Eective Capacity-Based Quality of Service Measures

for Wireless Networks

Dapeng Wu

Rohit Negi

Abstract
An important objective of next-generation wireless networks is to provide quality of service
(QoS) guarantees. This requires a simple and ecient wireless channel model that can easily translate into connection-level QoS measures such as data rate, delay and delay-violation probability. To
achieve this, in [8], we developed a link-layer channel model termed eective capacity, for the setting
of a single hop, constant-bit-rate arrivals, uid trac, and wireless channels with negligible propagation delay. In this paper, we apply the eective capacity technique to deriving QoS measures
for more general situations, namely, 1) networks with multiple wireless links, 2) variable-bit-rate
sources, 3) packetized trac, and 4) wireless channels with non-negligible propagation delay.
Key Words: Wireless channel model, QoS, delay, eective capacity, large deviations theory.

Please direct all correspondence to Dapeng Wu, University of Florida, Dept. of Electrical & Computer

Engineering, P.O.Box 116130, Gainesville, FL 32611, USA. Tel. (352) 392-4954, Fax (352) 392-0044, Email:
[email protected]. URL: http://www.wu.ece.ufl.edu.

Carnegie Mellon University, Dept. of Electrical & Computer Engineering, 5000 Forbes Avenue, Pittsburgh, PA 15213, USA. Tel. (412) 268-6264, Fax (412) 268-2860, Email: [email protected].
http://www.ece.cmu.edu/~negi.

URL:

Data
source

Link-layer channel
Transmitter

Data
sink
Instantanteous channel capacity
log(1+SNR)
Receiver

Network
access device

Network
access device

Channel
encoder

Channel
decoder

Modulator

Demodulator

Physical-layer channel

Received
SNR

Wireless
channel

Figure 1: A wireless communication system.

Introduction

Providing QoS guarantees is crucial in the development of next-generation packet-based

wireless communication networks [4]. To support QoS guarantees, QoS provisioning mechanisms are required. A major problem in designing QoS provisioning mechanisms is the high
complexity in characterizing the relation between the control parameters of QoS provisioning mechanisms, and the calculated QoS measures, based on existing channel models, i.e.,
physical-layer channel models (see Fig. 1). This is because the physical-layer channel models
(e.g., Rayleigh fading model with a specied Doppler spectrum) do not explicitly characterize a wireless channel in terms of the link-level QoS metrics specied by users, such as data
rate, delay and delay-violation probability. To use the physical-layer channel models for QoS
support, we rst need to estimate the parameters for the channel model, and then extract
the link-level QoS metrics from the model. This two-step approach is obviously complex,
and may lead to inaccuracies due to possible approximations in extracting QoS metrics from
the models.
Recognizing that the limitation of physical-layer channel models in QoS support, is the
1

diculty in analyzing queues using them, in [8], we proposed moving the channel model up
the protocol stack, from the physical-layer to the link-layer. We call the resulting model an
eective capacity (EC) channel model [8], because it captures a generalized link-level capacity
notion of the fading channel. Figure 1 illustrates the dierence between the conventional
physical-layer channel and the link-layer channel. In [8], we presented the EC channel model
under the setting of a single hop, constant-bit-rate arrivals, uid trac, and wireless channels
with negligible propagation delay; in this paper, we use the eective capacity technique to
derive QoS measures for more general situations, namely, 1) networks with multiple wireless
links, 2) variable-bit-rate sources, 3) packetized trac, and 4) wireless channels with nonnegligible propagation delay.
The remainder of this paper is organized as follows. In Section 2, we present preliminary
results to familiarize the reader with the eective capacity technique. Sections 3 to 6 present
eective capacity-based QoS measures for networks with multiple wireless links, variable-bitrate sources, packetized trac, and wireless channels with non-negligible propagation delay,
respectively. Section 7 concludes the paper.

Preliminaries

We rst formally dene statistical QoS, which characterizes the requirement of a user. First,
consider a single-hop system, where the user is allotted a single time varying channel. Assume
that the user source has a xed rate rs and a specied delay bound Dmax , and requires that
the delay-bound violation probability is not greater than a certain value , that is,
P r{D() > Dmax } ,

(1)

where D() is the steady-state delay experienced by a ow, and P r{D() > Dmax } is the
probability of D() exceeding a delay bound Dmax . Then, we say that the user is specied
by the (statistical) QoS triplet {rs , Dmax , }. Even for this simple case, it is not immediately
obvious as to which QoS triplets are feasible, for the given channel, since a rather complex
queueing system (with an arbitrary channel capacity process) will need to be analyzed. The
key contribution of [8] was to introduce a concept of statistical delay-constrained capacity
2

Instantaneous Channel Capacity

Date Rate

Dmax

r(t)
Buffer
Data
sink

Data
source
D(t)

Figure 2: A queueing system model.

termed eective capacity, which allows us to obtain a simple and ecient test, to check the
feasibility of QoS triplets for a single time-varying channel. That paper did not deal with
general situations, e.g., networks with multiple wireless links and multi-hops, variable-bitrate sources, packetized trac, and wireless channels with non-negligible propagation delay,
which we consider in this paper.
Next, we briey explain the concept of eective capacity, and refer the reader to [8] for
details.
Let r(t) be the instantaneous channel capacity at time t. Assume that, the asymptotic
log-moment generation function of r(t)

t
1
log E[eu 0 r( )d ]
t t

(u) = lim

(2)

exists for all u 0. Then, the eective capacity function of r(t) is dened as
(u) =

(u)
u

, u > 0.

(3)

That is,

t
1
log E[eu 0 r( )d ], u > 0.
t ut

(u) = lim

(4)

Consider a queue of innite buer size supplied by a data source of constant data rate
(see Fig. 2). It can be shown [8] that if (u) indeed exists (e.g., for ergodic, stationary,
Markovian r(t)), then the probability of D() exceeding a delay bound Dmax satises
P r{D() > Dmax } e()Dmax ,
3

(5)

where the function () of source rate depends only on the channel capacity process r(t).
() can be considered as a channel model that models the channel at the link layer (in
contrast to physical layer models specied by Markov processes, or Doppler spectra). The
approximation (5) is accurate for large Dmax .
In terms of the eective capacity function (4) dened earlier, the QoS exponent function
() can be written as [8]
() = 1 ()

(6)

where 1 () is the inverse function of (u). Once () has been measured for a given
channel, it can be used to check the feasibility of QoS triplets. Specically, a QoS triplet
.
{rs , Dmax , } is feasible if (rs ) , where = log /Dmax . Thus, we can use the eective
capacity model (u) (or equivalently, the function () via (6)) to relate the channel capacity
process r(t) to statistical QoS. Since our eective capacity method predicts an exponential
dependence (5) between and Dmax , we can henceforth consider the QoS pair {rs , } to be
equivalent to the QoS triplet {rs , Dmax , }, with the understanding that = log /Dmax .
In the following sections, we extend the eective capacity technique to more general
situations. The following property is needed in the propositions in the rest of this paper.
Property 1 (i) The asymptotic log-moment generation function (u) dened in (2) is nite
for all u R. (ii) (u) is dierentiable for all u R.

QoS Measures for Wireless Networks

In this section, we consider two basic network structures for wireless networks: one with
only tandem wireless links (see Figure 3) and the other with only parallel wireless links (see
Figure 4). In the following, Propositions 1 and 2 give QoS measures for these two network
structures, respectively.
Denote rk (t) (k = 1, , K) the instantaneous capacity of channel k at time t. For a

Node 1

Node 2

Data
source
Rate =

Wireless

Wireless

channel 1

channel 2

Q1

Data
sink

Q2

Figure 3: A network with tandem wireless links.

0 , t), for t 0 and any t0 [0, t], by
network with K tandem links, dene the service S(t
0 , t) =
S(t

K 


inf

t0 t1 tK1 tK =t

k=1

tk

rk ( )d

(7)

tk1

and the asymptotic log-moment generating function

1

log E[euS(0,t) ]
t t

tandem (u) = lim

(8)

t) is dened by (7); also dene the eective capacity of channel k by

where S(0,

t
1
log E[eu 0 rk ( )d ], u > 0.
t ut

k (u) = lim

(9)

Proposition 1 Assume that the log-moment generating function tandem (u) dened by (8)
satises Property 1. Given the eective capacity functions {k (u), k = 1, , K} of K
tandem links and an external arrival process with constant rate , the end-to-end delay
D() experienced by the trac traversing the K tandem links satises
lim sup

Dmax

1
Dmax

log P r{D() > Dmax } ,

if (/) > ,

(10)

where ( /) = ,

(11)

and
lim

Dmax

1
Dmax

log P r{D() > Dmax } = ,

where (u) = tandem (u)/u. Moreover, the eective capacity (u) satises
(u) min k (u).
k

(12)

Wireless
channel 1

Node

Data
source

Data
sink

Rate =

Wireless
channel K

Figure 4: A network with parallel wireless links.

For a proof of Proposition 1, see the Appendix. Note that the capacity processes of the
tandem channels are not required to be independent in Proposition 1.
Proposition 2 Assume that the log-moment generating function k (u) of each channel k in
the network satises Property 1. Given the eective capacity functions {k (u), k = 1, , K}
of K independent parallel links and an external arrival process with constant rate , the endto-end delay D() experienced by the trac traversing the K parallel links satises
lim sup

Dmax

1
Dmax

log P r{D() > Dmax } ,

if (/) > ,

(13)

where ( /) = ,

(14)

and
1

lim

Dmax

where (u) =

Dmax

log P r{D() > Dmax } = ,

K

k=1 k (u).

For a proof of Proposition 2, see the Appendix.

Propositions 1 and 2 suggest the following approximation
P r{D() > Dmax } e

max

(15)

for large Dmax . In addition, (u) specied in Propositions 1 and 2 can be regarded as the
eective capacity of the equivalent channel of the network, which consists of tandem links
6

only or independent parallel links only. In Sections 4 to 6, we will use (u) to characterize
the equivalent channel of the network; and we will use (14) only since (14) is tighter than
(13).

QoS Measures for Variable-Bit-Rate Sources

In this section, we develop QoS measures for the case where the sources generate trac at
variable bit-rates (VBR). We consider two classes of VBR sources: leaky-bucket constrained
arrival [2][9, page 15] and exponential process with its eective bandwidth function known
[1][9, page 16]. Propositions 3 and 4 provide QoS measures for these two classes of VBR
sources, respectively.
Proposition 3 Assume that a wireless network consists of tandem links only or independent
parallel links only; the eective capacity function of the equivalent channel of the wireless
network is characterized by (u); and the log-moment generating function k (u) of each
channel k in the network satises Property 1. Given an external arrival process constrained
(s)

by a leaky bucket with bucket size (s) and token generating rate s , the end-to-end delay
D() experienced by the trac traversing the network satises
lim

Dmax

1
Dmax

(s)

(s)
(s) /s

(s)

log P r{D() > Dmax } = , where ( /s ) = s .

(16)

For a proof of Proposition 3, see the Appendix. Eq. (16) suggests the following approximation
P r{D() > Dmax } e

(D

max

(s) /(s) )
s

(17)

for large Dmax .

For convenience, we replicate the denition of the eective bandwidth [1] here. Consider
an arrival process {A(t), t 0} where A(t) represents the amount of source data (in bits)
over the time interval [0, t). Assume that the asymptotic log-moment generating function

of a stationary process A(t), dened as

1
log E[euA(t) ],
t t

(u) = lim

(18)

exists for all u 0. Then, the eective bandwidth function of A(t) is dened as
(s) (u) =

(u)
u

, u > 0.

(19)

Proposition 4 Assume that a wireless network consists of tandem links only or independent
parallel links only; the eective capacity function of the equivalent channel of the wireless
network is characterized by (u); an external arrival process is characterized by its eective
bandwidth function (s) (u); and the log-moment generating function k (u) of each channel k
in the network and the log-moment generating function (s) (u) of the external arrival process
satisfy Property 1. Denote u the unique solution of the following equation
(s) (u) = (u).

(20)

The end-to-end delay D() experienced by the trac traversing the network satises
lim

Dmax

1
Dmax

log P r{D() > Dmax } = , where = u (s) (u ).

(21)

For a proof of Proposition 4, see the Appendix. Note that a single-link network is a special
case in Propositions 3 and 4.

QoS Measures for Packetized Trac

In previous sections, we assumed uid trac. In this section, we extend the QoS measures
obtained previously for the uid model to the case with packetized trac. This is important
since in practical situations, the packet size is not negligible (not innitesimal as in uid
model).

We assume the propagation delay of a wireless link is negligible, and the service at
a network node is non-cut-through, i.e., no packet is eligible for service until its last bit
has arrived. We also assume a wireless network consists of tandem links only or parallel
links only. For a network with tandem links only, the number of hops in the network is
determined by the number of tandem links in the network; for a network with parallel links
only, the number of hops in the network is one. We consider two cases: 1) a constant-bit-rate
source with constant packet size, and 2) a variable-bit-rate source with variable packet size.
Propositions 5 and 6 give QoS measures for these two cases, respectively.
Proposition 5 Assume that a wireless network consists of tandem links only or independent
parallel links only; the eective capacity function of the equivalent channel of the wireless
network is characterized by (u); the log-moment generating function k (u) of each channel
k in the network satises Property 1; and the network consists of N hops. Given an external
arrival process with constant bit rate and constant packet size Lc , the end-to-end delay
D() experienced by the trac traversing the network satises
1
log P r{D() > Dmax } = , where ( /) = .
Dmax Dmax N Lc /
lim

(22)

For a proof of Proposition 5, see the Appendix. Eq. (22) suggests the following approximation
P r{D() > Dmax } e

(D

max N Lc /)

(23)

for large Dmax .

Proposition 6 Assume that a wireless network consists of tandem links only or independent
parallel links only; the eective capacity function of the equivalent channel of the wireless
network is characterized by (u); the log-moment generating function k (u) of each channel
k in the network satises Property 1; and the network consists of N hops. Given a trac
ow having maximum packet size Lmax and constrained by a leaky bucket with bucket size
(s)

(s) and token generating rate s , the end-to-end delay D() experienced by the trac
traversing the network satises
lim

Dmax

1
Dmax N

(s)
Lmax /s

(s)
(s) /s

log P r{D() > Dmax } = ,

(24)

(s)

(s)

where ( /s ) = s .
For a proof of Proposition 6, see the Appendix. Eq. (22) suggests the following approximation
P r{D() > Dmax } e

(D

(s)
(s) /(s) )
max N Lmax /s
s

(25)

for large Dmax . Note that a single-link network is a special case in Propositions 5 and 6.

QoS Measures for Wireless Channels with

Non-negligible Propagation Delay

In previous sections, we assumed the propagation delay of a wireless link is negligible. In

this section, we extend the QoS measures obtained previously to the situation where the
propagation delay of a wireless link is not negligible. We consider two cases: 1) a uid
source with a constant rate, and 2) a variable-bit-rate source with variable packet size.
Propositions 7 and 8 give QoS measures for these two cases, respectively.
Proposition 7 Assume that a wireless network consists of tandem links only or independent
parallel links only; the eective capacity function of the equivalent channel of the wireless
network is characterized by (u); the log-moment generating function k (u) of each channel
k in the network satises Property 1; the network consists of N hops; and the i-th hop
(i = 1, , N) incurs a constant propagation delay di . Given a uid trac ow with constant
rate , the end-to-end delay D() experienced by the trac traversing the network satises
1
log P r{D() > Dmax } = , where ( /) = .
N
Dmax D

d
max
i=1 i
lim

(26)

For a proof of Proposition 7, see the Appendix. Eq. (26) suggests the following approximation
P r{D() > Dmax } e
for large Dmax .
10

(D

max

N

i=1

di )

(27)

Proposition 8 Assume that a wireless network consists of tandem links only or independent
parallel links only; the eective capacity function of the equivalent channel of the wireless
network is characterized by (u); the log-moment generating function k (u) of each channel
k in the network satises Property 1; the network consists of N hops; and the i-th hop
(i = 1, , N) incurs a constant propagation delay di . Given a trac ow having maximum
packet size Lmax and constrained by a leaky bucket with bucket size (s) and token generating
(s)

rate s , the end-to-end delay D() experienced by the trac traversing the network satises
lim

Dmax

1
Dmax N
(s)

(s)
Lmax /s

(s)
(s) /s

N
i=1

di

log P r{D() > Dmax } = , (28)

(s)

where ( /s ) = s .
For a proof of Proposition 8, see the Appendix. Eq. (28) suggests the following approximation
P r{D() > Dmax } e

(D

N
(s)
(s) /(s)
max N Lmax /s
s
i=1

di )

(29)

for large Dmax .

Concluding Remarks

The design of QoS provisioning mechanisms in wireless networks calls for a simple and
eective wireless channel model. In [8], we proposed and developed such a simple and
eective channel model, called eective capacity, for the setting of a single hop, constantbit-rate arrivals, uid trac, and wireless channels with negligible propagation delay. In this
paper, we employed the eective capacity technique to derive QoS measures for more general
situations, i.e., networks with multiple wireless links, variable-bit-rate sources, packetized
trac, and wireless channels with non-negligible propagation delay.
In our future work, the QoS measures developed in this paper will be used to design
ecient mechanisms to provide end-to-end QoS guarantees in a multihop wireless network.
This will involve developing algorithms for QoS routing, resource reservation, admission
control and scheduling.
11

Acknowledgment
This work was supported by the National Science Foundation under the grant ANI-0111818.

Appendix
Proof of Proposition 1
Denote Qk (t) the queue length at time t at node k (k = 1, , K), Q(t) the end-to-end queue
length at time t, Q() the steady state of the end-to-end queue length, A(t0 , t) the amount

t
of arrival to node 1 (see Figure 3) over the time interval [t0 , t]. Dene Sk (t0 , t) = t0 rk ( )d ,
which is the service provided by channel k over the time interval [t0 , t].
We rst prove an upper bound. It can be proved [10, page 81] that

Q(t) =

K

k=1

Qk (t) = sup

0t0 t



0 , t)
A(t0 , t) S(t

(30)

0 , t) is dened by (7). Without loss of generality, we consider the discrete time case
where S(t
only, i.e., t N, where N is the set of natural numbers. From (30) and Loynes Theorem [6],
we obtain





Q() = sup A(0, t) S(0, t) = sup t S(0, t)

tN

tN

12

(31)

Then, we have




P r {Q() > q} = P r sup t S(0, t) > q



(a)

Pr

tN



t) > q
t S(0,

(32)


(33)

tN

(b)



t) > q
P r t S(0,

(34)

tN

(c)

euq E[eu(tS(0,t)) ]

(35)

tN







t) > q 
t) > q , (b) is due
where (a) since the event suptN t S(0,
t

S(0,
tN
to the union bound, and (c) from the Cherno bound. Since (u) = tandem (u)/u, we
have
1

log E[euS(0,t) ], u > 0,

t ut

(u) = lim

(36)

Hence, for any  > 0, there exists a number t > 0 such that for t t, we have

E[euS(0,t) ] eu((u)+)t ,

u > 0.

(37)

If +  < (u), we have



uq

u(tS(0,t))

E[e

(a)

uq u((u)+)t

tt

tN

euq E[eu(tS(0,t)) ]

(38)

t=1

(b)

t1


euq

u((u)+)t

e
+
1 eu((u)+)

t1


E[eu(tS(0,t)) ]

(39)

t=1

where (a) from (37), and (b) from geometric sum. From (35) and (39), we have
P r {Q() > q} euq ,

13

if +  < (u),

(40)

where is a constant independent of q. Hence, we obtain

1
log P r {Q() > q} u,
q

lim sup
q

if +  < (u).

(41)

if < (u).

(42)

Letting  0, we have
lim sup
q

1
log P r {Q() > q} u,
q

Since Q(t) = D(t), t 0, (42) results in

lim sup

Dmax

1
Dmax

log P r{D() > Dmax } u,

if < (u).

(43)

Let = u. Then (43) becomes (10).

Next we prove a lower bound. Let q = t where > 0. Then, we have
lim inf
q

1
1
log P r {Q() > q} = lim inf
log P r {Q() > t}
t t
q



1

log P r sup t S(0, t) > t

= lim inf
t t
tN


1
t) > t
lim inf
log P r t S(0,
t t


t)
S(0,
1
log P r
>
= lim inf
t t
t
(a)

1
inf (x)
x>

(44)
(45)
(46)

(47)

(48)

where (a) from Gartner-Ellis Theorem [1] since tandem (u) satises Property 1, and the
Legendre-Fenchel transform (x) of tandem (u) is dened by
(x) = sup{u x tandem (u)}.
uR

14

(49)

Since (48) holds for any > 0, we have

lim inf
q

1
1
log P r {Q() > q} sup
inf (x)
x>
q

>0
(y)
y> y +

= inf

(50)

(51)

It can be proved [3] that

(y)
= u ,
y> y +

where tandem (u ) = u .

inf

(52)

From the denition of eective capacity (u) in (36), tandem (u ) = u implies

(u ) = . Then, applying (52) to (51) leads to
lim inf
q

1
log P r {Q() > q} u ,
q

where (u ) = .

(53)

Due to the continuity of (u), we have

lim (u) =

(54)

uu

Hence, letting u u , (42) and (53) result in

u lim inf
q

1
1
log P r {Q() > q} lim sup log P r {Q() > q} u
q
q q

(55)

Hence, we have
1
log P r {Q() > q} = u ,
q q
lim

where (u) = .

(56)

Since Q(t) = D(t), t 0, (56) results in

lim

Dmax

1
Dmax

log P r{D() > Dmax } = u,

Let = u . Then (57) becomes (11).

15

where (u) = .

(57)

From (7), it is obvious that

0 , t) min Sk (t0 , t).
S(t
k

(58)

Then we have for u > 0,

1

log E[euS(0,t) ]
t ut

(59)

log E[eu mink Sk (0,t) ]

t ut

(60)

min lim

log E[euSk (0,t) ]

t ut

(61)

= min k (u)

(62)

(u) = lim
(a)

lim

where (a) from (58). This completes the proof.

Proof of Proposition 2
Denote rk (t) (k = 1, , K) channel capacity of link k at time t. From Figure 4, it is clear
that the network has only one queue and multiple servers, each of which corresponds to a

wireless link. Since the total instantaneous channel capacity r(t) = K
k=1 rk (t), the eective
capacity function for the aggregate parallel links is
(a)

t
1
log E[eu 0 r( )d ]
t ut

(u) = lim

t K
1
log E[eu 0 k=1 rk ( )d ]
t ut

= lim

t
1 
= lim
log E[eu 0 rk ( )d ]
t ut
k=1
K

(b)

(c)

K


k (u)

(63)

k=1

where (a) from (4), (b) since {rk (t), k = 1, , K} are independent, and (c) from (9).
16

Given the eective capacity (u), we can prove (13) and (14) with the same technique
used in proving (10) and (11). This completes the proof.

Proof of Proposition 3
(s)

For the trac of constant rate s , denote Q()

the steady state of the end-to-end queue

length and D()

the end-to-end delay. Using the result in [5, page 30]), we can show

D() D()
(s) /(s)
s ,

(64)

Note that D() is the end-to-end delay experienced by the trac constrained by a leaky
(s)

bucket with bucket size (s) and token generating rate s . Hence, we have

D() D()
+ (s) /(s)
s ,

(65)

(s)

From (40) and Q()

= s D(),
we have



(s)

P r D()
> Dmax eus Dmax ,

(s)

if (u) > s ,

(66)

where is a constant independent of Dmax . Then, we have



(a)

> Dmax (s) /(s)

P r {D() > Dmax } = P r D()
s
(b)

(s)

eus

(s)

(Dmax (s) /s )

(67)

where (a) from (65), and (b) from (66). Hence, we have
lim sup

Dmax

1
Dmax

(s)
(s) /s

log P r{D() > Dmax } ,

(s)

(s)

if (/s ) > s .

(68)

Similar to the proof of Proposition 1, we can obtain a lower bound

lim inf

Dmax

1
Dmax

(s)
(s) /s

log P r{D() > Dmax } ,

Combining (68) and (69), we obtain (16). This completes the proof.
17

(s)

(s)

where ( /s ) = s . (69)

Proof of Proposition 4
The proof is similar to that of Proposition 1.
Denote Q(t) the end-to-end queue length at time t, Q() the steady state of the endto-end queue length, A(t0 , t) the amount of external arrival to the network over the time
interval [t0 , t]. From (30), we know

Q(t) = sup

0t0 t

0 , t)
A(t0 , t) S(t


(70)

0 , t) =
0 , t) is dened by (7) for the tandem links, and is dened by S(t
where S(t

t K
0

k=1 rk ( )d

for independent parallel links.

We rst prove an upper bound. Without loss of generality, we consider the discrete time
case only, i.e., t N, where N is the set of natural numbers. From (70) and Loynes Theorem
[6], we obtain


t)
Q() = sup A(0, t) S(0,


(71)

tN

Then, we have



t) > q
P r {Q() > q} = P r sup A(0, t) S(0,

Pr

tN




A(0, t) S(0, t) > q

(72)


(73)

tN




P r A(0, t) S(0, t) > q

(74)

tN

euq E[eu(A(0,t)S(0,t)) ]

(75)

tN

From the denition of eective capacity in (36), for any /2 > 0, there exists a number t > 0
such that for t t, we have

E[euS(0,t) ] eu((u)+/2)t ,
18

u > 0.

(76)

Similarly, from the denition of eective bandwidth in (19), for any /2 > 0, there exists a
number t > 0 such that for t t, we have
(s) (u)+/2)t

E[euA(0,t) ] eu(

u > 0.

(77)

Without loss of generality, here we choose the same t for both (76) and (77), since we can
always choose the maximum of the two to make (76) and (77) hold. Then, if (s) (u) +  <
(u), we have


uq

u(A(0,t)S(0,t))

E[e

(a)

uq u((s) (u)(u)+)t

tt

tN

t1


euq E[eu(A(0,t)S(0,t)) ]

(78)

t=1

euq

u((s) (u)(u)+)t

e
+
1 eu((s) (u)(u)+)

t1


E[eu(A(0,t)S(0,t)) ]

(79)

t=1

where (a) from (76) and (77). From (75) and (79), we have
lim sup
q

1
log P r {Q() > q} u, if (s) (u) +  < (u).
q

(80)

1
log P r {Q() > q} u, if (s) (u) < (u).
q

(81)

Letting  0, we have
lim sup
q

Similar to the proof of Proposition 1, we can obtain a lower bound

lim inf
q

1
log P r {Q() > q} u , where (s) (u ) = (u ).
q

(82)

Combining (81) and (82), we have

1
log P r {Q() > q} = u , where (s) (u ) = (u ).
q q
lim

Since Q() = (s) (u ) D(), (83) results in (21). This completes the proof.
19

(83)

Proof of Proposition 5
For the packetized trac, denote Qk (t) the queue length at time t at node k (k = 1, , N),
Q(t) the end-to-end queue length at time t, and Q() the steady state of the end-to-end
k (t)
queue length. Correspondingly, for the uid trac of constant arrival rate , denote Q
the end-to-end queue length at time t, Q()

the queue length at time t at node k, Q(t)

the

steady state of the end-to-end queue length, and D()

the end-to-end delay.
For each node k, we have the sample path relation as below [7]
k (t) Lc ,
Qk (t) Q

t 0.

(84)

Summing up over k, we obtain

N


k (t)] = Q(t) Q(t)

N Lc ,
[Qk (t) Q

t 0.

(85)

k=1

Hence, for the steady state, we have

Q() Q()
N Lc .

(86)

Since Q() = D() and Q()

= D(),
we have

D() D()
N Lc /.

(87)

Note that D() is the end-to-end delay experienced by the packetized trac with constant
bit rate and constant packet size Lc . Then, we can prove (22) in the same way as we prove
(16) in Proposition 3.

Proof of Proposition 6

Denote D()
the end-to-end delay experienced by the uid trac with constant arrival
(s)

rate s . Using the sample path relation in [5, page 35]), we obtain
(s)
(s)

D() D()
N Lmax /(s)
s + /s ,

20

(88)

Note that D() is the end-to-end delay experienced by the packetized trac having maximum packet size Lmax and constrained by a leaky bucket with bucket size (s) and token
(s)

generating rate s . Then, we can prove (24) in the same way as we prove (16) in Proposition 3.

Proof of Proposition 7

Denote D()
the end-to-end delay experienced by the uid trac with constant arrival rate
and without propagation delay. Using the sample path relation between the two cases
(with/without propagation delay), it is easy to show

D() D()

N


di ,

(89)

i=1

Then, we can prove (26) in the same way as we prove (16) in Proposition 3.

Proof of Proposition 8

Denote D()
the end-to-end delay experienced by the uid trac with constant arrival
(s)

rate s and without propagation delay. Using the sample path relation in [5, page 35]), we
obtain
(s)
(s)

D() D()
N Lmax /(s)
s + /s +

N


di ,

(90)

i=1

Then, we can prove (28) in the same way as we prove (16) in Proposition 3.

References
[1] C.-S. Chang, Performance guarantees in communication networks, Springer, 2000.
[2] R. L. Cruz, A calculus for network delay, Part I: network elements in isolation, IEEE Trans.
on Information Theory, vol. 37, no. 1, pp. 114131, Jan. 1991.

21

[3] G. de Veciana and J. Walrand, Eective bandwidths: call admission, trac policing and
ltering for ATM networks, Queuing Systems, vol. 20, pp. 3759, 1995.
[4] H. Holma and A. Toskala, WCDMA for UMTS: Radio Access for Third Generation Mobile
Communications, Wiley, 2000.
[5] J.-Y. Le Boudec and P. Thiran, Network calculus: a theory of deterministic queueing systems
for the Internet, Springer, 2001.
[6] R. M. Loynes, The stability of a queue with non-independent inter-arrivals and service times,
Proc. Camb. Phil. Soc., vol. 58, pp. 497520, 1962.
[7] A. K. Parekh and R. G. Gallager, A generalized processor sharing approach to ow control
in integrated services networks: the single node case, IEEE/ACM Trans. on Networking,
vol. 1, no. 3, pp. 344357, June 1993.
[8] D. Wu and R. Negi, Eective capacity: a wireless link model for support of quality of service,
IEEE Trans. on Wireless Communications, vol. 2, no. 4, pp. 630643, July 2003.
[9] D. Wu, Providing quality of service guarantees in wireless networks, Ph.D. Dissertation,
Dept. of Electrical & Computer Engineering, Carnegie Mellon University, Aug. 2003. Available
at http://www.wu.ece.ufl.edu/mypapers/Thesis.pdf.
[10] Z.-L. Zhang, End-to-end support for statistical quality-of-service guarantees in multimedia
networks, Ph.D. Dissertation, Department of Computer Science, University of Massachusetts,
Feb. 1997.

22