630 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO.

4, JULY 2003

Effective Capacity: A Wireless Link Model for

Support of Quality of Service
Dapeng Wu, Student Member, IEEE, and Rohit Negi, Member, IEEE

Abstract—To facilitate the efficient support of quality of service estimate the parameters for the channel model, and then extract
(QoS) in next-generation wireless networks, it is essential to model QoS metrics from the model. This two-step approach is obvi-
a wireless channel in terms of connection-level QoS metrics such as ously complex, and may lead to inaccuracies due to possible
data rate, delay, and delay-violation probability. However, the ex-
isting wireless channel models, i.e., physical-layer channel models, approximations in extracting QoS metrics from the models.
do not explicitly characterize a wireless channel in terms of these To address this issue, we propose and develop a link-layer
QoS metrics. In this paper, we propose and develop a link-layer channel model termed the effective capacity (EC) model. In
channel model termed effective capacity (EC). In this approach, we this approach, we first model a wireless link by two EC func-
first model a wireless link by two EC functions, namely, the prob- tions, namely, the probability of nonempty buffer, and the QoS
ability of nonempty buffer, and the QoS exponent of a connection.
Then, we propose a simple and efficient algorithm to estimate these exponent of the connection. Then, we propose a simple and
EC functions. The physical-layer analogs of these two link-layer efficient algorithm to estimate these EC functions. The phys-
EC functions are the marginal distribution (e.g., Rayleigh–Ricean ical-layer analogs of these two link-layer EC functions are the
distribution) and the Doppler spectrum, respectively. The key ad- marginal distribution (e.g., Rayleigh–Ricean distribution) and
vantages of the EC link-layer modeling and estimation are: 1) ease the Doppler spectrum, respectively. The key advantages of EC
of translation into QoS guarantees, such as delay bounds; 2) sim-
plicity of implementation; and 3) accuracy, and hence, efficiency link-layer modeling and estimation are: 1) ease of translation
in admission control and resource reservation. We illustrate the into QoS guarantees, such as delay bounds; 2) simplicity of im-
advantage of our approach with a set of simulation experiments, plementation; and 3) accuracy, and hence, efficiency in admis-
which show that the actual QoS metric is closely approximated by sion control and resource reservation. Simulation results show
the QoS metric predicted by the EC link-layer model, under a wide that the actual QoS metric is closely approximated by the esti-
range of conditions.
mated QoS metric obtained from our channel estimation algo-
Index Terms—Delay, Doppler spectrum, fading, quality of ser- rithm, under a wide range of conditions. This demonstrates the
vice (QoS), queueing theory, wireless channel model.
effectiveness of the EC link-layer model in guaranteeing QoS.
Conventional channel models directly characterize the
I. INTRODUCTION fluctuations in the amplitude of a radio signal. We call these
models physical-layer channel models, to distinguish them
T HE NEXT-GENERATION wireless networks such as the
third-generation (3G) and the fourth-generation wireless
systems are targeted at supporting diverse quality of service
from the link-layer channel model we propose. In this paper, we
consider a small-scale fading model [12] for the physical-layer
channel. Small-scale fading models describe the characteristics
(QoS) requirements and traffic characteristics [9]. The success
of generic radio paths in a statistical fashion. Small-scale fading
in the deployment of such networks will critically depend upon
refers to the dramatic changes in signal amplitude and phase
how efficiently the wireless networks can support traffic flows
that can be experienced as a result of small changes (as small as
with QoS guarantees [10]. To achieve this goal, mechanisms for
a half-wavelength) in the spatial separation between a receiver
guaranteeing QoS (e.g., admission control and resource reser-
and a transmitter. Small-scale fading can be slow or fast,
vation) need to be efficient and practical [6].
depending on the Doppler spread. The statistical time-varying
Efficient and practical mechanisms for QoS support require
nature of the envelope of a flat-fading signal is characterized
accurate and simple channel models [10]. Toward this end, it is
by distributions such as Rayleigh, Ricean, Nakagami, etc. [12].
essential to model a wireless channel in terms of QoS metrics
Physical-layer channel models provide a quick estimate of
such as data rate, delay, and delay-violation probability. How-
the physical-layer performance of wireless communications
ever, the existing channel models (e.g., Rayleigh-fading model
systems [e.g., symbol error rate versus signal-to-noise ratio
with a specified Doppler spectrum) do not explicitly charac-
(SNR)]. However, physical-layer channel models cannot be
terize a wireless channel in terms of these QoS metrics. To use
easily translated into complex link-layer QoS guarantees for a
the existing channel models for QoS support, we first need to
connection, such as bounds on delay. The reason is that these
complex QoS requirements need an analysis of the queueing
Manuscript received March 30, 2001; revised January 2, 2002 and June 3, behavior of the connection, which is hard to extract from
2002; accepted June 4, 2002.The editor coordinating the review of this paper
and approving it for publication is Y. Fang. This work was supported in part by physical-layer models. Thus, it is hard to use physical-layer
the National Science Foundation under Grant ANI-0111818. models in QoS support mechanisms, such as admission control
The authors are with the Department of Electrical and Computer Engi- and resource reservation.
neering, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail:
[email protected]; [email protected]). Recognizing that the limitation of physical-layer channel
Digital Object Identifier 10.1109/TWC.2003.814353 models in QoS support is the difficulty in analyzing queues
1536-1276/03$17.00 © 2003 IEEE
WU AND NEGI: EFFECTIVE CAPACITY: A WIRELESS LINK MODEL FOR SUPPORT OF QoS 631

II. MOTIVATION FOR USING LINK-LAYER CHANNEL MODELS

Physical-layer channel models have been extremely suc-
cessful in wireless transmitter–receiver design, since they can
be used to predict physical-layer performance characteristics
such as bit/frame error rates as a function of SNR. These are
very useful for circuit switched applications, such as cellular
telephony. However, future wireless systems will need to handle
increasingly diverse multimedia traffic, which are expected to
be primarily packet switched. For example, the new wideband
code-division multiple-access (WCDMA) specifications make
explicit provisions for 3G networks to evolve over time, from
circuit switching to packet switching. The key difference be-
tween circuit switching and packet switching, from a link-layer
design viewpoint, is that packet switching requires queueing
analysis of the link. Thus, it becomes important to characterize
the effect of the data traffic pattern, as well as the channel
behavior, on the performance of the communication system.
QoS guarantees have been heavily researched in the wired
Fig. 1. Packet-based wireless communication system. networks (e.g., asynchronous transfer mode (ATM) and Internet
protocol (IP) networks). These guarantees rely on the queueing
using them, we propose moving the channel model up the model shown in Fig. 2. This figure shows that the source traffic
protocol stack from the physical layer to the link layer. We and the network service are matched using a first-in first-out
call the resulting model an EC link model because it captures buffer (queue). Thus, the queue prevents loss of packets that
a generalized link-level capacity notion of the fading channel. could occur when the source rate is more than the service rate,
Fig. 1 illustrates the difference between the conventional at the expense of increasing the delay. Queueing analysis, which
physical-layer and our proposed link-layer model.1 For sim- is needed to design appropriate admission control and resource
plicity, we interchange “physical-layer channel” with “physical reservation algorithms [1], [13], requires source traffic charac-
channel” and interchange “link-layer channel” with “link” in terization and service characterization. The most widely used
the rest of the paper. approach for traffic characterization is to require that the amount
To summarize, the EC link model that we propose aims to of data (i.e., bits as a function of time ) produced by a source
characterize wireless channels in terms of functions that can be conform to an upper bound, called the traffic envelope . The
easily mapped to link-level QoS metrics, such as delay-bound service characterization for guaranteed service is a guarantee
violation probability. Furthermore, we propose a novel channel of a minimum service (i.e., bits communicated as a function of
estimation algorithm that allows practical and accurate measure- time) level, specified by an SC [7]. Functions and
ments of the EC model functions. are specified in terms of certain traffic and service param-
The remainder of this paper is organized as follows. In Sec- eters, respectively. Examples include the UPC parameters used
tion II, we elaborate on the QoS guarantees that motivate us to in ATM [1] for traffic characterization, and the traffic specifi-
search for a link-layer model. We describe usage parameter con- cation (T-SPEC) and the service specification (R-SPEC) fields
trol (UPC) traffic characterization, and its dual, the service curve used with the resource reservation protocol [2], [7] in IP net-
(SC) network service characterization. We show that these con- works.
cepts, borrowed from networking literature, lead us to consider To elaborate on this point, a traffic envelope charac-
the EC model of wireless channels. In Section III, we formally terizes the source behavior in the following manner: over any
define the EC link model in terms of two functions: probability window of size , the amount of actual source traffic does
of nonempty buffer and QoS exponent. We then describe an es- not exceed (see Fig. 3). For example, the UPC parameters
timation algorithm, which accurately estimates these functions, specify by
with very low complexity. Section IV shows simulation results
that demonstrate the advantage of using the EC link model to ac- (1)
curately predict QoS, under a variety of conditions. This leads to
efficient admission control and resource reservation. Section V where is the peak data rate, the sustainable rate, and
concludes this paper and points out future research directions. the leaky-bucket size [7]. As shown in Fig. 3, the curve
Table I lists the notations used in this paper. consists of two segments; the first segment has a slope equal to
the peak source data rate , while the second segment, has a
slope equal to the sustainable rate , with . is
the axis intercept of the second segment. has the property
1In Fig. 1, we use Shannon’s channel capacity to represent the instantaneous
that for any time .
channel capacity. In practical situations, the instantaneous channel capacity is
log(1+ SNR=0 ), where 0 is determined by the modulation scheme and Just as upper bounds the source traffic, a network SC
the channel code used. lower bounds the actual service that a source will
632 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 4, JULY 2003

TABLE I
NOTATIONS

Fig. 2. Queueing system model.

receive. has the property that for any time .

Both and are negotiated during the admission control
and resource reservation phase. An example of a network SC is
the R-SPEC curve used for guaranteed service in IP networks

(2)
Fig. 3. Traffic and service characterization.
where , is the constant service rate, and
the delay error term (due to propagation delay, link sharing,
and so on). This curve is illustrated in Fig. 3. consists of denoted by , is the queue length built up at time , due to
two segments; the horizontal segment indicates that no packet is packets that have not been served yet.
being serviced due to propagation delay, etc., for a time interval In contrast to packet-switched wireline networks, providing
equal to the delay error term , while the second segment has QoS guarantees in packet-switched wireless networks is a chal-
a slope equal to the service rate . In the figure, we also ob- lenging problem. This is because wireless channels have low re-
serve that: 1) the horizontal difference between and , liability, and time-varying capacities, which may cause severe
denoted by , is the delay experienced by a packet arriving QoS violations. Unlike wireline links, which typically have a
at time ; and 2) the vertical difference between the two curves, constant capacity, the capacity of a wireless channel depends
WU AND NEGI: EFFECTIVE CAPACITY: A WIRELESS LINK MODEL FOR SUPPORT OF QoS 633

upon such random factors as multipath fading, cochannel inter- III. EFFECTIVE CAPACITY MODEL OF WIRELESS CHANNELS
ference, and noise disturbances. Consequently, providing QoS Section II argued that QoS guarantees can be achieved if a
guarantees over wireless channels requires accurate models of statistical SC can be calculated for the given wireless link. Thus,
their time-varying capacity, and effective utilization of these we need to calculate an SC such that, for a given ,
models for QoS support. the following probability bound on the channel service is
The simplicity of the SCs discussed earlier motivates us to satisfied:
define the time-varying capacity of a wireless channel as in
(2). Specifically, we hope to lower bound the channel service (3)
using two parameters: the channel sustainable rate , and the
maximum fade duration .2 However, physical-layer wire- Further, is restricted to being specified by the parame-
less channel models do not explicitly characterize the channel
ters , as below [(2), which we reproduce, for con-
in terms of such link-layer QoS metrics as data rate, delay, and
venience]
delay-violation probability. For this reason, we are forced to
look for alternative channel models. (4)
A tricky issue that surfaces is that a wireless channel has a
capacity that varies randomly with time. Thus, an attempt to Therefore, the statistical SC specification requires that we
provide a strict lower bound [i.e., the deterministic SC , relate its parameters to the fading wireless
used in IP networks] will most likely result in extremely conser- channel. Note that a (nonfading) additive white Gaussian noise
vative guarantees. For example, in a Rayleigh or Ricean fading (AWGN) channel of capacity can be specified by the
channel, the only lower bound that can be deterministically triplet ; i.e., an AWGN channel can guarantee
guaranteed is a capacity3 of zero. This conservative guarantee constant data rate.
is clearly useless. Therefore, we propose to extend the concept At first sight, relating to the fading wireless
of deterministic SC to a statistical version, specified as channel behavior seems to be a hard problem. However, at this
the pair . The statistical SC specifies that point, we use the idea that the SC is a dual of the traffic
the service provided by the channel, denoted as , will envelope . A rich body of literature exists on the so-called
always satisfy the property that . In theory of effective bandwidth [3], which models the statistical
other words, is the probability that the wireless channel will behavior of traffic. In particular, the theory shows that the rela-
not be able to support the pledged SC . For most practical tion
values of , a nonzero SC can be guaranteed.
To summarize, we propose to extend the QoS mechanisms (5)
used in wired networks to wireless links, by using the traffic
and service characterizations popularly used in wired networks, is satisfied for large , by choosing two parameters (which are
namely, the traffic envelope and the SC , respectively. functions of the channel rate ) that depend on the actual data
However, recognizing that the time-varying wireless channel traffic, namely, the probability of nonempty buffer, and the ef-
cannot deterministically guarantee a useful SC, we propose to fective bandwidth of the source. Thus, a source model defined
use a statistical SC . by these two functions fully characterizes the source from a QoS
As mentioned earlier, it is hard to extract a statistical SC viewpoint. The duality between (3) and (5) indicates that it may
using the existing physical-layer channel models. In fact, in Sec- be possible to adapt the theory of effective bandwidth to SC
tion III-D, we show how physical-layer channel models can be characterization. This adaptation will point to a new channel
used to derive in an integral form. There, the reader model, which we call the EC link model. Thus, the EC link
will see that: 1) it is not always possible to extract model can be thought of as the dual of the effective bandwidth
from the physical-layer model (such as, when only the Doppler source model, which is commonly used in networking.
spectrum, but not the higher-order statistics are known); and 2) The rest of this section is organized as follows. In Sec-
even if it is possible, the computation involved may make the ex- tion III-A, we present the theory of effective bandwidth
traction extremely hard to implement. This motivates us to con- using the framework of Chang and Thomas [3]. An accurate
sider link-layer modeling, which we describe in Section III. The and efficient source traffic estimation algorithm exists [11],
philosophy here is that we want to model the wireless channel which can be used to estimate the functions of the effective
at the layer in which we intend to use the model. bandwidth source model. Therefore, we use a dual estimation
algorithm to estimate the functions of the proposed EC link
model in Section III-B. In Section III-C, we provide physical
interpretation of our link model. Section III-D shows that in
the special case of Rayleigh-fading channel at low SNRs, it is
2 and  are meant to be in a statistical sense. The maximum fade dura- possible to extract the SC from a physical-layer channel model.
tion  is a parameter that relates the delay constraint to the channel service; it For Rayleigh-fading channels at high SNRs, the extraction is
determines the probability sup Pr ( ) 9( )
fS t < t g. We will see later that  complicated, whereas the extraction may not even be possible
is specified by the source with = D , where D is the delay bound
for other types of fading. Therefore, our link-layer EC model
required by the source.
3The capacity here is meant to be delay-limited capacity, which is the max- has substantial advantage over physical-layer models, in
imum rate achievable with a prescribed delay bound (see [8] for details). specifying SCs, and hence QoS.
634 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 4, JULY 2003

(a) (b)
Fig. 4. (a) Effective bandwidth function (u) and (b) EC function u
( ).

A. Theory of Effective Bandwidth If the quantity of interest is the delay experienced by a

The stochastic behavior of a source traffic process can be source packet arriving at time (see Fig. 3), then the probability
modeled asymptotically by its effective bandwidth. Consider an of exceeding a delay bound satisfies4
arrival process , , where represents the amount
(10)
of source data (in bits) over the time interval . Assume that
the asymptotic log-moment generating function of , defined
as Thus, the key point is that for a source modeled by the pair
, which has a communication delay bound of
(6) , and can tolerate a delay-bound violation probability
of at most , the effective bandwidth concept shows that the
constant channel capacity should be at least , where is the
exists for all . Then, the effective bandwidth function of
solution to . In terms of the traffic envelope
is defined as
(Fig. 3), the slope and .
Fig. 4(a) shows a typical effective bandwidth function. It can
(7) be easily proved that is equal to the average data rate of the
source, while is equal to the peak data rate. From (10),
See [3] for details. note that a source that has a more stringent QoS requirement
Consider a queue of infinite buffer size served by a channel of (i.e., smaller or smaller ), will need a larger QoS expo-
constant service rate (see Fig. 2), such as an AWGN channel. nent .
Due to the possible mismatch between and , the queue Reference [11] shows a simple and efficient algorithm to es-
length (see Fig. 3) could be nonzero. Using the theory of timate the source model functions and . In the fol-
large deviations, it can be shown that the probability of lowing section, we use the duality between traffic modeling
exceeding a threshold satisfies [3] ( ), and channel modeling to propose an EC link
model, specified by a pair of functions . It
as (8) is clear that we intend to be the channel
duals of the source functions . Just as the constant
where means that . How- channel rate is used in source traffic modeling, we use the con-
ever, it is found that for smaller values of , the following ap- stant source traffic rate in modeling the channel. Furthermore,
proximation is more accurate [4]: we adapt the source estimation algorithm in [11] to estimate the
link model functions .
(9)
B. EC Link Model

where both and are functions of channel capacity . Let be the instantaneous channel capacity at time . De-
According to the theory, is the proba- fine , which is the service provided by the
bility that the buffer is nonempty for randomly chosen time , channel. Note that the channel service is different from the
while the QoS exponent is the solution of . Thus, actual service received by the source; only depends
the pair of functions model the source. Note that on the instantaneous channel capacity and, thus, is independent
is simply the inverse function corresponding to the effec- 4 (r ) in (10) is different from  (r ) in (9). The relationship between them
tive bandwidth function . 2
is  (r ) =  (r ) r [15, pp. 57].
WU AND NEGI: EFFECTIVE CAPACITY: A WIRELESS LINK MODEL FOR SUPPORT OF QoS 635

same. This can be illustrated by Fig. 5: whenever the curve

is below , the horizontal line will cross the line ;
i.e., we have an event , since the horizontal
distance between and is . Then, we have

(14)

where follows from the fact that

and are the same event, and from
the fact that for any . From (14), it can
be seen that implies
.
Fig. 4(b) shows that the EC decreases with increasing
QoS exponent ; that is, as the QoS requirement becomes more
Fig. 5. Relation between SC and delay-bound violation. stringent, the source rate that a wireless channel can support
with this QoS guarantee, decreases. The channel sustainable rate
of the arrival . Paralleling the development in Section III-A, is upper bounded by the AWGN capacity and lower
we assume that bounded by the minimum rate 0. Fig. 4(a) and (b) together il-
lustrate the duality of the effective bandwidth source model and
(11)
the EC link model.
exists for all . This assumption is valid, for example, for The function pair defines our proposed
a stationary Markov-fading process . Then, the EC function EC link model. The definition of these functions shows that the
of is defined as EC model is a link-layer model, because it directly characterizes
the queueing behavior at the link layer. From (13), it is clear that
(12) the QoS metric can be easily extracted from the EC link model.
Now that we have shown that QoS metric calculation is
Consider a queue of infinite buffer size supplied by a data
trivial, once the EC link model is known, we need to specify
source of constant data rate (see Fig. 2). The theory of effec-
a simple (and hopefully, accurate) channel estimation al-
tive bandwidth presented in Section III-A can be easily adapted
gorithm. Such an algorithm should estimate the functions
to this case. The difference is that whereas in Section III-A the
from channel measurements, such as the
source rate was variable while the channel capacity was con-
measured SNR or channel capacity .
stant, in this section, the source rate is constant while the channel
Let us take a moment to think about how we could
capacity is variable. Similar to (10), it can be shown that the
use the existing physical-layer channel models to estimate
probability of exceeding a delay bound satisfies
. An obvious fact that emerges is that if a
(13) channel model specifies only the marginal probability density
function (pdf) at any time (such as Ricean pdf), along with
where are functions of source rate . The the Doppler spectrum (such as Gans Doppler spectrum),
approximation in (13) is accurate for large , but we will which is second-order statistics, then the model does not have
show later, in the simulations, that this approximation is also enough information to calculate the EC function (12). Indeed,
accurate even for smaller values of . such a calculation would need higher order joint statistics,
For a given source rate , is again which cannot be obtained merely from the Doppler spectrum.
the probability that the buffer is nonempty at a randomly chosen Thus, only approximations can be made in this case. For a
time , while the QoS exponent is defined as Rayleigh-fading distribution, the joint pdf of channel gains
, where is the inverse function of . Thus, will be complex Gaussian, and hence, the Doppler spectrum
the pair of functions model the link. is enough to calculate the EC. This result is presented in
So, when using a link that is modeled by the pair Section III-D. However, it will be shown there that even in
, a source that requires a communication the Rayleigh-fading case, the calculation is complicated, and,
delay bound of , and can tolerate a delay-bound violation therefore, not likely to be practical.
probability of at most , needs to limit its data rate to a max- Assume that the channel-fading process is stationary
imum of , where is the solution to . and ergodic. Then, a simple algorithm to estimate the functions
In terms of the SC (Fig. 3), the channel sustainable rate is the following (adapted from [5], [11]):
and . It is clear that when the buffer
is empty, the extra service will be lost, i.e., (15)
not used for data transmission (see Fig. 5). So, we have
for any time . Furthermore, we know that the
event and the event are the (16)
636 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 4, JULY 2003

and
(17)
where is the average remaining service time of a packet
being served. Note that is zero for a fluid model (assuming
infinitesimal packet size). The intuition in (15) is that, since the
distribution of is approximately exponential for large
[see (13)], then is given by (15). Now, the delay
is the sum of the delay incurred due to the packet already in
service, and the delay in waiting for the queue to clear. This
results in (16), using Little’s theorem. Substituting
in (13) results in (17).
Solving (16) for , we obtain

(18)

Equations (17) and (18) show that the functions and can
be estimated by estimating , , and . Fig. 6. Marginal CDF and () versus source rate .
The latter can be estimated by taking a number of samples,
say , over an interval of length , and recording the fol-
to estimate , which can then be used in (13).
lowing quantities at the th sampling epoch: the indicator of
This completes the specification of our link-layer model.
whether a packets is in service ( ), the number
The EC link model and its application are summarized below.
of bits in the queue (excluding the packet in service), and the
remaining service time of the packet in service (if there is one EC link model:
in service). The following sample means are computed: is the EC link model, which exists
if the log-moment generating function in (11)
(19) exists [e.g., for a stationary Markov-fading process
In addition to its stationarity, if is also ergodic,
then can be estimated by (19) through
(20) (22).
Given the EC link model, the QoS can
and
be computed by (23), where
(21) The resulting QoS corresponds directly
to the SC specification with
Then, from (18), we have and

(22)
C. Physical Interpretation of Our Model
Equations (19)–(22) constitute our channel estimation
algorithm, to estimate the EC link model functions We stress that the model presented in the previous section,
. They can be used to predict the QoS , is not just a result of mathematics (i.e., large
by approximating (13) with deviation theory). But rather, the model has direct physical inter-
pretation; i.e., corresponds to marginal cu-
(23) mulative distribution function (CDF) and Doppler spectrum of
the underlying physical-layer channel. This correspondence can
Furthermore, if the ultimate objective of EC link modeling is be illustrated as follows:
to compute an appropriate SC , then as mentioned ear- • The probability of nonempty buffer is similar to
lier, given the delay bound and the target delay-bound the concept of marginal CDF (e.g., Rayleigh–Ricean dis-
violation probability of a connection, we can find tribution), or equivalently outage probability (the proba-
by: 1) setting ; and 2) solving (23) bility that the received SNR falls below a certain speci-
for and setting . A fast binary search procedure that fied threshold). As shown later in Fig. 6, different mar-
estimates for a given and , is shown in the Appendix. ginal CDF of the underlying physical-layer channel cor-
This section introduced the EC link model, which is parame- responds to different . However, the two functions,
terized by the pair of functions . It was shown marginal CDF (i.e., outage probability) and , are
that these functions can be easily used to derive QoS guarantees not equal. The reason is that the probability of nonempty
(13), such as a bound that uses . Furthermore, this buffer takes into account the effect of packet accumula-
section specified a simple and efficient algorithm [(19)–(22)] tion in the buffer, while the outage probability does not
WU AND NEGI: EFFECTIVE CAPACITY: A WIRELESS LINK MODEL FOR SUPPORT OF QoS 637

channel models cannot be easily used to extract QoS guar-

antees. In [14], the authors attempt to model the wireless
physical-layer channel using a discrete state representation.
For example, [14] approximates a Rayleigh flat-fading channel
as a multistate Markov chain, whose states are characterized
by different bit-error rates. With the multistate Markov chain
model, the performance of the link layer can be analyzed, but
only at expense of enormous complexity. In this section, we
outline a similar method for the Rayleigh flat-fading channel
at low SNRs.
As mentioned earlier, cannot be calcu-
lated using (12), in general, if only the marginal pdf at any time
and the Doppler spectrum are known. However, such an ana-
lytical calculation is possible for a Rayleigh flat-fading channel
in AWGN, albeit at very high complexity.
Suppose that the wireless channel is a Rayleigh flat-fading
channel in AWGN with Doppler spectrum . Assume that
we have perfect causal knowledge of the channel gains. For ex-
Fig. 7.  versus Doppler rate f .
ample, the Doppler spectrum from the Gans model [12] is
the following:
(i.e., an arrival packet will be immediately discarded if the
SNR falls below a threshold). Therefore, the probability of (24)
nonempty buffer is larger than the outage probability, be-
cause buffering causes longer busy periods compared with
the nonbuffered case.
From Fig. 6, we observe that and marginal CDF where is the maximum Doppler frequency, is the carrier
have similar behavior: 1) both increases with the source frequency, and .
rate ; and 2) a large outage probability at the physical We show how to calculate the EC for this channel. Denote
layer results in a large at the link layer. Thus, a sequence of measurements of the channel gain, spaced
does reflect the marginal CDF of the underlying at a time-interval apart, by , where
wireless channel. are the complex-valued channel gains
• , defined as the decay rate of the probability ( are, therefore, Rayleigh distributed). Without loss of
, corresponds to the Doppler generality, we have absorbed the constant noise variance into
spectrum. This can be seen from Fig. 7. As shown in the definition of . The measurement is a realization of
the figure, different Doppler rates give different . a random variable sequence denoted by , which can be
In addition, the figure shows that increases with written as the vector . The pdf of a
the Doppler rate. The reason for this is the following. random vector for the Rayleigh-fading channel is
As the Doppler rate increases, the time diversity of the
channel also increases. This implies lower delay-violation (25)
probability , which leads to a
larger decay rate . Therefore, does reflect where is the covariance matrix of the random vector ,
the Doppler spectrum of the underlying physical channel. the determinant of matrix , and the conjugate of
Note that a stationary Markov-fading process (as is com- . Now, to calculate the EC, we first need to calculate
monly assumed, for a physical wireless channel), will
always have a log-moment generating function .
Therefore, the EC link model is applicable to such a case. On
the other hand, in Section IV-B, the simulation results show
that the delay-violation probability
does decrease exponentially with the delay bound (see
Figs. 8–10). Therefore, our link model is reasonable, not only
from a theoretical viewpoint (i.e., Markovian property of fading
channels) but also from an experimental viewpoint (i.e., the
actual delay-violation probability decays exponentially with
the delay bound).
(26)
D. QoS Guarantees Using Physical-layer Channel Models
In this section, we show that in contrast to our approach, where approximates the integral by a sum from the
which uses a link-layer model, the existing physical-layer standard result on Gaussian channel capacity (i.e.,
638 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 4, JULY 2003

(a) (b)
Fig. 8. Actual delay-violation probability versus D for various Doppler rates. (a) Rayleigh fading and (b) Ricean fading.

(a) (b)
Fig. 9. Prediction of delay-violation probability when the average SNR is (a) 15 and (b) 0 dB.

, where is the modulus of ), and from

(25). This gives the EC (12) as

(27)

In general, the integral in (27) is of high dimension (i.e.,

dimensions) and it does not reduce to a simple form, except for (28)
the case of low SNR, where approximation can be made. Next,
where using the approximation for
we show a simple form of (27), for the case of low SNR. We
(26) (if is small, that is, low SNR), by the definition of
first simplify (26) as follows:
the norm of the vector , and by the relation
(where is identity matrix). We consider three cases of interest
for (28).
• Case 1 (Special Case) Suppose , where
is the average channel capacity. This case happens
WU AND NEGI: EFFECTIVE CAPACITY: A WIRELESS LINK MODEL FOR SUPPORT OF QoS 639

• Case 3 (General Case) Denote the eigenvalues of

matrix by . Since is sym-
metric, we have , where is a unitary
matrix, is its Hermitian, and the diagonal matrix
. From (28), we have

(32)
Fig. 10. Prediction of delay-violation probability, when f = 5 Hz.
Case 3 is the general case for a Rayleigh flat-fading
channel at low SNRs.
when a mobile moves very fast with respect to the sample We now use the calculated to derive the log-mo-
period. From (28), we have ment generating function as

(29)

where follows from the fact that the sample period

is .
As the number of samples , we have
(33)
(30)
where follows from (32), follows from the fact that the
frequency interval and the bandwidth , and
Thus, in the limiting case, the Rayleigh-fading channel re- from the fact that the power spectral density
duces to an AWGN channel. Note that this result would not and that the limit of a sum becomes an integral. This gives the
apply at high SNRs because of the concavity of the EC (12) as
function. Since Case 1 has the highest degree of diversity,
it is the best case for guaranteeing QoS; i.e., it provides (34)
the largest EC among all the Rayleigh-fading processes
with the same marginal pdf. It is also the best case for high Thus, the Doppler spectrum allows us to calculate .
SNR. The EC function (34) can be used to guarantee QoS using (13).
• Case 2 (Special Case) Suppose , where Remark 1: We argue that even if we have perfect knowledge
denotes matrix transpose and . Thus, about the channel gains, it is hard to extract QoS metrics from
all the samples are fully correlated, which could occur if the physical-layer channel model, in the general case. The EC
the wireless terminal is immobile. From (28), we have function (34) is valid only for a Rayleigh flat-fading channel, at
low SNR. At high SNR, the EC for a Rayleigh-fading channel
is specified by the complicated integral in (27). To the best of
our knowledge, a closed-form solution to (27) does not exist. It
is clear that a numerical calculation of EC is also very difficult,
(31) because the integral has a high dimension. Thus, it is difficult to
extract QoS metrics from a physical-layer channel model, even
Since Case 2 has the lowest degree of diversity, it is the for a Rayleigh flat-fading channel. The extraction may not even
worst case. Specifically, Case 2 provides zero EC because be possible for more general fading channels. In contrast, the
a wireless terminal could be in a deep fade forever, making EC link model that we have proposed can be easily translated
it impossible to guarantee any nonzero capacity. It is also into QoS metrics for a connection, and we have shown a simple
the worst case for high SNR. estimation algorithm to estimate the EC model functions.
640 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 4, JULY 2003

Fig. 11. Queueing model used for simulations.

IV. SIMULATION RESULTS TABLE II

SIMULATION PARAMETERS
In this section, we simulate a queueing system and demon-
strate the performance of our algorithm for estimating the func-
tions of the EC link model. Section IV-A describes the simula-
tion setting, while Section IV-B illustrates the performance of
our estimation algorithm.

A. Simulation Setting
We simulate the discrete-time system depicted in Fig. 11. In where the noise is zero-mean complex Gaussian with unit
this system, the data source generates packets at a constant rate variance per dimension and is statistically independent of .
. Generated packets are first sent to the (infinite) buffer at the The coefficient can be determined by the following procedure:
transmitter, whose queue length is , where refers to the 1) compute the coherence time by [12, pp. 165]
th sample interval. The head-of-line packet in the queue is
transmitted over the fading channel at data rate . The fading (39)
channel has a random channel gain (the noise variance is ab-
sorbed into ). We use a fluid model that is the size of a packet where the coherence time is defined as the time, over which the
is infinitesimal. time autocorrelation function of the fading process is above 0.5;
We assume that the transmitter has perfect knowledge of 2) compute the coefficient by5
the channel gain (the SNR, really) at each sample interval. (40)
Therefore, it can use rate-adaptive transmissions and strong
where is the sampling interval.
channel coding to transmit packets without errors. Thus, the
Table II lists the parameters used in our simulations.
transmission rate is equal to the instantaneous (time-varying)
capacity of the fading channel, as below B. Performance of the Estimation Algorithm
(35) We organize this section as follows. In Section IV-B1, we es-
timate the functions of the EC link model
where is the channel bandwidth.
from the measured . Section IV-B2 provides simulation re-
The average SNR is fixed in each simulation run. We define
sults that demonstrate the relation between the physical channel
as the capacity of an equivalent AWGN channel, which
and our link model. In Section IV-B3, we show that the esti-
has the same average SNR, i.e.
mated EC functions accurately predict the QoS metric, under a
SNR (36) variety of conditions.
1) EC Model Estimation: In the simulations, a
where SNR is the average SNR, i.e., . Then, we can Rayleigh flat-fading channel is assumed. We simulate four
eliminate using (35) and (36) as cases: 1) SNR dB and the maximum Doppler rate
Hz; 2) SNR dB and Hz; 3)
(37) SNR dB and Hz; and 4) SNR dB
SNR
and Hz. Figs. 12 and 13 show the estimated EC
In our simulations, the sample interval is set to 1 ms. This functions and . As the source rate increases
is not too far from reality, since 3G WCDMA systems already from 30 to 85 kb/s, increases, indicating a higher buffer
incorporate rate adaptation on the order of 10 ms [9]. occupancy, while decreases, indicating a slower decay
Each simulation run is 1000 s long in all scenarios. Since the of the delay-violation probability. Thus, the delay-violation
channel sample rate is 1000 samples/s, 1 000 000 samples of probability is expected to increase with increasing source rate
Rayleigh–Ricean flat-fading were generated for each 1000-s . From Fig. 12, we also observe that SNR has a substantial
run, using a first-order autoregressive (AR) model. Specifically, impact on . This is because higher SNR results in larger
is generated by the following AR(1) model:
5The autocorrelation function of the AR(1) process is  , where n is the
(38) number of sample intervals. Solving  = 0:5 for , we obtain (40).
WU AND NEGI: EFFECTIVE CAPACITY: A WIRELESS LINK MODEL FOR SUPPORT OF QoS 641

Rayleigh channel, i.e., the probability that the SNR falls below
a threshold SNR , is
SNR SNR SNR SNR (41)
Similar to (37), we have the source rate
SNR
(42)
SNR
Solving (42) for SNR , we obtain
SNR SNR (43)
Using (41) and (43), we plot the marginal CDF of the Rayleigh
channel, as a function of source rate . Similarly, we plot the
marginal CDF of the Ricean channel as a function of source rate
.
As shown in Fig. 6, different marginal CDF at the physical
layer yields different at the link layer. We observe that
and marginal CDF have similar behavior: 1) both increases with
Fig. 12. Estimated function ^() versus source rate . the source rate ; and 2) if one channel has a larger outage prob-
ability than another channel, it also has a larger than the
other channel. For example, in Fig. 6, the Rayleigh channel has
a larger outage probability and a larger than the Ricean
channel. Thus, the probability of nonempty buffer is sim-
ilar to marginal CDF, i.e., outage probability.
Figs. 7 and 8 show the result for Scenario B. From Fig. 7,
it can be seen that different Doppler rate at the physical layer
leads to different at the link layer. In addition, the figure
shows that increases with the Doppler rate. This is rea-
sonable since the increase of the Doppler rate leads to the in-
crease of time diversity, resulting in a larger decay rate of
the delay-violation probability. Therefore, corresponds to
the Doppler spectrum of the physical channel.
Fig. 8 shows the actual delay-violation probability
versus the delay bound ,
for various Doppler rates. It can be seen that the actual
delay-violation probability decreases exponentially with the
delay bound , for all the cases. This justifies the use of an
exponential bound (23) in predicting QoS, thereby justifying
Fig. 13. Estimated function ^() versus source rate .
our link model .
3) Accuracy of the QoS Metric Predicted by and : In
channel capacity, which leads to smaller probability that a the previous section, the simulation results have justified the
packet will be buffered, i.e., smaller . In contrast, Fig. 12 use of in predicting QoS. In this section, we evaluate
shows that has little effect on . The reason is that the accuracy of such a prediction. To test the accuracy, we
reflects the marginal CDF of the underlying fading process, use and to calculate the delay-bound violation probability
rather than the Doppler spectrum. [using (23)], and then compare
2) Physical Interpretation of Link Model : To illus- the estimated probability with the actual (i.e., measured)
trate that different a physical channel induces different parame- .
ters , we simulate two kinds of channels, i.e., a Rayleigh To show the accuracy, we simulate three scenarios. In the first
flat-fading channel and a Ricean flat-fading channel. For the scenario, the source rates are 75/80/85 kb/s, which loads the
Rayleigh channel, we set the average SNR to 15 dB. For the system as light/moderate/heavy, respectively. For all three cases,
Ricean channel, we set the factor6 to 3 dB. We simulate two we simulate a Rayleigh flat-fading channel with SNR
scenarios: A) changing the source rate while fixing the Doppler dB, kb/s, and Hz. Fig. 9(a) plots
rate at 30 Hz; and B) changing the Doppler rate while fixing the the actual and the estimated delay-bound violation probability
source rate, i.e., kb/s. as a function of . As predicted by
The result for Scenario A is shown in Fig. 6. For comparison, (23), the delay-violation probability follows an exponential de-
we also plot the marginal CDF (i.e., Rayleigh–Ricean CDF) of crease with . Furthermore, the estimated
the physical channel in the same figure. The marginal CDF for is close to the actual .
6The K factor is defined as the ratio between the deterministic signal power In the second scenario, we also set kb/s and
A and the variance of the multipath 2 ; i.e., K = A =(2 ). Hz, but change the average SNR to 0 dB. Fig. 9(b)
642 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 2, NO. 4, JULY 2003

shows that the conclusions drawn from the first scenario still APPENDIX
hold. Thus, our estimation algorithm gives consistent perfor- We show that the functions that specify
mance over different SNRs also. our EC link model can be easily used to obtain the SC specifi-
In the third scenario, we set SNR dB and
cation . The parameter is simply equal
kb/s, but we change the Doppler rate to 5 Hz. Fig. 10
to the source delay requirement . Thus, only the channel
shows that the conclusions drawn from the first scenario still
sustainable rate needs to be estimated. is the source
hold. Thus, our estimation algorithm consistently predicts the
rate at which the required QoS (delay-violation probability )
QoS metric under different Doppler rate .
is achieved.
In summary, the simulations illustrate that our EC link model,
The following binary search procedure estimates for
together with the estimation algorithm, predict the actual QoS
a given (unknown) fading channel and source specification
accurately.
. In the algorithm, is the error between the target
and the estimated , the precision
tolerance, the source rate, a lower bound on the source
V. CONCLUDING REMARKS rate, and an upper bound on the source rate.
Efficient bandwidth allocation and QoS provisioning over
wireless links demand a simple and effective wireless channel Algorithm 1 (Estimation of the channel
model. In this paper, we have modeled a wireless channel from sustainable rate )
the perspective of the communication link layer. This is in con- /* Initialization */
trast to existing channel models, which characterize the wireless Initialize , , and ;
channel at the physical layer. Specifically, we modeled the wire- /* E.g., ; ; ; */
less link in terms of two “EC” functions, namely, the probability ; /* An obvious lower bound on the
of nonempty buffer and the QoS exponent . The rate */
QoS exponent is the inverse of a function which we call EC. ; /* An obvious upper bound on
The EC link model is the dual of the effective bandwidth source the rate is the AWGN capacity */
traffic model used in wired networks. Furthermore, we devel- ;
oped a simple and efficient algorithm to estimate the EC func- /* Binary search to find a that is
tions . Simulation results show that the actual conservative and within */
QoS metric is closely approximated by the QoS metric predicted While (( ) or ( )) {
by the EC link model and its estimation algorithm, under var- The data source transmits at the
ious scenarios. output rate ;
We have provided key insights about the relations between Estimate and using (19) to (22);
the EC link model and the physical-layer channel, i.e., Use (23) to obtain ;
corresponds to the marginal CDF (e.g., Rayleigh–Ricean distri- ;
bution) while is related to the Doppler spectrum. The EC if ( ) {/* Conservative */
link model has been justified not only from a theoretical view- if ( ) {/* Too conservative */
point (i.e., Markov property of fading channels) but also from ; /* Increase rate */
an experimental viewpoint (i.e., the delay-violation probability ;
does decay exponentially with the delay). }
The QoS metric considered can be easily translated into }
traffic envelope and SC characterizations, which are popular else {/* Optimistic */
in wired networks, such as ATM and IP, to provide guaranteed ; /* Reduce rate */
service. Therefore, we believe that the EC link model, which ;
was specifically constructed keeping in mind this QoS metric, }
will find wide applicability in future wireless networks that }
need QoS provisioning. .
In summary, our EC link model has the following features:
simplicity of implementation, efficiency in admission control,
and flexibility in allocating bandwidth and delay for connec- Algorithm 1 uses a binary search to find the channel sustain-
tions. In addition, our link model provides a general framework, able rate . An alternative approach is to use a parallel search,
under which physical-layer fading channels such as AWGN, such as the one described in [11]. A parallel search would re-
Rayleigh-fading, and Ricean-fading channels can be studied. quire more computations, but would converge faster.
Armed with the new link model, we are now investigating its
use in designing admission control, resource reservation, and REFERENCES
scheduling algorithms, for efficient support of a variety of traffic [1] “Traffic management specification (version 4.0),” ATM Forum Tech.
flows that require guaranteed QoS. We are also exploring the Committee, 1996.
[2] L. Zhang, S. Berson, S. Herzog, and S. Jamin, “Resource ReSerVation
implementation of our link model in 3G wireless systems, and Protocol (RSVP) version 1, functional specification,” in RFC 2205, R.
its implications in QoS support for such networks. Braden, Ed: Internet Eng. Task Force, Sept. 1997.
WU AND NEGI: EFFECTIVE CAPACITY: A WIRELESS LINK MODEL FOR SUPPORT OF QoS 643

[3] C.-S. Chang and J. A. Thomas, “Effective bandwidth in high-speed dig- Dapeng Wu (S’98) received the B.E. degree in
ital networks,” IEEE J. Select. Areas Commun., vol. 13, pp. 1091–1100, electrical engineering from Huazhong University
Aug. 1995. of Science and Technology, Wuhan, China, and the
[4] G. L. Choudhury, D. M. Lucantoni, and W. Whitt, “Squeezing the most M.E. degree in electrical engineering from Beijing
out of ATM,” IEEE Trans. Commun., vol. 44, pp. 203–217, Feb. 1996. University of Posts and Telecommunications,
[5] A. E. Eckberg, “Approximations for bursty and smoothed arrival delays Beijing, China, in 1990 and 1997, respectively. From
based on generalized peakedness,” in Proc. 11th Int. Teletraffic Con- July 1997 to December 1999, he conducted graduate
gress, Kyoto, Japan, Sept. 1985. research at Polytechnic University, Brooklyn, New
[6] P. Ferguson and G. Huston, Quality of Service: Delivering QoS on the York. Since January 2000, he has been working
Internet and in Corporate Networks. New York: Wiley, 1998. toward the Ph.D. degree in electrical and computer
[7] R. Guerin and V. Peris, “Quality-of-service in packet networks: Basic engineering at Carnegie Mellon University, Pitts-
mechanisms and directions,” Computer Netw. ISDN, vol. 31, no. 3, pp. burgh, PA.
169–179, Feb. 1999. During summers 1998, 1999, and 2000, he conducted research at Fujitsu Lab-
[8] S. Hanly and D. Tse, “Multiaccess fading channels: Part II: Delay-lim- oratories of America, Sunnyvale, CA, on architectures and traffic management
ited capacities,” IEEE Trans. Inform. Theory, vol. 44, pp. 2816–2831, algorithms in the Internet and wireless networks for multimedia applications.
Nov. 1998. His current interests are in the areas of networking, information theory, commu-
[9] H. Holma and A. Toskala, WCDMA for UMTS: Radio Access for Third nication theory, and multimedia communications over the Internet and wireless
Generation Mobile Communications. New York: Wiley, 2000. networks.
[10] B. Jabbari, “Teletraffic aspects of evolving and next-generation wireless Mr. Wu received the IEEE Circuits and Systems for Video Technology
communication networks,” IEEE Pers. Commun., vol. 3, pp. 4–9, Dec. (CSVT) Transactions Best Paper Award for 2001.
1996.
[11] B. L. Mark and G. Ramamurthy, “Real-time estimation and dynamic
renegotiation of UPC parameters for arbitrary traffic sources in ATM
networks,” IEEE/ACM Trans. Networking, vol. 6, pp. 811–827, Dec. Rohit Negi (S’98–M’00) received the B.Tech. degree
1998. in electrical engineering from the Indian Institute of
[12] T. S. Rappaport, Wireless Communications: Principles & Prac- Technology, Bombay, India, in 1995. He received the
tice. Englewood Cliffs, NJ: Prentice-Hall, 1996. M.S. and Ph.D. degrees in electrical engineering from
[13] S. Shenker, C. Partridge, and R. Guerin, “Specification of guaranteed Stanford University, Stanford, CA, in 1996 and 2000,
quality of service,” in RFC 2212: Internet Eng. Task Force, Sept. 1997. respectively.
[14] Q. Zhang and S. A. Kassam, “Finite-state Markov model for Rayleigh Since 2000, he has been with the Electrical and
fading channels,” IEEE Trans. Commun., vol. 47, pp. 1688–1692, Nov. Computer Engineering Department at Carnegie
1999. Mellon University, Pittsburgh, PA, where he is an
[15] Z.-L. Zhang, “End-to-end support for statistical quality-of-service guar- Assistant Professor. His research interests include
antees in multimedia networks,” Ph.D. dissertation, Dept. Comput. Sci., signal processing, coding for communications
Univ. Massachusetts, 1997. systems, information theory, networking, and cross-layer optimization.
Dr. Negi received the President of India Gold Medal in 1995.