives. Therefore the minimum number of MIH packets
sent per message is two if MIH ACKs are being used. As
the packet loss rate increases, the MIH User will retransmit
messages up to two times. The maximum number of packets
for each MIH Indication is therefore 6 (3 copies of the message
and 3 corresponding MIH ACKs), as seen in Fig. 20. In
the case of Request/Response exchanges, if the TCP delay
is too large, Requests will arrive late at the MN, whose
Response will be ignored by the PoS, which will in turn create
another Request. The maximum overhead is 4.5 since there
are at most 3 Requests, 3 ACKs, and 3 Responses for each
Request/Response pair. If we examine both graphs, we observe
that the amount of overhead increases as we decrease the
maximum RTO, which is what we would expect. Furthermore,
significant increases in overhead to not occur until the packet
loss probability exceeds 0.1.
V. CONCLUSIONS
In this paper, we developed characteristic functions that
let us obtain the distribution and moments of the time to
exchange an arbitrary number of MIH messages using either
UDP with MIH retransmissions and acknowledgments or using
15
Fig. 20. Simulation values of MIH overhead for Indication messages over
TCP with MIH ACKs.
Fig. 21. Simulation values of MIH overhead for Request/Response exchanges
over TCP with MIH ACKs.
TCP. We also developed expressions for the average overhead
expressed as extra message transmissions due to packet losses.
Our results verify that there is a tradeoff between latency and
overhead. Adjusting timeout values to reduce latency, whether
in the MIH User if UDP is being used or at the transport
layer if TCP is being used, invariably results in an increase in
the expected number of messages that are transmitted during
an exchange. The delay and overhead penalties are of course
smaller if the packet loss probability is small. This illustrates
the importance of proactive handover signaling that begins
well before the signal quality on the wireless link degrades
to the point where the packet loss probability is on the order
of 0.1 or greater. We also note that TCP is best suited for
longer MIH message exchanges since it will hold data until
there is enough to fill a segment, which is undesirable for
messages that require low latency.
Our simulation results agreed with our mathematical model.
We observed low delays and less overhead in the case where
we used UDP, but we also experienced a higher rate of
failure for a MIH message exchange. TCP displayed a lower
exchange failure rate than UDP with retransmissions, largely
because the number of retransmissions over UDP was limited
while the number of TCP retransmissions was not. TCP’s
shortcoming was that it resulted in greater overhead than UDP
and performed inefficiently when the MIH message generation
rate was low. Our results show that using MIH retransmission
timers with TCP produces conflicts that severely degrade
performance.
ACKNOWLEDGMENT
The authors would like to thank the members of the IETF
MIPSHOP MIH design team for many helpful discussions,
especially Juan Carlos Z´ u˜ niga and Telemaco Melia.
REFERENCES
[1] D. Hong, Daehyoung and S.S. Rappaport, “Traffic Model and Perfor-
mance Analysis for Cellular Mobile Radio Telephone Systems with
Prioritized and Nonprioritized Handoff Procedures,” IEEE Transactions
on Vehicular Technology, Vol. 35, no. 3, pp. 77-92, Aug 1986.
[2] M. Sidi and D. Starobinski, “New Call Blocking versus Handoff Blocking
in Cellular Networks,” IEEE INFOCOM 96, pp. 35-42, Mar 1996.
[3] D. Johnson, C. Perkins, and J. Arkko, “Mobility Support in IPv6,” IETF
RFC 3775, Jun 2004.
[4] “Fast Handovers for Mobile IPv6,” R. Koodli, Ed., IETF RFC 4068, Jul
2005.
[5] H. Soliman, C. Castelluccia, K. El Malki, and L. Bellier, “Hierarchical
Mobile IPv6 Mobility Management (HMIPv6),” IETF RFC 4140, Aug
2005.
[6] “Draft IEEE Standard for Local and Metropolitan Area Networks:
Media Independent Handover Services,” IEEE LAN/MAN Draft IEEE
P802.21/D06.00, Jun 2007.
[7] “Mobility Services Transport: Problem Statement,” T. Melia, Ed., IETF
RFC 5164, Mar 2008.
[8] “Mobility Services Framework Design,” T. Melia, Ed., draft-ietf-mipshop-
mstp-solution-01, Feb 2008. IETF Internet Draft, work in progress.
[9] Abate, J. and Whitt, W., ”The Fourier-series method for inverting trans-
forms of probability distributions,” Queuing Systems, Vol. 10, pp. 5–88,
1992.
[10] A. Rahman, U. Olvera-Hernandez, J.C. Z´ u˜ niga, M. Watfa, and
H.W. Kim, “Transport of Media Independent Handover Messages over
IP,” draft-rahman-mipshop-mih-transport-03, Jul 2007. IETF Internet
Draft, work in progress.
[11] J. Padhye, V. Firoiu, D. Towsley, and J. Krusoe, “Modeling TCP
Throughput: A Simple Model and its Empirical Validation,” IEEE/ACM
Transactions on Networking, Vol. 8, No. 2, pp. 133–145, Apr 2000.
[12] N. Cardwell, S. Savage, T. Anderson, “Modeling TCP latency,” INFO-
COM 2000: Proceedings of the Nineteenth Annual Joint Conference of the
IEEE Computer and Communications Societies , Vol.3, pp. 1742–1751,
26-30 Mar 2000.
[13] M. Mellia, I. Stoica, and H. Zhang, “TCP Model for Short Lived Flows,”
IEEE Communications Letters, Vol. 6, No. 2, pp. 85–87, Feb 2002.
[14] National Institute of Standards and Technology. NS-2 Mobility Package.
http://www.antd.nist.gov/seamlessandsecure.shtml, Jul 2007.

2 Mobile Node (MN) τ1 τ1 + τ1 + τ1 + 2( 1 1 Mobility Manager (MM) MSG1. Each message in the exchange sequence is generated at time τi = τi−1 + i−1 + Di−1 + Tproc (i). we consider TCP. For the exchange of M messages to be successful. For the case of the first message in the handover sequence. the delays will be correlated. at time τ2 = τ1 + 1 + D1 + Tproc (i). We use this simplification because the total time to complete the exchange. To model the case where packet losses occur. The amount of time required to move the ith message out of the transmission buffer is i = Li /B s. We assume that the delays {Di }M +1 are independent ideni=1 tically distributed random variables with cumulative distribution function FD (t) and characteristic function ΦD (ω) = ∞ fD (t) exp(jωt)dt. each message must be received successfully. . not M + 1. is not affected by the amount of time spent sending MIH ACKs. The node that receives the first MIH message immediately transmits an MIH ACK and. it arrives at its destination after some random delay that we denote as Di . MIH maintains a retransmission timer for each sent message. up to a limit of Ri retransmissions. M . If the ACK for the ith message is not received by the time the retransmission timer expires. TMIH (i). We assume that an MIH User that receives an MIH message immediately generates an MIH ACK followed by an MIH response message that constitutes the next part of the information exchange. Thus the probability that a message exchange fails is M Pfail = 1 − i=1 (1 − pRi +1 ).2 1 + TMIH (1)) ACK1. we assume independence to simplify the analysis. and the network transit time delay distribution. occurs. 1. we show results obtained using simulation tools that we developed that allow network planners to consider various of deployment scenarios. (1) where DM +1 is the time for the M th MIH message’s ACK to traverse the network. ΦH (ω). because we assume that the loss of the final MIH ACK message will not prevent completion of the message exchange. (2) The product is taken over M messages. Sample signal flow between MN and MM. 1. We denote the timer duration for the ith MIH message by TMIH (i). We assume that each . where fD (t) = dFD (t)/dt is the 0 density of D. The ith message traverses the network and.2 D1 τ2 MSG1.1 τ2 + Tproc (2) + ACK2.1 + TMIH (1) MSG1. and because the exchange’s state advances only when each of the two endpoints receives MIH messages. for i = 1. In Section II we consider the UDP transport where MIH uses ACKs and retransmission timers. in the course of this analysis. such as a handover signaling sequence. A sender will also re-transmit a message if the sum of the message’s transit time and the transit time of its corresponding ACK is greater than the timeout. II. we examine the effect of introducing MIH retransmission timers in addition to TCP’s recovery mechanisms.3 τ2 + Tproc (2) D2 τ3 MSG2. where Li is the length of the first message in bits and B is the bandwidth of the channel in bits/s. We use simulation results to verify our models in Section IV. MIH S IGNALING OVER UDP We first consider the case where the MIH User uses UDP to carry signaling messages. in which packet loss message size. The exchange of a message and ACK is successful with probability Ps = (1−p)2 . sends the second MIH message back to the node that initiated the message exchange. . MIH ensures reliability by requesting that each outgoing message be acknowledged by its recipient. This requires that at least one of Ri + 1 attempts to transmit the ith message be successful. let p be the probability that an MIH message or MIH ACK is lost while transiting the network. if it is not lost.1 2 t t Fig. and the exchange finishes when the MIH ACK for the M th MIH message is received. since the forward and reverse paths pass through the same network elements. 2. and examine some of the design trade-offs associated with the transport layer protocols. In Section III. The rest of this paper is organized as follows. In addition. MIH over UDP Latency We can derive the distribution of the time required to complete an exchange of M MIH messages. The time between the generation of the ith ACK and the transmission of the subsequent (i + 1)th MIH message is a deterministic quantity that we denote as Tproc (i). In addition. . τ1 is the time when the message is generated. the characteristic function of the M +1 M handover delay. In this case. In the absence of packet loss. is ΦD (ω) exp jω i=1 [ i + Tproc (i)] . In practice. The probability that a series of message transmission attempts does not fail is 1 − pRi +1 . we show the MIH ACKs as requiring no time to transmit. We define τi to be the time when the (i − 1)th message in the handover message exchange is received. . In the figure. We summarize our results in Section V. H. The total time required to complete the exchange is H = τM + M + DM + DM +1 − τ1 M = DM +1 + i=1 Di + i + Tproc (i) . This reflects the amount of time used by the receiving node to perform any tasks associated with the arriving MIH message. A. MIH will send a new copy of the message.3 ACK1. We show an example message exchange between a mobile node (MN) and mobility manager (MM) in Fig.

The characteristic function of the additional delay ∆D associated with timeouts caused by losses of the ith MIH i message is therefore Φ∆D (ω) i = = 1−p 1 − pRi +1 Ri pk ejkωδi k=0 1 − p 1 − (pejωδi )Ri +1 . The probability that k transmission attempts fail to reach the destination prior to a final.4) from [9]: FH (t) = FH (t) ≈ 2 ht + π π Nh Re[ΦH (kh)] k=1 sin(kht) . Ri . This does not add to the handover delay because the MM is already responding to the first MIH message. We consider an example to illustrate our results. the i characteristic function for H is M ΦH (ω) = (ΦD (ω))M +1 ejω(TP +T M ) i=0 Φ∆D (ω). . which is unity. k (7) + TMIH (i) units of time to the total time required to complete the signaling for the handover. and are given as multiples of µD . Consider the first message in the exchange. and h must be small enough to guarantee accuracy while not being so small that it results in rounding errors. and then computing 1 − FH (t). Decreasing the timeout duration to TMIH = 5µD produces the behavior shown in the red curve with squares. thus the probability that the ith message transmission is preceded by a delay of k · TMIH (i) is pk (1 − p)/(1 − pRi +1 ) for k = 0. then Equation (4) becomes ΦH (ω) = (1 − p)M (ΦD (ω))M +1 ejω(TP +T 1 − Pfail × (1 − (pejωδi )R+1 ) . The one-way transit delay is assumed to be exponential with an average value of 1 normalized unit of time. (1 − pejωδi ) i=0 M ) (5) where h is the step size for the summation on the ω-axis. At that time. We condition on the event that the message is successfully delivered during one of its allowed Ri + 1 transmission attempts. i (4) where TP = i=1 Tproc (i) is the total message processing time spent by both the connection endpoints. The value of 1−FH (t) at this plateau is two orders of magnitude below the plateau . 1 − pRi +1 1 − pejωδi (3) where the D superscript indicates that MIH is in datagram mode (i. and T = M i=1 i is the total time spent transmitting the messages that successfully reach their destinations. The performance curve has a staircase shape with plateaus that correspond to the packet loss probability and whose location on the time axis is determined by the MIH timeout. as shown in Fig. 1. Some plateauing is visible in this curve. We compute this probability by getting FH (t) = Pr{H ≤ t}. we can approximate this integral as follows. so that it is not possible for a message retransmission to arrive at the destination before a retransmission that was sent earlier. The number of retransmissions that are allowed for each message is R = 2. Each failure to send the ith message adds an additional δi i Because the characteristic function of H is the Fourier transform of the density fH (t). (6) π 0 ω Using the trapezoidal rule.01 and TMIH = 15µD . If the message is lost. We plot 1 − FH (t). up to a limit of R1 + 1 attempts in total. the cumulative distribution function of H. .4 of [10]. which is sin(ωt) 2 ∞ Re[ΦH (ω)] dω. If we decrease the packet loss probability to p = 10−4 while keeping the MIH timeout at TMIH = 15µD . successful attempt is pk (1 − p). 1 shows an MIH ACK being lost during the the first message’s second transmission attempt. which is Equation (4. The effect of message retransmissions is most clearly visible in the black curve with diamonds that is associated with p = 0. and Nh is the number of points taken in the summation. we can use it to get any desired statistic for H.11 WLAN network to a cellular network. For the computations that we performed for this paper we found that h ≈ 5 × 10−3 radians/s and Nh = 2 × 104 produced good results. we obtain behavior shown by the curve with blue circles. we assume that the loss performance is independent of that of all the other messages. t. If the MIH User allows the same number of retransmissions for each message. which is much steeper and which does not exhibit the same plateau effect that can be seen in the curve associated with p = 0. The MIH timeout values for all three messages are the same for each case plotted in the figure. We can get the distribution function by inverting the characteristic function using Equation (3. The processing time associated with each message is assumed to be zero. which may be associated with a desired level of Quality of Service (QoS) for the MN. the expected value of D. 1. the sender will generate a second copy of the first message. For instance. which was first transmitted at time τ1 . .3 retransmitted message arrives after its preceding copy. . Each additional message that is lost will cause a timeout at the sender’s MIH User and a retransmission of the original message.6) from [9]. we consider the case described in Section 10. Fig. In Fig.01 and TMIH = 15µD . using UDP transport). the sender will not receive an MIH ACK and its MIH event timer will expire at time τ1 + 1 + TMIH (1). 2. For each of the M messages in the handover. Since the total exchange time H accounting for dropped M +1 M messages is H = i=1 Di + i=1 i + Tproc (i) + ∆D . Nh must be large enough to produce an accurate approximation. so that Ri = R for all i. in which a mobile node (MN) signals a remote mobility manager (MM) to perform a handover from an IEEE 802. the first plateau is located at the same position on the time access as the first plateau the same packet loss probability and the larger MIH timeout.e. the probability that the handover time exceeds a given time. we may want to know the probability that the exchange time exceeds some threshold t. Lost MIH ACKs also trigger retransmissions. The number of messages required to complete the handover is M = 3. for three cases in which the probability of packet loss and the MIH timeout duration vary.

i i=1 which are themselves mutually independent.T H MIH MIH MIH = 10µ =µ D D 10 −1 p = 10 . For the plots in the figure. and µ∆D is i µ∆D i = = 1 dΦ∆D (ω) i j dω ω=0 (1 − p)Ri + 1 Ri + 1 − δi . as TMIH (i) → 0. ω=0 Each of the variances of {∆D }M are: i i=1 2 σ∆D i = E{(∆D )2 } − µ2 D i ∆ i = = (−1) d2 Φ∆D (ω) i dω 2 ω=0 − µ2 D ∆ i (Ri + 1)2 pRi +1 2 p − δi . i (9) where µD is the expected one-way transit time. allows consistent handover performance over a wide range of MIH timeout values. setting the timeout interval to be too small will result in unnecessary generation of duplicate copies of messages. p −2 10 −1 10 0 Fig.01. 1−p 1 − pRi +1 2 2 where limp→1 σ∆D = (Ri + 2)Ri δi /12. The three curves in Fig. associated with the packet loss probability value of 0. Fig. 2. 2 (1 − p) (1 − pRi +1 )2 (12) (8) which we can evaluate in the case of Equation (4) as M µH = (M + 1)µD + TP + T + i=1 µ∆D . Probability that the handover time. We can quantify the overhead penalty associated with a particular set of values for the timeouts {TMIH (i)}M by computing i=1 the expected number of messages generated per attempted exchange. i We can use Eq. including initiating the handover before the message loss probability becomes too large. H 2 σH. Indeed. 2. T . for short and long MIH timeout values (TMIH = µD and TMIH = 10µD . (9)-(12) to plot the mean and variance of H for MIH signaling over UDP versus the p.4 10 0 10 p = 10−2. 3. setting the timeout interval to be on the order of the network delay will minimize the effect of packet loss on the handover delay. we used the network parameters given in Table I. Thus a proactive approach to handovers. MIH over UDP Overhead (10) As indicated by Fig. the value of TMIH does not have an impact on the statistics of H for p < 10−3 . 3.T 10 1 D D p = 10−4. A small packet loss probability will also eliminate the effect of retransmissions. time units 30 35 40 10 0 10 −4 −1 10 10 −5 10 −6 10 −2 10 −4 10 −3 10 Packet loss probability. Thus the variance of H is M 2 2 σH = (M + 1)σD + i=1 2 σ∆D . The expected value of H is µH = 1 dΦH (ω) j dω . T = 10µ D TMIH = µD 10 −3 Time. adding to the overhead associated with the handover. respectively). Assuming that there are retransmission opportunities remaining. T = 5µ µ . and are independent of the additional delays {∆D }M . T = 15µ 10 Pr{H > t } −2 H 2 σ . exceeds t seconds for three sets of values of packet loss probability p and MIH timeout. the probability of message loss. We can get the first and second order moments of H from ΦH (ω). i (11) . T = 15µ D −2 2 µ . and the effect on µH is not noticeable for p < 10−2 . in Fig. However. H. the additional time required to complete a handover vanishes. the sending node will transmit an additional copy of a message if the previous copy is lost. The processing times and packet transmit times are deterministic. i Recall that the transit times {Di }M are mutually indepeni=1 dent. p. B. We let TMIH take values of µD and 10µD to show the effect of varying the MIH timeout. but we let µD = 100 ms. sec 0 5 10 15 20 25 t. Mean and variance of the handover time H versus packet loss probability. 2 demonstrate that the performance of MIH over UDP does not depend on the values of the retransmission timeout if the timeout period is on the order of the round-trip time between the MIH connection endpoints. For this set of network parameters. If the previous copy is where limp→1 µ∆D = Ri δi /2.

given that Ri retransmissions are allowed. the sending node will produce m copies with probability πm as defined in Equation (14). particularly when the timeout is on the order of the average packet transit time across the network. if D is exponential with mean µD . p: 0. an ACK is generated by a node each time its MIH User receives an MIH message. . In the second case. As the packet loss probability increases to near unity. Ri R Prt i (i). (17) = (1 − p) To get the expected number of messages of both types sent during the exchange. and 10−6 in Fig. we sum ai over all messages {i} and combine the resulting expression with the expected number of MIH messages from Equation (16). a retransmission will take place if the corresponding MIH ACK is lost. For ease of computation. N → 2 i=1 (Ri + 1). 1 − Prt (i) R m−1 mPrt (i) + (Ri + 1)Prt i (i) (15) and the average total number of messages required for the exchange is M M N MSG = i=1 ni = i=1 R 1 − Prt i +1 (i) . For example. a retransmission will occur if the time from the message transmission to the ACK reception is greater than TMIH (i). the characteristic function of RTT is ΦRTT (ω) = (ΦD (ω))2 . and µD = 1. (13) connection with the ith message is Ri +1 m ai = m=1 πm k=0 Ri +1 m=1 m kpm−k (1 − p)k k mπm = (1 − p)ni . which happens when p → 1 or when TMIH (i) → 0. Ri +1 ni = m=1 mπm Ri = = (1 − Prt (i)) m=1 R 1 − Prt i +1 (i) . . (14) From this expression we can get the expected number of messages sent. m = Ri + 1. m = 1  m−1 Rm +1 Ri +1 p ). . 1 − (1 + t/µD )e−t/µD . t<0 t ≥ 0. however. accounting for the loss of all Ri +1 copies . ni → Ri + 1 as Prt (i) → 1. the cumulative distribution of RTT is FRTT (t) = 0. exchanges fail if all of the allowed number of transmission attempts at one of the stages result in the MIH message’s not reaching its destination.1. The curves in Fig. Thus the probability that the mth message is the last message in a given exchange is  pR1 +1 . In general. so that all times are normalized to the average transit time. 2. m = 1. N = (2 − p) i=1 R 1 − Prt i +1 (i) . The probability that any one of these m copies is lost is p. In order to assess the load that the MIH Users collectively offer to the network. is πm = m−1 (1 − Prt (i))Prt (i). we see a significant deviation in N . In these cases. Equivalently. if we let p → 1 we would obtain a horizontal line corresponding to N = M · R = 9. we assume that the RTT is the sum of two independent one-way transit times that each have the cumulative distribution FD (t). Thus the probability that a given message transmission attempt results in another copy of the message being sent is Prt (i) = p + (1 − p)[p + (1 − p)(1 − FRTT (TMIH (i)))] = (2 − p)p + (1 − p)2 [1 − FRTT (TMIH (i))] = 1 − (1 − p)2 FRTT (TMIH (i)).5. our treatment of the message exchange overhead has considered only those cases where the exchange completes successfully. 4. . The probability that all copies of the ith MIH message are lost is pRi +1 . 1 − Prt (i) (18) The probability that the sending node generates m copies of the ith message. the probability that k copies out of m are successfully received by the destination node is m Ck pm−k (1 − p)k .5 not lost. 0. giving M where FRTT (t) is the cumulative distribution of the message round-trip time. The resulting curves show that the average number of messages per exchange is insensitive to changes in p over a wide range of values for the MIH timeout. 1 < m < M χm = i=1 (1 − p  M −1 (1 − pRi +1 ). Therefore the expected number of MIH ACKs that are generated in From Equation (15). If neither the message nor its ACK are lost. We can count those as well. m = M i=1 The expected number of MIH messages that are generated during an exchange. We ignore the first case because the probability of the exchange failure is M 1. which is the maximum possible number of messages that can be generated. For the ith message in a M -message sequence. Thus ∞ FRTT (t) = 0 fD (u)FD (t − u)du where fD (t) is the density of D. We consider the example from Section II-A in which M = 3 and R = 2. at least one of the Ri + 1 transmission opportunities succeeds at each of the M stages of the exchange. we must consider the effect of transmissions associated with incomplete exchanges. 4 suggest that we can prevent the number of messages generated per exchange from being excessively large if we let the MIH timeout be at least two times larger than the measured packet transit time. RTT. and each copy results in an MIH ACK’s being sent back. 1 − Prt (i) (16) This expression does not account for the MIH ACKs that are generated each time a node receives an MIH message. Ri +1 copies of the ith message are generated for all i because the timer immediately expires after each message transmission. So far. We plot N for three values of the packet error probability.

If there is no packet loss. Conversely. We assume that there are no unacknowledged segments when the exchange begins. . M }. Our latency analysis considers the degenerate case in the model [12] in which the number of segments to be transferred at any time is one. are in use. we assume that the SYN/SYN-ACK exchange has already taken place during the start up of the MIH User. RTOmax = 2K RTO. . M } as well. the first portion of the MIH message. the MIH message may not get sent until enough duplicate copies are put into the TCP transmit buffer to fill a segment and trigger its transmission by TCP. [11] and the work that builds on it. TCP will not send the segment until enough bits have been put into the transmission buffer by the MIH User. depending on whether it is in Slow Start mode or Congestion Avoidance mode. . such as those associated with MIH signaling. This is because TCP is a stream-oriented protocol rather than a packet-oriented one. . instead. [12]. if the MIH message is larger than a segment. Here. . In particular. This mismatch will present problems with regard to completing a message exchange in a timely manner. which will be sent as soon as it is put into TCP’s transmit buffer. A TCP model for short-lived flows. We can obtain a simplified version of this expression if we assume that Ri = R and TMIH (i) = TMIH for i ∈ {1. Then. the total number of MIH messages in the exchange is N = (2 − p)N MSG . We compute the performance metrics associated with a message exchange between MIH Users by using the techniques from Section II. If the MIH message is smaller than a TCP segment. the MIH message may not get sent at all. We consider the case of MIH over TCP. MIH allows TCP to be solely responsible for guaranteeing message reliability and does not use its own timeouts. . For this discussion.6 18 p = 5 × 10−1 Mean number of messages per handover 16 14 12 10 8 6 4 2 0 p = 10 p = 10 −1 −6 0 2 4 6 TMIH. the source node will continue to double its RTO value until it reaches a maximum. the destination node will generate an acknowledgment for the source node as soon as its TCP layer receives the source’s segment. 2. 4. Thus. . which is a connection-based. is M m N MSG = m=1 χm i=1 ni . a source node will not respond to a packet loss by going into Fast Retransmit/Fast Recovery mode. If MIH timeouts. TCP will automatically retransmit messages if its own retransmission timer expires or if it receives three duplicate acknowledgments from the destination node at the other endpoint of the connection. M }. . Therefore. up to an integer multiple of a segment length. In addition. If subsequent retransmitted messages are also lost. 2. was developed in [13] but does not incorporate a . A. Because we assume that the source node does not have any outstanding unacknowledged segments when it begins the exchange. it will wait until its retransmission timer expires. . we assume for the sake of tractability in the analysis that the size of each MIH message is equal to the size of the data payload of a TCP segment. (1 − Prt )pR+1 (19) Since ai = (1 − p)ni . the node initiating the message exchange will be able to transmit the first message immediately. multiples of mean transit time 8 10 Fig. If MIH is not using timeouts. 2. The time to send each message is therefore s for all i ∈ {1. III. When this happens. 3. will get sent while the remainder of the message will remain in the TCP’s buffer until enough additional bits are generated by the MIH User to cause the segment to be transmitted. it will advance its own transmission window and possibly expand its congestion window. Mean number of messages generated by MIH signaling endpoints during an exchange where M = 3 and Ri = 2 for i = 1. where K is the maximum N MSG = = = (1 − q)n m=1 mq m−1 + M q M −1 n (1 − q M )n 1−q R+1 (1 − Prt )(1 − (1 − pR+1 )M ) . MIH S IGNALING OVER TCP MIH provides reliable transport over UDP by using internal timers to trigger retransmission of lost or excessively delayed messages. we get M −1 maximum retransmission timeout (RTO). . the source node will shrink its congestion window to one segment. have respectively characterized the loss and delay behavior of TCP during a steady-state bulk data transfer and during the transmission of an arbitrarily small amount of data. each segment contains exactly one MIH message. Prt (i) = Prt for i ∈ {1. . described in Section II. it will not receive any duplicate acknowledgments from the destination if the first message is lost. 2. In practice. and double the RTO value. stream-oriented protocol that provides reliable transport of application data. There has been voluminous work on TCP behavior. Defining R+1 1−Prt q 1 − pR+1 and n 1−Prt . an MIH message may not fit exactly within a TCP segment. Because the TCP transmission window will be at least one segment long. of the ith message at any point in the exchange. retransmit the message. When the source node receives the acknowledgment. . MIH over TCP Latency In the absence of MIH timeouts.

and the MN’s TCP will send another copy if no ACK arrives by the time τi + 3 + 7TRTO . Our latency analysis is similar to our treatment of MIH over UDP from Section II. with each successive timeout and the lack of a limit on the number of retransmission attempts. However the TCP ACK is lost on the way back to the MN. We plot performance curves for the case where M = 3 in Fig. The expression for µH is nearly identical to Equation (9). H. 6. and number of RTO stages. it generates a response MIH message at time τi+1 + Tproc (i + 1). In the figure. We can use the characteristic function to obtain the expected value and the variance of the exchange time. 1−p (1 − p)(1 − 2p) (21) .3 Di τi+1 0 p = 10−4. RTO. the ith message in the message sequence. we observe the same plateau effects that we saw in the UDP case. exceeds t seconds for six sets of values of packet loss probability p. 2. Fig. K = 6 10 −3 MSGi+1.1 + ACK τi+1 + Tproc (i + 1) 10 −4 τi+2 τi + 3 + 7TRTO t ACKi+1. H. Thus the only change is a new expression for the additional time penalty imposed on the ith message in the message sequence.1 10 −5 t 10 −6 0 5 10 15 20 25 t. When the MN receives this message at time τi+2 . We assume that TCP will continue to send copies of the lost segment until it receives an acknowledgment from the destination endpoint. R = 2. This message is contained within a TCP segment whose header’s acknowledgment field contains the byte number that follows the last byte of the ith message. R = 4. number of times that TCP doubles the RTO value. K = 6 p = 10 . after which the RTO expands to RTOmax . The lack of a TCP ACK causes the sending node’s TCP to retransmit the segment and double the RTO. For subsequent retransmissions (r > K). This segment is also lost. 5.2 −2 p = 10−4.2 MSGi. latency is minimized by using small timeouts and fewer backoff stages. so the MN’s TCP sends a second copy of the segment at time τi + 2 + 3TRTO and doubles the RTO interval again. the only change is replacing µ∆D with the expected value of ∆C . R = 2. 1 − pejω( +RTOmax ) (20) Inserting this expression into Equation (4) in place of Φ∆D (ω) i gives us the characteristic function for H. The first retransmission attempt succeeds in the sense that the MM receives the message and begins preparing a response message.7 Mobile Node (MN) MSGi. Because the RTO doubles with each retransmission. 6. up to a maximum amount. K = 6 p = 10−4. they are most noticeable when the packet loss rate is high. Because the destination node receives the first transmission of the MIH message at time τi+1 = τi + + Di . Sample MIH signal flow between MN and MM when TCP transport is used and when some messages and TCP ACKs are lost in transit. The characteristic function of the extra delay ∆C caused by packet loss when i MIH is using TCP transport (i. . R = 2. K = 2 10 −1 p = 10−4. nor does it affect the processing times for each message at the MIH layer. Immediately after the rth retransmission attempt. which shows that. K = 6 p = 10 . Note that the use of TCP does not affect the time penalty associated with each successful transit. In addition. we can again get Pr{H > t} = 1 − FH (t) for various values of t. K = 6 −2 −2 10 Pr{H > t } τi + 2 + 3TRTO ACKi. the time when the rth retransmission occurs (given the first transmission attempt occurred at time τi ) is τi +r +TRTO +2TRTO +· · ·+2r−1 TRTO = τi + r + (2r − 1)TRTO for r = 1.1 Mobility Manager (MM) 10 τi τi + + TRTO MSGi. K. We show an example of an MIH message transmission over TCP in Fig. Probability that the MIH message exchange time. . . which is transmitted at time τi . similar to the UDP case. R = 4. TCP expands the RTO to 2r TRTO . K. time units 30 35 40 Fig. By inverting ΦH (ω) to obtain FH (t) using Equation (6).e. the retransmission occurs at time τi + K + 2K TRTO − TRTO = τi + K + RTOmax − TRTO . . is lost in transit. At the Kth retransmission attempt. which we i i obtain from Equation (20) as follows: µ∆C = i (1 − p − 2K pK+1 )p p + TRTO . R = 1. the RTO remains at its maximum value and the rth retransmission occurs at time τi + r + (r − K + 1)RTOmax − TRTO . in connection mode) is thus Φ∆C (ω) i K = (1 − p) r=0 pr ejω[r ∞ +(2r −1)TRTO ] + (1 − p) r=K+1 K pr ejω[r pr ejω[r +(r−K+1)RTOmax −TRTO ] = (1 − p) r=0 +(2r −1)TRTO ] + (1 − p)pK+1 ejω[(K+1) +2RTOmax −TRTO ] . The principal differences lie in the exponential expansion of TCP’s timeout window. its TCP layer notes the acknowledgment and stops retransmission of the ith message. 5.

(11) with σ∆C replacing σ∆D . Note that the performance of UDP and TCP is identical for small values of p. We can also relate Equation (21) 1 to the expression for C1 (the mean time to send one segment with an initial congestion window size of one segment) in [13] by letting K → ∞. k ≤ K 1 − (1 − p)2 FRTT (RTOmax ). When TCP is used. Defining psi 1 − (1 − p)2 FRTT (RTOmax ). 7. p −2 10 −1 10 0 Fig. As in the case of µ∆C .T 10 1 H 2 σ . 7. (25) (22) This quantity becomes undefined when p = 1/2 or p = 1/4. µD and 10µD . (4p)K p(7−29p−4p3 +pK+2 −4pK+3 ) (1−p)2 (1−2p)2 (1−4p) B. for short and long TCP RTO values (TRTO = µD and TRTO = 10µD . RTOmax . −(2p)K p(2+K−3(K+1)p+2Kp2 ) p (1−p)2 (1−2p)2 TRTO the mean and variance of H increase without bound as p → 1. G(p) = 1 − 2p The cause of the additional factor of p is that Equation (18) gives the average duration of Z T O . Mean and variance of the handover time H versus packet loss probability. given that a timeout occurred. we can take limits and get i 2 lim σ∆C = i p→1/4 4 9 2(−5 + 9 · 2K ) − 3K TRTO 9 · 2K 4K (27K − 1) + 9 · 2K+1 − 1 2 + TRTO 9 · 4K+1 2 + (23) and p→1/2 2 lim σ∆C = 2 i 8 + 5K + K 2 TRTO 2 2(−9 + 13 · 2K ) − 8K − K 2 2 + TRTO . 3. which itself comes from [11]. p.T H RTO RTO RTO = 10µ =µ D D lim µ∆C i K+ 2 TRTO . We can use L’Hˆ pital’s Rule to get o p→1/2 10 2 µ . respectively). because we are computing the additional delay due to message losses. with an additional factor of p. using the same parameters that we used to generate Fig. the probability that the kth transmission attempt fails and requires another attempt is φk = 1 − (1 − p)2 FRTT (2k−1 TRTO ). by letting → 0 and observing that the general form of K Equation (19) from [12]. + − 1−p (1−2p)2 (1−4p) + (2p)K+1 (1−6p−(2K+4 −9)p2 −4p3 ) (1−p)2 (1−2p)2 (1−4p) 2 pTRTO + p 2 (1−p)2 . Similar to the UDP case. . G(p) = 1 + k=1 2k−1 pk .8 The second fraction assumes the indeterminate form 0/0 when p = 1/2. the occurrence of a timeout is not a given condition. (24) 4 2 + 2 We show plots of µH and σH in Fig. We can manipulate our Equation (21) to produce Equation (18). and that the value of p for which there is a noticeable 2 increase in the value of µH and σH is around the same for both the UDP and TCP transport cases. but this benefit is less meaningful in the case of signaling for handovers. The gain from doing this is the eventual success of the transmission as long as p = 1. H 2 σH. However. T = 10µ D TRTO = µD Equation (21) is closely related to Equation (18) in [12]. so that degrading link conditions do not expand the amount of time required to successfully deliver signaling messages. We get the variance of i i C ∆i in the same way that we got Equation (12). Our Equation (21) gives the average time to transmit a segment in addition to the network transit delay. since their model assumes no upper bound on the RTO value. its TCP ACK is lost or the RTO timer expires before the sender receives the ACK. MIH over TCP overhead We also can compute the expected number of messages generated during a message exchange as a function of the RTO value TRTO and the packet loss probability. thus Z T O must be at least TRTO time units long. We assume that there is no limit on the number of times TCP attempts to retransmit a segment. the TCP timeout period. we have to account for the exponential dilation of the timeout interval until it reaches its final value. where we assume that TRTO is the same for all M messages in the message sequence. sec µ . and so ∆C i has a minimum value of zero. this figure demonstrates the importance of beginning critical handover signaling exchanges early. which gives us a more complex but still closed-form expression: 2 σ∆C = 2 (1−p) i 2 10 0 10 −1 10 −2 10 −4 10 −3 10 Packet loss probability. The probability that m copies of the ith message gets sent is the probability that the TCP ACK for the mth transmitted copy of the MIH message arrives within the time limit after m − 1 failures. The variance of H when MIH messages use TCP is given 2 2 by Eq. p. can be written as 1 − p − 2K pK+1 . Again. similar to Equation (13). because the amount of time available to complete message delivery is short. Here we consider two values of the TCP RTO value. = + 2 Time. because TCP attempts to deliver an unacknowledged segment until it receives an ACK. k > K. retransmission of an MIH message occurs if the message is lost. The expression in [13] has an additional factor of p as well. Thus.

we know that τ2 − τ1 = D1 + ∆C . An important difference is the asymptote at RTO = 0. In addition to the MIH messages. There are two cases involving the relative values of the MIH timeout TMIH and the TCP timeout. µD . the set of positive integers. The handover exchange is composed of M = 3 MIH messages where TCP is used for transport and there are no timeouts and retransmissions at the MIH User. φk = ψ for k > K) we have  1 − φ1 . The MIH/TCP case is similar to the MIH/UDP case in that reducing the timeout to be on the order of the network round-trip time significantly increases N in both cases. Because the retransmission parameters do not change from message to message. and the total number of messages created in connection with the ith phase of the exchange is (2 − p)ni . This exchange involves one MIH message and one MIH ACK. 4. where n is given in Equation (27). for various RTO values. resulting in unnecessary extra traffic. The MIH layer will produce m additional copies of the 1st message. it follows that the probability of exactly R1 retransmissions is the probability that U1 is greater than R1 TMIH (1).e. 8. ni is a constant. The probability of this event is Pr{U1 ∈ mTMIH (1). and the mean total number of messages sent during an exchange is N = (2 − p)M n. k out of m segments generated by a source will arrive at their destination (and cause ACKs to be sent back) with probability m−k (1−p)k .9 Mean number of messages per handover (i. We consider only the case where the MIH timeout occurs later than the initial value of the TCP RTO (TMIH ≥ TRTO ). we present an analytical model that quantifies some of the negative effects that result from combining the TCP layer and MIH User retransmission mechanisms. C. Letting the MIH timeout occur before the RTO would introduce additional copies of the message into the TCP queue before TCP has a chance to time out and retransmit. which is 1 − FTM (R1 TMIH (1)). R1 . In this situation. 0 ≤ m ≤ R1 . lies in the half-open interval mTMIH (1). From our analysis in Section III. In Fig. U1 . Thus the expected number of ACKs is (1− m Ck p p)ni . 1 < m ≤ K πm =  K (1 − ψ)(ψ)m−K−1 k=1 φk . Our model examines the case of an MIH Indication message transmission from a MN to the MM. The number of messages that the MIH User generates depends on the amount of time that TCP spends in Timeout retransmitting the original copy of the message. We first examine the Request message from the MN to the MM. m = 1  m−1 (1 − φm ) k=1 φk . multiples of mean transit time 8 10 p = 5 × 10−1 p = 10 p = 10 −1 −6 (26) for m ∈ Z+ . the connection endpoints generate TCP ACKs every time they receive a TCP segment. if the time from the first message transmission attempt until the TCP layer receives an ACK. (m+1)TMIH (1) . the MIH User’s timeout occurs after the TCP RTO. but this is a TCP function that is not controllable by the MIH User. so we have already characterized the random 1 . In contrast. we plot the average number of messages in a handover versus RTO. so the MIH layer will not begin generating duplicate messages until TCP has been in Timeout mode for some period of time. m > K 30 27 24 21 18 15 12 9 6 3 0 0 2 4 6 RTO. The MIH User generates an MIH ACK message for each MIH message that it receives. = 1 − φ1 + m=2 m(1 − φm ) k=1 K φk (27) + 1 + K(1 − ψ) 1−ψ φk . for a total of two messages generated jointly by both nodes’ MIH Users. Using this expression we can get the mean number of copies sent for a given message: ∞ ni = m=1 mπm K m−1 = 1 − φ1 + m=2 ∞ m(1 − φm ) k=1 φk K + m=K+1 m(1 − ψ)(ψ)m−K−1 k=1 K m−1 φk Fig. TRTO . The figure bears a strong resemblance to Fig. Mean number of messages generated by MIH signaling endpoints using TCP transport during a handover where M = 3 and K = 6. The Effect of Combining MIH Timeouts with TCP In this subsection. k=1 As was the case with MIH transport over UDP. which plots N versus the MIH timeout for the MIH/UDP case. (28) Because MIH is limited to a finite number of retransmissions. MIH is limited to a finite number of retransmission attempts when it uses UDP. Thus a conservative RTO estimate reduces overhead. which is due to our assumption that TCP never stops trying to send an un-ACKd segment. where the RTO is expressed as multiples of the average one-way network transit time. (m + 1)TMIH (1) } (m+1)TMIH (1) = mTMIH (1) fU1 (u) du = FU1 ((m + 1)TMIH (1)) − FU1 (mTMIH (1)). 8.

Because of this. we show the exchange of messages between the MN and MM when MIH reliability mechanisms (ACKs and retransmissions) are used in addition to TCP. The delay experienced by the MIH ACK is D1. the characteristic function for ∆C 1. There are two ways the MN can get a TCP ACK from the MM.ACK . We next get the characteristic function of U1 . the MM will send the 2nd MIH message before the RTO expires. this assumes that the MM’s TCP uses delayed ACKs. In this case. the additional time required to send the MIH ACK for MSG1 from the MM to the MN. when Tproc (2) ≤ TRTO . the MN’s TCP immediately responds with a 40-byte TCP ACK. and when Tproc (2) > TRTO . Recalling that the maximum RTO is RTOmax = 2K TRTO . we see that ∆C 1.ACK ∆C 1 D1 Mobility Manager (MM) time from the first transmission attempt to the segment’s successful reception at the MM. so ∆C + TRTO with probability (1 − p)p2 . the MM’s MIH layer sends an MIH ACK to the MN whose TCP header contains the sequence number of the last octet in MSG1 . TCP will continue normal operations. and the MM’s TCP transmission window is at least two segments long. We assume that all MIH messages and MIH ACKs have the same transmission duration of units of time. 1. We compute the characteristic function for ∆C 1. and begin retransmitting the MIH ACK while the 2nd MIH message waits in the transmission queue. (9). Immediately upon receiving MSG1 . 2. 9. Failed transmission attempts by the MM’s TCP cause an additional delay of ∆C 1. If Tproc (2) ≥ TRTO and the MIH ACK gets lost in transit. The MM’s TCP will behave exactly like the MN’s TCP. Like [12]. There is an additional time penalty of ∆C units associated with failed 1 attempts by the MN’s TCP to transmit MSG1 . the message will go into the transmission queue until the MM’s TCP receives an ACK.ACK . The MN can also receive a TCP ACK in the header of the next MIH message in the handover sequence.ACK takes the values 0 and Tproc (2) with probabilities (1−p) and (1−p)p.ACK also has characteristic function ΦD (ω).ACK = . respectively.ACK . .ACK If Tproc (2) ≤ TRTO . There are two possibilities. K. 1 1. If the ACK is received before TRTO units of time have expired. Mobile Node (MN) τ1 lost MSG1 copies MSG1 MSG1 ACK U1 MSG2 add’l lost MSG1 ACKs MSG1 ACK TCP ACK τ2 τ2 + Tproc (2) τ2 + TRTO Mobility Manager (MM) t t Fig. If the MN receives the ACK after TCP has gone into timeout. RTOmax = 2K TRTO .ACK D1. TCP will end the timeout and begin Slow Start. ∆C 1.ACK 2 D1. as long as the MM’s transmission window is large enough to send both the MIH ACK and the second MIH message. . as shown in Fig. The MN can receive a TCP ACK in the header of the MIH ACK that the MM’s MIH User generates when it receives the MN’s MIH message. If no ACK comes back from the MN by the time t = τ2 + TRTO .ACK t t Fig. Signal flow for an exchange of an MIH message and MIH ACK message between MN and MM over TCP. we assume that TCP ACKs are never lost in transit. where D1. 1.ACK = Continuing this pattern and using the results of our analysis r in Section III. The MN has two chances to receive a segment containing an ACK for the 1st MIH message and send an ACK back to the MM. the local TCP will note that the transmitted segment was correctly received.ACK changes slightly from its form in Equation (20). as D1 . . for r = 1. the MM’s TCP will enter timeout. which has the same characteristic function. 10.ACK = r + (2 − 1)TRTO with r+1 probability (1−p)p . (29) where Φ∆C (ω) is given by Equation (20). we see that ∆C 1. The first retransmission of the MIH ACK happens at time t = τ2 + + TRTO if no ACK is received for either the MIH ACK or the 2nd MIH message. Signal flow for an exchange of an MIH message and MIH ACK message between MN and MM over TCP. when Tproc (2) > TRTO . we show the case where both the MM’s MIH ACK and MSG2 transmissions fail. . shrink its congestion window to one segment. retransmitting the MIH ACK’s segment and expanding the RTO at each attempt until it reaches its maximum value.ACK time units. its characteristic function is ΦD (ω)Φ∆C (ω). When any one of these packets first arrives at the MN. From Fig. When the MIH ACK finally arrives at the MN. ΦU1 (ω) = ΦD (ω) Φ∆C (ω)Φ∆C (ω)ej2ω . The 1st MIH message travels from the MN to the MM and is delayed by D1 units of time. ΦD (ω). So the MM will have one attempt to send the 2nd MIH message before its TCP goes into timeout. In Fig. 1 the time from the 1st message’s reception at the MM to the arrival of a TCP ACK at the MN. depending on whether Tproc (2) is larger than TRTO .10 Mobile Node (MN) τ1 lost MSG1 copies MSG1 U1 lost MSG1 ACKs MSG1 ACK TCP ACK τ2 ∆C 1. In the figure. as we now show. the MM’s TCP will time out before the 2nd MIH message is ready to be sent. 9. which reaches the MM after a delay of D1. 10.

B is insensitive to p for p < 0. . T RTO RTO RTO = 200 ms = 200 ms = 200 ms TMIH =150 msec. IV.ACK With the characteristic function ΦU1 (ω) in hand. Thus we obtain the following expression for the characteristic function when Tproc (2) < TRTO : Φ∆C (ω) 1. the small TRTO curves are flatter and begin increasing toward R at a larger value of p.5 (1 − p)[1 + pe +(1 − p) + (1 − p)p ] +(2r −1)TRTO ] 0 −2 10 10 p −1 pr+1 ejω[r e 1 − pejω( 10 0 r=1 K+2 jω[(K+1) +2RTOmax −TRTO ] +RTOmax ) . TRTO = 2 ms 2 TMIH =200 msec. 1. the probability of message loss for different values of the MIH timeout. 12 that consists of a mobility manager located at a remote Point of Service (PoS) connected to an access point via a backbone network that we have abstracted as a single router and a lossy link.ACK = (1 − p) + (1 − p)pejωTproc (2) K 3 T T 2. Computing B using different values for Tproc does not result in a significant change in the curves.5 second. . 1. T =200 msec. 2. S IMULATION R ESULTS To further quantify the performance of the MIH User over different transport layers. thus reducing the backlog. This is because decreasing TRTO = 200 increases the TCP retransmission rate and it becomes more likely that a TCP ACK will arrive at the MN before the MIH timeout expires. In each scenario that we considered. Then. Within the coverage area of the access point is a single mobile node that communicates with the access point over a IEEE 802. TRTO = 2 ms TMIH =400 msec. 200 ms. TMIH . the average size of the message backlog generated by MIH retransmissions due to expiration of the MIH timer. we can work out the characteristic function of U1 by conditioning on the whether MSG2 is lost in transit. depending on whether we are simulating Indications or Request/Response exchanges. TRTO = 2 ms 1. as follows: ΦU1 (ω) = (1 − p) ΦD (ω) Φ∆C (ω)Φ∆C (ω)e 1 1. TMIH : 150 ms.5 B +(1 − p) + (1 − p) = pr+1 ejω[r r=1 ∞ +(2r −1)TRTO ] pr+1 ejω[r r=K+1 jωTproc (2) K +(r−K+1)RTOmax −TRTO ] 1 0.5 T MIH MIH MIH =150 msec. 1 where Φ∆C (ω) is given by Equation (30). (10).01. Thus the average number of MIH messages put into the TCP transmission queue for each MIH message that the MIH User needs to send is the original message plus the ones in the backlog. Using Equation (28) and the inversion formula in Equation (6). π 0 ω (32) 2 2 +Tproc (2)) 2 j2ω Fig. we get β(m) Pr{m messages in backlog}: = Pr{mTMIH (1) ≤ U1 < (m + 1)TMIH (1)} = FU1 (m + 1)TMIH (1) − FU1 mTMIH (1) 2 ∞ Re[ΦU1 (ω)] sin (m + 1)ωTMIH (1) = π 0 dω − sin mωTMIH (1) (31) ω for m = 0. . we can get an estimate of the MIH overhead by computing the size of the backlog of copies of the MIH Indication message at the MN at the time the MN receives a TCP ACK for the MIH Request message. or 1 + B.11 wireless LAN link.5 second or the PoS generates requests every 0. The figure also shows that increasing the MIH timeout reduces the size of the backlog. including TRTO = 200 ms. We used the set of parameters shown in Table I. we used a simple network topology shown in Fig. we also used TRTO = 2 ms to show the effect of increasing the TCP retransmission rate. . Note that using a smaller value of TRTO produce curves that have the same value for small values of p as the curves where TRTO = 200 ms. the MN generates Indications every 0. although it will increase the TCP segment transmission rate. (30) From Fig. and 400 ms.ACK + p ΦD (ω) Φ∆C (ω)ejω(2 1 = × ΦD (ω) Φ∆C (ω)ej2ω .ACK of the MIH timeout. we performed a set of simulations using the ns-2 tool.11 r + (r − K + 1)RTOmax − TRTO with probability (1 − p)pr+1 for r ≥ K. respectively. which we enhanced by extending the tool’s mobility framework [14]. we we plot the average backlog size versus p. Plot of B. (1 − p)Φ∆C (ω) + pejωTproc (2) 1. R1 − 1 and β(m) β(R1 ) = Pr{U1 > R1 TMIH (1)} = 1 − FU1 R1 TMIH (1) 2 ∞ dω = 1− Re[ΦU1 (ω)] sin ωR1 TMIH (1) . For all three values of TMIH and for TRTO = 200 ms. 11. T =400 msec. 11. the MN first connects to the AP and registers with the PoS. versus p. In Fig. We take performance measurements by taking traces of the relevant output parameters and averaging those traces over The expected number of messages in the backlog is B = R1 m=0 mβ(m). the probability of MIH message loss for three different values . To generate our results.

11 Data rate (Mb/s) 11Mb/s Coverage area radius (m) 50 Links Speed (Mb/s) 100 Delay (s) 0. The graph further shows that the probability that a message exchange is unable to complete successfully decreases as the number of the packets in the exchange increases. 0.75.2 Max retransmission Unlimited Queue size Unlimited Header size (bytes) 20 IP header IPv6 header (bytes) 40 MIH Function Transaction timeout (s) none Ri 2 Tproc (2) (s) 0. 14 we plot the average time required to complete the simple Indication transmission as well as the Request/Response message exchange. In the figure. we examine the delay performance of MIH over both UDP and TCP transport layers.5.5 s for Indications and TMIH = 0. In Fig. we obtained excellent agreement between the values predicted by the model and the results that we obtained from the simulations. The parameters that we used in our simulations are shown in Table I. the sets of parameters that we used are indicated in the legend appearing in the figure. In all the scenarios that we considered. 0. We examine the delay with MIH acknowledgment messages and retransmissions (for which TMIH = 0.3. We compare the results from ns-2 with theoretical results that we obtained by computing 1 − Pfail from Equation (2). Network topology used in simulations of Indication and Request/Response MIH message exchanges. The first type consists of a Indication message sent by the mobile node to the mobility manager. 13. We used a variety of values for the TCP maximum RTO. In addition.5. as shown in the table. 12. 0.3 s for Request/Response) and without.01 UDP Max packet size (byte) 1000 Header size (bytes) 8 TCP Max Segment Size (bytes) 1280 Min RTO (s) 0. We plot values for the probability that MIH transaction completes successfully in Fig. The typical standard deviation of the data in each of the graphs in this section is on the order of 1% of the mean data value plotted. 50%] RTOmax (s) 0. i. Both types of MIH message exchanges occur according to a Poisson process with an average exchange arrival rate of two events per second.12 TABLE I S IMULATION PARAMETERS IEEE 802. using MIH acknowledgments significantly increased the success rate. figures in this subsection. MIH Delay In this subsection. A. the upper terminus of a bar is located at a value one standard deviation greater than the corresponding mean. between the 5 s and 4005 s marks. rather they show the value of the standard deviation of the completion time. The bars in the figure do not show the variance of the simulation data. Two different transport layers were used in the simulations: UDP and TCP. and varied the link loss rate for each maximum RTO value that we used. thus error bars would not be visible and so are not plotted. Theoretical and simulated values for MIH transaction completion probability. The packet loss rate on the link connecting the router to the access point was allowed to vary between 0 and 0. 4000 seconds of simulation time. The second type of exchange involves a Request message generated by the mobility manager which produces a Response message from the mobile node. The parameters for both transport layers are given in Table I. p variable [0.e.2. 1 Indications/s 2 Requests/s 2 MIH Packet size (bytes) 200 Fig.2 Simulation configuration Duration (s) 6005 loss probability. there are two sets of horizontal marks that corre- . Each simulation run covered a time interval of 6005 seconds. as we would expect. In all of the Fig. The simulations examined two types of MIH message exchanges. 13.

Thus.13 Fig. We observe that the exchange completion time decreases when the packet loss reaches 45% for larger values of RTOmax . 14. we see that the average completion time is much higher in the case of TCP. the message exchange will fail unless the initial attempt to transmit each MIH message is successful. the MIH User sends a Request and waits for a Response. the transaction delay for a given maximum RTO value is approximately 2. These duplicate packets cause additional delays because all of them must be transmitted and acknowledged before the next MIH Indication message can be sent. If we compare the average delay to what we observed in connection with UDP. 15. We have an average delay of 20 ms and 240 ms for the Indication and Request/Response exchanges. To further show the negative effect of combining MIH retransmissions with TCP’s reliable delivery. Thus the delay shown in the figure is associated only with the minority of successful exchanges. 18 we show the average Request/Response completion time when MIH retransmission are not being used over TCP. The delay curves also show strong agreement between the simulation results and the mathematical model’s prediction from Equation (9) and Equation (10). In addition. 16. 16. The MIH timeout TMIH is set to 3RTOmax .25 times greater than it is in the case when only a Indication message is being transmitted. especially for small values of p. the Responses received are always considered valid although that might not be the case in real implementations. using MIH ACKs over TCP further increases the average delay. 15 we show simulation results showing the mean message exchange time for Indications when TCP is used in addition to MIH ACKs and timeouts. Mean MIH Indication transaction completion time over TCP transport with MIH timeouts and acknowledgments. Theoretical and simulation values of mean MIH handover time for MIH handovers where M = 1 (Indication) and M = 2 (Request/Response) over UDP. in this case. spond to the scenarios in which MIH Indications and MIH Request/Response exchanges take place without the use of MIH ACKs. In our scenario. respectively. This is because the delays at this level of packet loss are so large and most Request/Response transactions do not complete within the ACK timeout interval. largely because the TCP transmission queue is no longer being filled with extra copies of messages that were lost during transmission attempts. 17. In Fig. using TCP results in a much lower probability that a given exchange will fail. In Fig. As we expect. 15. We also show simulation results showing the mean transaction completion time for Request/Response exchanges when TCP is used in addition to MIH ACKs and timeouts in Fig. TCP transport without MIH reliability features in Fig. If no MIH ACK is used. Mean MIH Request/Response transaction completion time over TCP transport with MIH timeouts and acknowledgments. In both cases. The figure does not exhibit the type of severe performance degradation that we observed in Fig. are very small. as . However. the only variance in the exchange completion time comes from random delays in the network which. When MIH ACKs are used and the average TCP roundtrip delay is greater than the MIH retransmission timeout interval. we plot the average Indication completion time for the case where we use Fig. MIH will create duplicate packets that go into the TCP queue. The figure shows much greater sensitivity to the packet loss probability as the transaction delay increases exponentially with respect to the value of TRTO . Fig.

we examine the amount of overhead associated with the various transport layer options. C ONCLUSIONS In this paper. When there is no packet loss (p = 0). if the TCP delay is too large. If MIH reliability is not in use. Fig. In Fig. as seen in Fig. 20.5 since there are at most 3 Requests. MIH Overhead In this subsection. 21. In the case of Request/Response exchanges. 17. If retransmissions by the MIH layer are turned off. which will in turn create another Request. If we examine both graphs. The maximum number of packets for each MIH Indication is therefore 6 (3 copies of the message and 3 corresponding MIH ACKs). As the packet loss rate increases.1. With MIH acknowledgments and retransmissions. Note that in all four figures we do not begin to see serious increases in the delay for either type of message exchange until the packet loss probability is on the order of 0. Mean MIH Request/Response transaction completion time over TCP transport without MIH timeouts and acknowledgments. 20 and Fig. which shows that increasing M . an MIH ACK message gets sent for each indication/request/response that a node receives. V. the MIH User will retransmit messages up to two times. Fig.25. there is no overhead penalty although this will result in a lower success rate for both types of exchanges. we developed characteristic functions that let us obtain the distribution and moments of the time to exchange an arbitrary number of MIH messages using either UDP with MIH retransmissions and acknowledgments or using . Requests will arrive late at the MN. we observe that the amount of overhead increases as we decrease the maximum RTO. 3 ACKs. Theoretical and simulation values of MIH overhead for Indication and Request/Response exchanges over UDP. Also.5 are an indication that a greater number of transactions are able to successfully complete within the time limit than we observed when we used MIH reliability features in conjunction with TCP. 18.14 Fig. and 3 Responses for each Request/Response pair. 19. significant increases in overhead to not occur until the packet loss probability exceeds 0. 19 we plot theoretical and simulation results for transport over UDP for both the Indication and Request/Response exchanges. This follows from Equation (19). we observed a slightly higher overhead penalty in connection with Indication exchanges relative to the Request/Response case. decreases N MSG . with N MSG → n/(1 − q) as M becomes large. The maximum overhead is 4. Therefore the minimum number of MIH packets sent per message is two if MIH ACKs are being used. the number of messages in an exchange. we plot the expected number of MIH packets sent to the transport layer per MIH message generated by the MIH User for Indications and Request/Response exchanges. whose Response will be ignored by the PoS. and a maximum error of 7% for the Request/Response exchange when p = 0. We observed very close agreement between the theoretical and simulation results for the Indication exchange. shown in Fig. which is what we would expect. the large average delay values associated with packet loss rates of near 0. respectively. this number is 1. 17. Next we examine overhead associated with using TCP transport. B. Mean MIH Indication transaction completion time over TCP transport without MIH timeouts and acknowledgments. Furthermore. In Fig.1.

We also note that TCP is best suited for longer MIH message exchanges since it will hold data until there is enough to fill a segment. Towsley.” T. [12] N. Krusoe. 35-42. “Modeling TCP latency.. Jun 2007. C. Hong. 35.” draft-rahman-mipshop-mih-transport-03. Aug 1986. [2] M.21/D06.shtml. Sidi and D. [8] “Mobility Services Framework Design. and J. 5–88.gov/seamlessandsecure. IETF Internet Draft. and Whitt. [5] H. [3] D. “Traffic Model and Performance Analysis for Cellular Mobile Radio Telephone Systems with Prioritized and Nonprioritized Handoff Procedures. pp. Vol. 1742–1751. Our results verify that there is a tradeoff between latency and overhead. IETF Internet Draft. Jul 2007. NS-2 Mobility Package. I. V.. pp. Arkko.” IEEE LAN/MAN Draft IEEE P802. Firoiu. W.. whether in the MIH User if UDP is being used or at the transport layer if TCP is being used. Daehyoung and S. Mellia. [9] Abate. Our simulation results agreed with our mathematical model.” R. Feb 2008. The delay and overhead penalties are of course smaller if the packet loss probability is small. Ed. and J.” IEEE Transactions on Vehicular Technology. work in progress. Jul 2007. Apr 2000.1 or greater. K. Padhye.nist. no. Simulation values of MIH overhead for Request/Response exchanges over TCP with MIH ACKs. u˜ R EFERENCES Fig. 3. Rappaport. Zhang. 2. Koodli. Vol. No. Adjusting timeout values to reduce latency. “Transport of Media Independent Handover Messages over IP. 21. but we also experienced a higher rate of . 133–145. “Hierarchical Mobile IPv6 Mobility Management (HMIPv6). which is undesirable for messages that require low latency. Simulation values of MIH overhead for Indication messages over TCP with MIH ACKs. 77-92. and H.antd. pp. Mar 1996. especially Juan Carlos Z´ niga and Telemaco Melia. [6] “Draft IEEE Standard for Local and Metropolitan Area Networks: Media Independent Handover Services. Vol. [14] National Institute of Standards and Technology.15 failure for a MIH message exchange. Watfa. 20. Mar 2008. U. 1992. Feb 2002. D.” IETF RFC 3775. TCP. invariably results in an increase in the expected number of messages that are transmitted during an exchange. Ed. Castelluccia. J. 10.S.W. and L. Ed. TCP’s shortcoming was that it resulted in greater overhead than UDP and performed inefficiently when the MIH message generation rate was low. This illustrates the importance of proactive handover signaling that begins well before the signal quality on the wireless link degrades to the point where the packet loss probability is on the order of 0. “Modeling TCP Throughput: A Simple Model and its Empirical Validation. [7] “Mobility Services Transport: Problem Statement. Savage.3. pp. ”The Fourier-series method for inverting transforms of probability distributions. “New Call Blocking versus Handoff Blocking in Cellular Networks. [1] D. Our results show that using MIH retransmission timers with TCP produces conflicts that severely degrade performance. M.C. We observed low delays and less overhead in the case where we used UDP. We also developed expressions for the average overhead expressed as extra message transmissions due to packet losses. ACKNOWLEDGMENT The authors would like to thank the members of the IETF MIPSHOP MIH design team for many helpful discussions. Stoica. Aug 2005. pp.” IEEE Communications Letters. [4] “Fast Handovers for Mobile IPv6. [13] M.” IEEE INFOCOM 96. [10] A. “TCP Model for Short Lived Flows. 8. draft-ietf-mipshopmstp-solution-01. Z´ niga. Fig. TCP displayed a lower exchange failure rate than UDP with retransmissions. El Malki. No. Melia.” INFOCOM 2000: Proceedings of the Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies .” IEEE/ACM Transactions on Networking.” IETF RFC 4140. “Mobility Support in IPv6. [11] J. pp.. Cardwell. Vol. Vol. Kim. http://www. Jul 2005. C. Bellier. Perkins. S. Rahman. T.” T. and u˜ H. 85–87. Johnson. work in progress. 6. Olvera-Hernandez. Anderson. 26-30 Mar 2000. IETF RFC 5164. Jun 2004. Melia. Starobinski.” Queuing Systems. IETF RFC 4068. 2. Soliman.00. J. largely because the number of retransmissions over UDP was limited while the number of TCP retransmissions was not.

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->