Proportional degradation services in wireless/mobile adaptive multimedia networks

WIRELESS COMMUNICATIONS AND MOBILE COMPUTINGWirel. Commun. Mob. Comput. 2005; 5:219–243Published online 23 August 2004 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/wcm.211

Proportional degradation services in wireless/mobile adaptivemultimedia networks

Yang Xiao1*,y, Haizhon Li1, C. L. Philip Chen2, Bin Wang3 and Yi Pan4

1Computer Science Division, The University of Memphis, 373 Dunn Hall, Memphis, TN 38152, U.S.A.2Department of Electrical and Computer Engineering, The University of Texas, San Antonio, Texas 78249-0669, U.S.A.3Department of Computer Science and Engineering Wright State University, Dayton, OH 45435, U.S.A.4Department of Computer Science, Georgia State University, Atlanta, GA 30303, U.S.A.

Summary

Adaptive multimedia services are very attractive since resources in wireless/mobile networks are relatively scarce

and widely variable, and more importantly the resource fluctuation caused by mobility and channel fading can be

mitigated using adaptive services. Therefore, there are extensive research activities on Quality of Service (QoS),

call admission control, as well as bandwidth degradation and adaptation for adaptive multimedia services in

wireless/mobile networks in recent years. However, fairness of bandwidth degradation has largely been ignored in

previous work and remains an important issue in adaptive multimedia service provisioning. In this paper, we

propose and study proportional degradation service provisioning in wireless/mobile networks that offer multiple

classes of adaptive multimedia services. The proposed proportional degradation fairness model guarantees the

proportional bandwidth degradation among different classes of services. Two proportional degradation scenarios

are studied in this paper. In the first scenario, we study the proportional degradation provisioning with two QoS

parameters for adaptive multimedia: the degradation ratio (DR) and the degradation degree (DD). In the second

scenario, we study the proportional degradation provisioning with a new QoS parameter for adaptive multimedia:

the degradation area (DA). For each scenario, based on the QoS parameters, proportional degradation adaptation

algorithms are proposed to approximate the proportional degradation model, to fairly adapt calls’ degradations, to

utilize the system resource efficiently, as well as to optimize QoS parameters. Performance studies show that in the

first scenario, proportional DR has been achieved very well, whereas proportional DD has not been well achieved.

In other words, DR outperforms DD in terms of proportional degradation. In the second scenario, proportional DA

has been well achieved. Furthermore, bandwidth resources have been efficiently utilized and DA has been

minimized. Copyright # 2004 John Wiley & Sons, Ltd.

KEY WORDS: adaptive multimedia; proportional degradation; quality of service; wireless networks

1. Introduction

Recently, there have been great demands for multi-

media applications with QoS in wireless/mobile net-

works. QoS provisioning in wireless/mobile networks

is more challenging than in wired networks due to

channel fading, inherent mobility and so forth [1].

Although channel fading can be improved with better

*Correspondence to: Yang Xiao, Computer Science Division, The University of Memphis, 373 Dunn Hall, Memphis, TN38152, U.S.A.yE-mail: [email protected]

Copyright # 2004 John Wiley & Sons, Ltd.

transmission and reception systems, mobility may

cause severe fluctuation of network resources [2].

For adaptive multimedia networking, the bandwidth

of an ongoing call is variable during its lifetime.

Adaptive multimedia services are very attractive since

resources in wireless/mobile networks are relatively

scarce and widely variable, and more importantly the

resource fluctuation caused by mobility and fading

can be mitigated using adaptive services. However,

with bandwidth adaptation, some of the ongoing calls

may be forced to operate under a degraded mode to

accommodate more calls in an overloaded system. In

other words, bandwidth adaptation entails bandwidth

degradation for some applications.

Adaptive multimedia was originally introduced in

wired networks to overcome resource fluctuation

caused by network congestion. Many adaptive multi-

media approaches were proposed such as layered

coding (or hierarchical encoding) [3–7] and filters

[8]. Adaptive multimedia can be further classified

into discrete adaptive multimedia and continuous

adaptive multimedia. In discrete adaptive multimedia,

the bandwidth takes on a set of discrete values in

approaches such as layered coding (or hierarchical

coding) [4,5], whereas in continuous adaptive multi-

media, the bandwidth takes on continuous values [6–

7]. In this paper, we consider only discrete adaptive

multimedia as in References [1,2,9–11]. With the

introduction of adaptive multimedia in wireless/mo-

bile networks, there has been extensive research

[1,2,9–14] on QoS, call admission control, as well

as bandwidth degradation and adaptation. While

much research has been done on fair wireless band-

width scheduling in wireless/mobile networks [15–

28], fairness of bandwidth degradation has largely

been ignored in previous work and remains an im-

portant issue in adaptive multimedia service provi-

sioning. To the best of our knowledge, this paper is the

first such effort towards providing proportional fair

bandwidth degradation in mobile/wireless networks.

Note that a fair bandwidth allocation algorithm cannot

produce fair bandwidth degradation.

The literature on bandwidth adaptation for adaptive

multimedia services in wireless/mobile networks is

abundant [1,2,9–11]. A bandwidth adaptation frame-

work for continuous multimedia services was pro-

posed in Reference [29]. In Reference [13], an optimal

adaptation algorithm for continuous adaptive multi-

media has been proposed at the expense of a large

message overhead. The tradeoff between the network

overhead and optimal bandwidth allocation has been

studied in Reference [10]. In Reference [11], we

proposed a k-level bandwidth adaptation algorithm

for one class of users without considering fairness

issues. In Reference [30], several bandwidth alloca-

tion algorithms with different optimization objectives

(e.g. revenue, complexity, quality and fairness) have

been studied. However, fairness is in terms of band-

width allocation, which is quite different from our

goal in this paper, i.e. providing fair bandwidth

degradation. As we discuss in Section 5.1, fair band-

width allocation cannot produce fair bandwidth de-

gradation.

Much research has been done on bandwidth degra-

dation and adaptation [1,2,9–11]. In Reference [2], a

QoS parameter, the degradation period ratio (DPR), is

proposed. DPR represents the portion of a call’s life-

time during which the call is degraded. However, DPR

does not capture the degree of bandwidth degradation.

Several other researches studied the bandwidth de-

gradation [29,31–34]. Odyssey [31] is a platform for

mobile data access and focuses on end-host adapta-

tion, where adaptation is defined as the trading of data

quality for resource consumption. Application-aware

adaptation is provided through collaboration between

concurrent applications and the operating system.

Total bandwidth is divided based on recent use; those

applications that have consumed a larger share in the

immediate past are assumed to need larger shares in

the immediate future. A small portion of the band-

width is reserved and divided fairly; this is to avoid

unduly punishing applications that do not use the

network for extended periods of time. Fugue [32]

provides interactive video on hand-held, mobile de-

vices through a division along time scales of adapta-

tion: per packet, per frame and per video. It also

focuses on end-host adaptation and considers single

wireless channel quality for finer-grained adaptations.

Adaptation is continuous. Bandwidth fairness issues

are not addressed. Mobiware [34], on the other hand,

places the responsibility for supporting adaptation in

the network instead of end-hosts. It is based on

programmable, active services placed throughout the

network. Each application submits a utility curve and

centralized points provide utility-fair bandwidth allo-

cations. A utility function captures the adaptive nature

over which an application can successfully adapt to

available bandwidth in terms of a utility curve that

represents the range of observed quality to bandwidth.

An adaptation policy allows the application to control

how it moves along its utility curve as resource

availability fluctuates. Bandwidth is allocated fairly

to all the flows in the sense that the same utility value

is achieved at an access point. Mobiware supports

220 Y. XIAO ET AL.

Copyright # 2004 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2005; 5:219–243

both discrete and continuous adaptation. MOBIWEB

[33] is deployed locally between or near the two ends

of a wireless link. It envisions a set of levels of quality

(LoQ) for discrete adaptation. MOBIWEB allows

only one stream at a time to adapt until no further

adaptation is necessary. It combines an admission

control algorithm and a dynamic prioritization

scheme. The admission control reserves resources

for each stream in order to guarantee at least the

minimum acceptable level of quality in the stream’s

performance. The dynamic prioritization is designed

to ensure fairness. The scheme relates the priority of a

stream inversely proportionally to its level of quality.

That is, whenever a stream is forced to adapt back-

wards by dropping a LoQ, its priority is increased by a

level and vice versa. This provides robustness and

fairness to streams that are already receiving a low

LoQ. None of the previous work explicitly considers

the frequency of bandwidth degradation among dif-

ferent service classes. Fairness of bandwidth alloca-

tion was considered in some of the previous work.

However, the fairness of bandwidth degradation was

not addressed.

Frequent small changes around an average applica-

tion quality may be annoying to many applications.

Some applications may wish to limit the frequency of

adaptation to change, e.g. multi-resolution applica-

tion, while others may wish to exploit any opportunity

for adaptation, e.g. real-time data applications. In

order to effectively characterize the bandwidth degra-

dation and to provide better QoS to service users, in

our previous work [11], we proposed two novel QoS

parameters for single class of adaptive multimedia

service: the degradation ratio (DR) and the degrada-

tion degree (DD). The two new QoS parameters

capture the frequency of degraded calls and the degree

of bandwidth degradation respectively. Our study

shows that DD and DR outperform other QoS para-

meters in terms of effectively characterizing band-

width adaptation [11]. Two sets of QoS parameters are

studied in this paper. In the first set, we study DR and

DD. In the second set, we study a new QoS parameter

for adaptive multimedia: the degradation area (DA).

We believe that this new QoS parameter is better in

terms of characterizing the bandwidth degradation.

In wireless/mobile networks that offer multiple

classes of services, bandwidth degradations suffered

by calls of different classes are different. It is desirable

to distribute the bandwidth degradation across differ-

ent classes of services in a fair manner. In particular,

we argue for proportional fair bandwidth degradation,

which will be elaborated in Section 5. Much of the

research focused on fair wireless bandwidth schedul-

ing algorithms [15–28], but none of them pays any

attention to fairness of bandwidth degradation. Our

study is not about a scheduling algorithm but band-

width allocation and adaptation to achieve fairness of

bandwidth degradation. Two proportional degradation

scenarios are studied in this paper. In the first scenario,

we study the proportional degradation provisioning

with DR and DD. In the second scenario, we study the

proportional degradation provisioning with DA. In

order to achieve both fairness and prioritization

among classes, we propose a proportional degradation

service model. Our proportional degradation service

model guarantees the ratios of service degradation

among classes. In other words, the relative amount of

bandwidth degradation incurred by different classes

should be proportional to a predefined ratio, and

service users can select the class that best meets their

quality and pricing constraints. Moreover, higher

priority classes in our proportional degradation ser-

vice model are consistently given better service than

lower priority classes even on short time-scales. To

this end, we propose proportional degradation adapta-

tion algorithms based on different sets of QoS para-

meters to approximate the proportional degradation

model and to fairly adapt calls’ degradations, as well

as to optimize QoS parameters.

A proportional approach may outperform the strict

prioritization approach proposed in References

[35,36], in which higher priority classes are serviced

before lower priority classes. The reason is that the

strict prioritization approach can result in long starva-

tion periods for lower priority classes if higher priority

classes are persistently backlogged [37]. A propor-

tional approach may also outperform the weighted fair

queuing (WFQ) approach [38], in which higher prior-

ity classes are allocated a larger amount of bandwidth

relative to the expected load in each class. The reason

is that the WFQ approach can often provide worse

QoS to higher priority classes than lower priority

classes on short time-scales [37].

The rest of this paper is organized as follows.

Adaptive multimedia traffic model is introduced in

Section 2. We show that such a model is a very general

one since it can express all the service categories:

constant bit rate (CBR), available bit rate (ABR), and

unspecified bit rate (UBR). Therefore, the adaptive

multimedia traffic model can handle both non-adap-

tive traffic and adaptive traffic seamlessly. Section 3

introduces QoS parameters for degradation character-

ization. Section 4 describes a measurement-based

CAC scheme. Section 5 presents the proportional

PROPORTIONAL DEGRADATION SERVICES 221


fair bandwidth degradation model. Section 6 presents

bandwidth degradation adaptation algorithms to ap-

proximate the proportional model. Simulation results

are reported in Section 7. The paper concludes in

Section 8.

2. A General Traffic Model

2.1. Adaptive Multimedia Traffic Model

In adaptive multimedia networking, a multimedia call

can dynamically adjust its bandwidth depending on

the network load situation during its lifetime.

We assume that there are m classes of users: {1,

2, . . . , m}, and the bandwidth of a call in class-i takes

its value from fbi;1; bi;2; . . . ; bi;j; . . . ; bi;Kig [1,2,9–11]

for 1 � i � m, where bi; j < bi; jþ1 for all 1 � j �Ki � 1, and Ki � 1, where Ki is the number of multi-

media layers for class-i. Note that the above ‘layer’

concept is not the ‘layer’ concept used in the layered

coding approaches [3–7]. Therefore, it does not ne-

cessarily stand for a particular video encoding method

or any particular multimedia encoding methods, but is

a general abstraction of adaptive multimedia that can

be expressed in terms of layers. For example, if the

bandwidth of a low quality voice call is bi;1 and

the bandwidth of a high quality voice call is bi;2,

then the bandwidth of a voice call may take its value

from fbi;1; bi;2g. If a cell is not overloaded, a call canbe assigned bi;2, that is, the bandwidth of a high

quality voice call. Otherwise, the call may be assigned

bi;1. Such a scheme can be implemented by the sub-

rating method [39].

Another example is the layered multi-resolution

coding approach, where multimedia receivers can

selectively choose to receive a subset of the layer-

encoded information depending on receivers’ capacity

[2,9]. Under such a coding approach, bi;jþ1 � bi;j is the

bandwidth of the j-th enhanced layer of the multi-

media stream. Such layered multi-resolution coding

techniques could be subband [3,13] or pyramid coding

[10]. The substream filtering function used to filter out

the higher enhanced layers can be implemented in the

base station [1]. How the filtering function is imple-

mented is beyond the scope of the paper. Readers are

referred to References [1,8] for details.

Figure 1 illustrates an example [30] of an

adaptive multimedia stream where Ki ¼ 3. If a cell

is lightly loaded and sufficient bandwidth is available,

a call will be allocated its maximum bandwidth bi;3.

Otherwise, depending on the cell’s load condition, the

call may be allocated bi;2 or bi;2. �bi;1 and �bi;2 are

respectively the 1st and the 2nd enhanced segments of

the multimedia stream in addition to the minimum

bandwidth bi;1. Examples could be {low quality video,

medium quality video, high quality video} or {low

quality audio, medium quality audio, high quality

audio} etc.

The wireless network, which is capable of imple-

menting such adaptive multimedia services, can be

High-Speed Circuit-Switched Data (HSCSD) in GSM

[39], or the next generation wireless network. HSCSD

is a data service for GSM to support large file transfer

and multimedia applications, such as mobile video

applications.

2.2. Service Categories

In order to efficiently support a multi-application

environment, we assume that services are classified

into three categories: CBR service, ABR service, and

UBR service. CBR service guarantees the required

bandwidth and can be used in non-adaptive multi-

media applications with strict bandwidth requirements

or non-multimedia applications, where the bandwidth

of an ongoing call is fixed during its lifetime. ABR

service only guarantees a minimum amount of band-

width and can be used by adaptive multimedia appli-

cations, where the bandwidth of an ongoing call is time

varying during its lifetime. UBR is the same as the best

effort service and does not guarantee any amount of

bandwidth. Here, we make an assumption that a call in

this service can be suspended and be reactivated since

the call’s bandwidth could be zero during its lifetime.

UBR service can be used for non-real-time multimedia

applications, e.g. email etc. If a user does not want his/

her call’s bandwidth to be adaptive, he/she can request

CBR services instead of ABR services.

2.3. A General Abstract Model

In this section, we show that the adaptive multimedia

traffic model in Section 3.1 can express all the traffic

categories in Section 2.2.

Fig. 1. Example of an Adaptive Multimedia Stream.

222 Y. XIAO ET AL.


For i-th class, its traffic model is

fbi;1; bi;2; . . . ; bi;j; . . . ; bi;Kig. IfKi ¼ 1 and bi;1 > 0,

we have fbi;1g, which can be used to describe CBR

service traffic. If Ki > 1 and bi;1 > 0, we obtain

fbi;1; bi;2; . . . ; bi;j; . . . ; bi;Kig, which can be used to

describe ABR service traffic. If Ki > 1 and bi;1 ¼ 0,

we obtain f0; bi;2; . . . ; bi;j; . . . ; bi;Kig, which can be

used to describe UBR service traffic. Therefore, the

adaptive multimedia traffic model in Section 2.1 can

express all the traffic categories in Section 2.2.

An adaptive multimedia call can dynamically

change its bandwidth depending on the network situa-

tion during its lifetime. If a user does not want his/her

call’s bandwidth to be adaptive, CBR service can be

used instead of ABR service. CBR traffic is a special

case of ABR service traffic. In other words, the non-

adaptive traffic model is a special case of the adaptive

traffic model. Moreover, adaptive traffic may include

some real adaptive traffic (ABR) and some non-adap-

tive traffic (CBR). Since the CBR service is a special

case of the ABR service, in later sections, we will not

make a distinction between them. Table I shows an

example of multi-class traffic with seven classes. The

1st and 4th classes belong to CBR service; the 2nd,

3rd, 5th and 6th classes belong to ABR service and the

7th class belongs to UBR service.

System events include arrival events and service

departure events. Arrival events include new call

arrival events and handoff call arrival events. Service

departure events include call completion events and

events of call handoffs to other cells. We consider

fixed capacity in each cell as in previous related work.

The fixed total number of channels is C. With this

background, we propose degradation QoS parameters

in the following section.

3. QoS Parameters for DegradationCharacterization

The most significant QoS parameter in non-multi-

media or non-adaptive multimedia services is the

forced termination probability (FTP), the probability

of terminating an ongoing call before the completion

of the service. However, in adaptive multimedia

services, FTP can be near zero [1,12] as long as we

design the CAC scheme and the bandwidth adaptation

algorithm properly. There are two other common QoS

parameters. The first one is the call blocking prob-

ability (CBP), the probability of a new arriving call

being blocked. The second is the handoff dropping

probability (HDP), the probability of a handoff arriv-

ing call being dropped. HDP is more important than

CBP since it is directly related to FTP, and service

users do not like their calls terminated suddenly just

because of changing cells. On the other hand, the fact

that a new call is rejected is easier to be accepted by

service users compared to handoff dropping. We can

see from our simulation results that HDP can be near

zero too due to the adaptive nature coupled with our

proposed CAC scheme and the proposed bandwidth

adaptation algorithms. For adaptive multimedia ser-

vices, the problems of handoff dropping and forced

termination in traditional wireless networks are con-

verted to the problem of bandwidth degradation

caused by the adaptation. Therefore, new QoS para-

meters are needed to effectively characterize the

bandwidth degradation and to provide better QoS to

service users.

3.1. Degradation Ratio and Degradation Degree

In our previous work [11], we proposed and studied

two new QoS parameters: the DR and the DD for

single class of adaptive multimedia services. The new

QoS parameters characterize the frequency of de-

graded calls and the degree of bandwidth degradation

respectively. These two parameters have been found

to outperform other QoS parameters in terms of

effectively characterizing bandwidth adaptation [11].

In this work, we apply DD and DR in wireless/mobile

networks that provide multiple classes of services.

The requested bandwidth of a call in class-i

is denoted as bi;req, where bi;req 2fbi;1; bi;2; . . . ;bi;j; . . . ; bi;Ki

g. We assume that all the users in the

same class use the same requested bandwidth so that

the requested bandwidth for each class is predefined.

We also assume bi;req > bi;1. Otherwise, the adaptive

multimedia service degenerates to a non-adaptive one

and the new QoS parameters will be so trivial that they

are always satisfied.

Let xiðtÞ denote the number of calls of class-i in a

cell at time t. Let bi;assiðk; tÞ denote the assigned

bandwidth for a call k of class-i at time t, where

Table I. An example of multi-class traffic.

Class 1: fb1g, /* a voice traffic with a fixed bandwidth */Class 2: fc1; c2; c3g, /* {low quality video, medium

quality video, high quality video} */Class 3: fb0; b1g, /* {low quality voice, normal

quality voice} */Class 4: fb3g,Class 5: fc1; c2; c3; c4g,Class 6: fd1; d2; d3; d4; d5g,Class 7: f0; f2g



bi;assiðk; tÞ 2fbi;1; bi;2; . . . ; bi;j; . . . ; bi;Kig and 1 � k

� xiðtÞ. If bi;assiðk; tÞ < bi;req, we refer to the call as

degraded.

Let I(�) denote the indicator function defined as

follows:

IðexpÞ ¼ 1; if exp ¼ true

0; if exp ¼ false

�ð1Þ

Let xi;dðtÞ denote the number of degraded calls out

of xiðtÞ class-i calls at time t. Therefore, we have

xi;dðtÞ ¼XxiðtÞk¼1

Iðbi;assiðk; tÞ < bi;reqÞ; for i ¼ 1; 2; . . . ;m

ð2Þ

At time t, the instant degraded ratio (IDR) is defined

as:

IDRiðtÞ ¼ xi;dðtÞxiðtÞ ¼

PxiðtÞk¼1 Iðbi;assiðk; tÞ < bi;reqÞ

xiðtÞ ;

for i ¼ 1; 2; . . . ;m

ð3Þ

The amount of bandwidth degradation for call k of

class-i is bi;req � bi;assiðk; tÞ; if bi;assiðk; tÞ < bi;req. At

time t, the (IDD) is defined as:

IDDiðtÞ ¼ 1

xi;dðtÞ

XxiðtÞk¼1

½bi;req � bi;assiðk; tÞ�Iðbi;assiðk; tÞ < bi;reqÞ1

xi;dðtÞXxiðtÞ

k¼1ðbi;req � bi;1ÞIðbi;assiðk; tÞ < bi;reqÞ

¼XxiðtÞ

k¼1½bi;req � bi;assiðk; tÞ�Iðbi;assiðk; tÞ < bi;reqÞ

ðbi;req � bi;1ÞXxiðtÞ

k¼1Iðbi;assiðk; tÞ < bi;reqÞ

;

for i ¼ 1; 2; . . . ;m

ð4Þ

The denominator of IDDiðtÞ is for the normaliza-

tion purpose. Both IDDiðtÞ and IDRiðtÞ are random

processes. The DRiðtÞ and the (DDiðtÞ are defined as

the time averages of IDRiðtÞ and IDDiðtÞ respectively,and they reflect the observed history of the system.

DRið�Þ ¼ 1

�T

ð��T

XxiðtÞk¼1

Iðbi;assiðk; tÞ < bi;reqÞxiðtÞ dt

for i ¼ 1; 2; . . . ;m

ð5Þ

DDið�Þ

¼ 1

�T

ð��T

XxiðtÞk¼1

½bi;req � bi;assiðk; tÞ�Iðbi;assiðk; tÞ < bi;reqÞðbi;req � bi;1Þ

XxiðtÞk¼1

Iðbi;assiðk; tÞ < bi;reqÞdt

for i ¼ 1; 2; . . . ;m

ð6Þ

Here, �T is a time interval for measurement and �is a time variable. Both DRiðtÞ and DDiðtÞ are randomprocesses. Since events of xiðtÞ changes happen dis-

cretely in time, the above integrations become discrete

summations when implemented. Both DR and DD

take on values ranging from 0.0 to 1.0. The smaller the

values of DR and DD are, the better the QoS is.

3.2. Degradation Area: a New QoS Parameter

In order to provide better QoS, in this paper, a new

QoS parameter is designed to effectively characterize

the bandwidth degradation: DA, illustrated in Figure 2.

Figure 2 shows the relationship of the assigned

bandwidth and the required bandwidth, as well as

the DA. From Figure 2, we observe that DA fully

characterizes the bandwidth degradation. For class-i

(i ¼ 1; . . . ;m), the DAi is defined as follows.

DAið�Þ

¼ 1

�T

ð��T

XxiðtÞk¼1

½bi;req � bi;assiðk; tÞ�Iðbi;assiðk; tÞ < bi;reqÞðbi;req � bi;1Þ�xiðtÞ dt

ð7Þ

DAiðtÞði ¼ 1; . . . ;mÞ are random processes. Since

events of xiðtÞ changes happen discretely in time,

the above integrations become discrete summations

when implemented. The DA takes on values ranging

Fig. 2. Degradation area (DA).

224 Y. XIAO ET AL.


from 0.0 to 1.0. A smaller value of DA means better

QoS.

4. Measurement-based Call AdmissionControl (CAC)

As we have mentioned in ‘Introduction’, we will work

on two scenarios in this paper with different sets of

QoS parameters. In the first scenario, we will use DD

and DR. In the second scenario, we will use DA.

The CAC schemes in both scenarios are the same

except that different sets of QoS parameters are used.

4.1. The 1st Scenario: Measurement-Based CACwith DD and DR

We allow a handoff call to be always accepted. For a

new call request, the QoS requirements are upper

bounds in terms of the DR and the DD introduced in

the previous section. Let DRi;qos and DDi;qos denote

the pre-specified upper bounds of DR and DD re-

spectively. The proposed CAC scheme seeks to main-

tain the DR value and the DD value in wireless/mobile

networks to be statistically less than the predefined

values of DRi;qos and DDi;qos respectively. Momenta-

rily, the system can have DR larger than DRi;qos and/or

DD larger than DDi;qos but in the long run, the system

will be such that

DRið�Þ � DRi;qos ð8Þ

DDið�Þ � DDi;qos ð9Þ

We measure the system resource usage at regular

intervals (every �T time units). During each

measurement window, we calculate DD and DR using

Equations (5) and (6). As stated in the previous

section, the integrations become discrete summations

when implemented. Let us denote DDki and DR

ki as the

k-th measurement of DDi and DRi. The proposed

scheme takes into account the history of previous

measurements with different weights. A factor �,where 0 < � < 1, is introduced to reduce the impact

of the historical measurements. Initially, let

Pi;DDðlÞ ¼ DD1i and Pi;DRðlÞ ¼ DR1

i . For the jth mea-

surement, where j > 1,

Pi;DD ð jÞ ¼ �Pi;DD ð j� 1Þ þ ð1� �ÞDDji ð10Þ

Pi;DR ð jÞ ¼ �Pi;DR ð j� 1Þ þ ð1� �ÞDRji ð11Þ

The weight � determines how fast the estimated

average adapts to the new measurement. A smaller �results in a faster reaction to the network changes.

As a special case when � goes to zero, we have

Pi;DD ð jÞ ¼ DDji and Pi;DR ð jÞ ¼ DR

ji which are the

results of current measurement window. In the pro-

posed scheme, the effects of old measurements dis-

appear eventually. With a larger �, such a scheme can

reflect quickly the current status of the system.

We need to further consider that ongoing calls may

be non-uniformly distributed among the cells so that

averaging Pi;DD ð jÞ and ðPi;DR ð jÞÞ among neighbor-

ing cells is necessary. The averaging is conducted by

putting a larger weight on its own cell and smaller

weights on the neighboring cells. The measurement-

based CAC algorithm is shown in Table II, where

‘average’ means averaging over neighboring cells.

4.2. The 2nd Scenario: Measurement-Based CACwith DA

We allow a handoff call to be always accepted. For a

new call request, the QoS requirements are upper

bounds in terms of the DAi. Let DAi;qos denote the

pre-specified upper bound of DAi. The proposed CAC

scheme seeks to maintain the DAi value to be statis-

tically less than the predefined value of DAi;qos.

Momentarily, the system can have DAi larger than

DAi;qos, but in the long run, the system will be such

that

DAið�Þ � DAi;qos ð12Þ

We measure the system resource usage at regular

intervals (every�T time units). During each measure-

ment window, we calculate DAi as in Equation (7). As

stated in the previous section, the integrations become

discrete summations when implemented. Let DAki

denote the k-th measurement of DAi. The proposed

scheme takes into account the history of previous

measurements with different weights. A factor �,where 0 < � < 1, is introduced to reduce the impact of

the historical measurements. Initially, let Pi;DAð1Þ ¼DA1

i . For the jth measurement, where j > 1,

Pi;DA ðjÞ ¼ �Pi;DA ð j� 1Þ þ ð1� �ÞDAji ð13Þ

Table II. Measurement-based CAC with DD and DR.

1. If (Handoff arrival) Accepted;2. else if (average (Pi;DRð jÞÞ � DRi;QoS and average

(Pi;DDð jÞÞ � DDi;QoSÞ Accepted;3. else Rejected;



The weight � determines how fast the estimated average

adapts to the new measurement. A smaller � results in a

faster reaction to the network changes. As a special case

when � goes to zero, we have Pi;DAð jÞ ¼ DAji which is

the result of current measurement window. In the

proposed scheme, the effects of old measurements

disappear eventually. With a larger �, such a scheme

can reflect quickly the current status of the system.

We need to further consider that ongoing calls may

be non-uniformly distributed among the cells so that

averaging Pi;DAð jÞ among neighboring cells is neces-

sary. The averaging is conducted by putting a larger

weight on its own cell and smaller weights on the

neighboring cells. The measurement-based CAC al-

gorithm is shown in Table III, where ‘average’ means

averaging over neighboring cells.

5. Proportional Degradation Model

Definitions of fairness are very diverse in literature.

We argue for the case of providing fairness in terms of

proportionally degrading bandwidth of adaptive mul-

timedia calls from different service classes. In this

section, we discuss the proportional degradation

model and its advantages as well as disadvantages

compared to other models. In the next section, we

describe greedy bandwidth degradation adaptation

algorithms to approximate the proportional degrada-

tion model described in this section.

5.1. Proportional Degradation Model

It is expected that users with widely varied service

expectations will use wireless/mobile networks. Some

service users may like to pay a higher price for a better

service and other users may prefer to pay less money

for an acceptable service. This boils down to provi-

sioning differentiated QoS assurance to users. Two

types of approaches, IntServ and DiffServ, have

been proposed. In the context of wireless networks,

researchers [15–23] have looked into wireless channel

bandwidth scheduling mechanisms for providing ser-

vice assurance on a per-flow basis [16–18,20–21] or a

per-class basis [19]. In general, these mechanisms try

to guarantee a minimal amount of bandwidth (e.g. by

approximating weighted fair queuing scheduling) for

a flow or a service class. Moreover, some mechanisms

[17–19,22–23] also try to compensate for wireless

channel bit errors. While guaranteeing the minimal

available bandwidth to a service is important, the

availability of the minimum amount of bandwidth

alone may not be enough or may not even be desirable

for adaptive multimedia services, especially under

congestion conditions in which service users have to

adjust their bandwidth requirements and accept ser-

vices with degraded quality. Sometimes, it is more

desirable to provide such an assurance that higher

classes will always receive better services relative to

lower classes (i.e. providing service prioritization or

differentiation), but not entirely at the cost of provid-

ing no services or useless services to lower classes.

That is, it is desirable that service degradation to

different classes be regulated. In particular, we argue

for the case of proportionally degrading service qua-

lities (e.g. providing proportional bandwidth degrada-

tion across service classes). In this way, service

classes will be provided some sort of fairness in

service degradation. Therefore, fairness and prioriti-

zation both become important issues for network

operators that offer adaptive multiple classes of ser-

vices. However, fairness and prioritization cannot

always come hand in hand.

We believe that obsessive pursuit of either fairness

or prioritization is not reasonable. To achieve prior-

itization or differentiation only, we can use strict

prioritization approaches proposed in References

[25–36], in which higher classes are always serviced

before lower classes. The drawbacks of strict prior-

itization approaches are numerous. If higher classes

are persistently backlogged, lower classes may have

long starvation periods [37]. Moreover, strict prioriti-

zation schemes do not provide a tuning mechanism for

adjusting the quality spacing among classes and the

system operating point depends only on the load

distribution among classes [37]. On the other hand,

absolute fairness is not reasonable, nor is it desirable

since some service users do expect better services by

paying a higher price. Another possible approach to

providing prioritization is to allocate a larger amount

of bandwidth to higher classes, relative to the ex-

pected load in each class, for example approaches

based on the WFQ scheduler [38] can use different

weights for service differentiation. The main

drawback of WFQ based approaches is that higher

classes may provide worse QoS than lower cases in

shorter time-scales [37].

In order to achieve a balance between fairness

and prioritization among classes, we propose the

Table III. Measurement-based CAC with DA.

If (Handoff arrival) Accepted;else if (average (Pi;DAð jÞÞ � DAi;QoS) Accepted;else Rejected;

226 Y. XIAO ET AL.


proportional degradation service model, which guar-

antees the ratios of service degradation among classes.

In other words, service degradation incurred by classes

should be proportional to their predefined ratios, and

service users can select the class that best meets their

quality and pricing constraints. In the next section, we

will propose greedy bandwidth degradation adaptation

algorithms to approximate this proportional degrada-

tion model. Our proposed approach will avoid long

starvation periods for lower classes even if higher

classes are persistently backlogged, and at the same

time, higher classes in our proportional degradation

service model consistently receive better services than

lower classes do even on short time-scales. Moreover,

service providers can use our approach as a ‘tuning

knob’ to tune or adjust the quality spacing among

classes.

The proportional approach can also be coupled with

a pricing scheme to make higher classes more costly

than lower classes and a higher class will receive a

better service than a lower class. The amount of

service received by a class and the resulting QoS

perceived by an application depend on the current

network load in each class, the CAC decisions and the

resource allocation scheme. Users can either adapt

their needs based on the current network performance

level in their class or switch to a better class if their

cost constraints allow this transition [37].

Another point that we want to make is that a fair

bandwidth allocation algorithm may not produce fair

bandwidth degradation. First of all, a fair bandwidth

allocation algorithm cannot fairly allocate the DR

among classes. Moreover, while a good fair band-

width allocation algorithm might improve the fairness

of the DD among classes, it cannot achieve good

bandwidth degradation fairness since it does not

take achieving fair degradation as the major goal.

Finally, a fair bandwidth allocation algorithm cannot

fairly allocate the DA among classes. Therefore, a

study on fair bandwidth degradation as proposed in

this paper is important.

We assume that given m classes, class-i is better or

at least no worse than class-(i� 1) for 1 � i � m in

terms of the two QoS parameters. A generic descrip-

tion of the proportional fairness model follows.

Among m classes of services, constraints of the

following form should be satisfied for all pairs of

classes [37,40]:

QoSiðt; t þ �ÞQoSjðt; t þ �Þ ¼

Qi

Qj

ð14Þ

where QoSiðt; t þ �Þ is the mean service obtained by

class-i during the time period ðt; t þ �Þ; Qi is the

predefined service parameter of class-i, and Q1 < Q2

< � � � < Qm are the generic quality differentiation

parameters [37]. � is the measurement period and

can be reconfigured.

5.2. The 1st Scenario: Proportional DD and DR

A better network service is one with a lower DD value

and a lower DR value. Particularly, we consider

QoS parameters DR and DD to provide propor-

tional adaptive degradation multimedia services. Let

QoSiðt; t þ �Þ ¼ 1=DDiðt; t þ �Þ or QoSiðt; t þ �Þ ¼1=DRiðt; t þ �Þ. Then, Equation (14) can be written

as:

DDiðt; t þ �ÞDDjðt; t þ �Þ ¼

Ri;DD

Rj;DDð15Þ

DRiðt; t þ �ÞDRjðt; t þ �Þ ¼

Ri;DR

Rj;DRð16Þ

where R1;DD > R2;DD > � � � > Rm;DD, R1;DR > R2;DR

> � � � > Rm;DR, and Ri;DDð1 � i � mÞ and Ri;DR

ð1 � i � mÞ are the ratios called the DD differentia-

tion parameters and the DR differentiation parameters

respectively. The basic idea is that, even though the

actual quality level of each class will vary with the

class loads, the quality degradation ratio among

classes will remain fixed and controllable by the

network operator, independent of the class load. It is

also predictable since, if the value of � is sufficiently

small, higher classes are consistently better than lower

classes even on short time-scales. We can produce

½ðm� 1Þm�=2 predefined DD ratios as defined in

Equation (15) and shown in the upper right corner

of the following table.

R1;DD : R2;DD R1;DD : R3;DD R1;DD : R4;DD . . . R1;DD : Rm;DD

� R2;DD : R3;DD R2;DD : R4;DD . . . R2;DD : Rm;DD

� � R3;DD : R4;DD . . . R3;DD : Rm;DD

� � . . .� � � � Rm�1;DD : Rm;DD

266664

377775 ð17Þ



Similarly, we can produce ½ðm� 1Þm�=2 predefined

DR ratios as defined in Equation (16) and shown in the

upper right corner of the following table.

Note that the predefined values of Ri;DD

ð1 � i � mÞ and Ri;DRð1 � i � mÞ are different from

the predefined values of DDi;QoSð1 � i � mÞ and

DRi;QoS (1 � i � mÞ respectively. The former ones

are the values for QoS ratios and the latter ones are the

QoS upper bounds used by CAC. In other words,

while the measurement-based QoS parameters have

upper bounds for CAC, their relative ratios should also

be kept by predefined ratios for all pairs of classes.

5.3. The 2nd Scenario: Proportional DA

A better network service is one with a lower DAvalue.

Particularly, we consider QoS parameters DA to

provide proportional adaptive degradation multimedia

services. Let QoSiðt; t þ �Þ ¼ 1=DAiðt; t þ �Þ. Then,Equation (14) can be written as:

DAiðt; t þ �ÞDAjðt; t þ �Þ ¼

Ri;DA

Rj;DAð19Þ

where R1;DA > R2;DA > � � � > Rm;DA, and

Ri;DAð1 � i � mÞ are the ratios called the DA differ-

entiation parameters. We can produce ½ðm� 1Þm�=2predefined DA ratios as defined in Equation (15) and

shown in the upper right corner of the following table.

6. Bandwidth Degradation AdaptationAlgorithms

A bandwidth degradation adaptation algorithm deci-

des how to adjust the calls’ bandwidth in a cell

adaptively. The algorithm is activated whenever there

is a call arrival acceptance event or a call departure

event. With respect to different QoS objectives,

several bandwidth adaptation algorithms [1–2,9–11]

have been proposed and studied. In the first scenario,

our objectives are to minimize DR and DD at any time

instant, to minimize DD with a higher priority than to

minimize DR, to guarantee the ratios of service

degradation among classes, and to efficiently utilize

the system resource. The reason to place a higher

priority in minimizing DD is that, from the propor-

tional service provisioning point of view, it is more

important to fairly distribute bandwidth degradation

among all classes and all calls with low degradation

degrees than to reduce the degradation ratio by letting

a few calls having high degradation degrees. In the

second scenario, our objectives are to minimize DA at

any time instant, to guarantee the ratios of service

degradation among classes and to efficiently utilize

the system resource.

Ideally, every call in a cell should be allocated the

maximum bandwidth (bi;Ki) whenever possible. How-

ever, if a cell is overloaded, some of the calls in the

cell might receive a bandwidth, which is lower than

bi;Kifor a call in class-i. In other words, if a new call

or a handoff call arrives, some of the calls already in

the cell might be forced to lower their bandwidth (the

minimum bandwidth is bi;1 for a call in class-i) to

accommodate the newly arrived call. On the other

hand, when a call completes or handoffs to other cells,

some of the remaining calls in the cell might increase

their bandwidths (the maximum bandwidth is bi;Kifor

a call in class-i). We can use a long couch with a fixed

capacity in a meeting room as an analogy. We assume

that this couch is the only place where people can sit.

R1;DR : R2;DR R1;DR : R3;DR R1;DR : R4;DR . . . R1;DR : Rm;DR

� R2;DR : R3;DR R2;DR : R4;DR . . . R2;DR : Rm;DR

� � R3;DR : R4;DR . . . R3;DR : Rm;DR

� � . . .� � � � Rm�1;DR : Rm;DR

266664

377775 ð18Þ

R1;DA : R2;DA R1;DA : R3;DA R1;DA : R4;DA . . . R1;DA : Rm;DA

� R2;DA : R3;DA R2;DA : R4;DA . . . R2;DA : Rm;DA

� � R3;DA : R4;DA . . . R3;DA : Rm;DA

� � � . . .� � � � Rm�1;DA : Rm;DA

266664

377775 ð20Þ

228 Y. XIAO ET AL.


If there are more people coming to the meeting room,

everyone will sit much closer together so that new-

comers can be accommodated. If someone leaves, the

people remaining on the couch get more space.

We propose K-level proportional degradation band-

width adaptation algorithms, which use above ideas

to allocate, increase and decrease bandwidth for a

call and at the same time provides proportional

bandwidth degradation if necessary. If we assume a

call’s bandwidth in class-i is bi,j, where bi;j 2 fbi;1;bi;2; . . . ; bi; j; . . . ; bi;Ki

g, we call it one level of decre-

ment if its new bandwidth becomes bi,j-1 and we call it

one level of increment if its new bandwidth becomes

bi,jþ 1. A call’s bandwidth can have at most K levels

of increments and decrements, where K¼max (K1,

K2; . . . ;Km).

Since the bandwidth value of each adaptive multi-

media is not continuous, it is not easy to keep exact

ratios. We need to approximate the proportional

degradation model. Therefore, the bandwidth adapta-

tion algorithm should adjust calls’ bandwidths to

approximate the ratios as much as possible.

6.1. The 1st Scenario: Bandwidth DegradationAdaptation Algorithm with DD and DR

For any two classes, define the DD difference as the

absolute value of the difference between their pre-

defined DD ratio in Equation (17) and their actual

ratio, and define the DR difference as the absolute

value of the difference between their predefined DR

ratio in Equation (18) and their actual ratio. We call

the DD difference or the DR-difference HIGH if their

actual ratio is larger than their predefined ratio,

otherwise we call it LOW.

The pseudo code of KL-PBA is shown in Table IV.

In this algorithm, there are at most K-levels of

bandwidth increment and K-levels of bandwidth

decrement. We explain the method as follows. For

accepted arrivals, if there is enough available band-

width, allocate the requested bandwidth. Otherwise,

some calls’ bandwidth will be decreased. Upon a call

departure, the amount of available bandwidth will

increase and we can then increase the bandwidth of

some calls in the cell. In either case, following steps

take place. First find a class that is constrained by the

proportional degradation Equations (15) and (16).

Then find a call in the chosen class. Finally, increase

or decrease the chosen call by one level. The algo-

rithm considers DD proportional fairness and DR

proportional fairness alternatively. The design philo-

sophy of the algorithm is to minimize DR and DD at

Table IV. K-level proportional degradation adaptation algorithmwith DD and DR.

Accepted Arrivals in class-iif (A� bi,req) Assign_as_most (bi;Ki

);while (A< bi,req){/* Finding the chosen class according to DRproportional fairness */

1. Find a pair of classes, whose DR difference is maximumamong all pairs of classes;

2. if (the chosen pair’s DR difference is HIGH) Choose the2nd class in the pair;

3. else Choose the 1st class in the pair;4. Find a call in the chosen class who has the highestbandwidth, which is larger than bi,1;

5. if (fail to find the call), break;6. else Decrease this call with one level;7. if (A� bi,req) break;/* Finding the chosen class according to DDproportional fairness */

8. Find a pair of classes, whose DD difference is maximumamong all pairs of classes;

9. if (the chosen pair’s DD difference is HIGH) Choose the2nd class in the pair;

10. else Choose the 1st class in the pair;11. Find a call in the chosen class who has the highest

bandwidth, which is larger than bi,1;12. if (fail to find the call), break;13. else Decrease this call with one level;}if (A< bi,1) Reject the call;else Assign_as_most (bi,req);Departureswhile (A> 0){/* Finding the chosen class according to DDproportional fairness */

14. Find a pair of classes, whose DD difference is maximumamong all pairs of classes;

15. if (the chosen pair’s DD difference is LOW) Choose the2nd class in the pair;

16. else Choose the 2nd class in the pair;17. Find a call in the chosen class who has the

lowest bandwidth;18. Increase this call in one level;19. if (fail to increase because there is not enough bandwidth

to increase one level) break;20. if (A¼ 0) break;/* Finding the chosen class according to DRproportional fairness */

21. Find a pair of classes, whose DR difference is maximumamong all pairs of classes;

22. if (the chosen pair’s DR difference is LOW) Choose the2nd class in the pair;

23. else Choose the 1st class in the pair;24. Find a call in the chosen class who has the

lowest bandwidth;25. Increase this call in one level;26. if (fail to increase because there is not enough

bandwidth to increase one level) break;}Assign_at_most(int level){Assign a bandwidth as much as possible but at most level;}Legend: A: available bandwidth.



any time, to minimize DD with a higher priority than

to minimize DR, to guarantee the ratios of service

degradation among classes and to efficiently utilize

the system resource.

Three aspects are taken in the algorithm to insure

that a higher priority is put on DD. First, if a chance

arises to increase bandwidth, DD is always the first to

be considered (lines 14–20 in Table IV) and if a

chance arises to decrease bandwidth, DD is always

the last to be considered (lines 8–13 in Table IV).

Second, there are at most K-levels of bandwidth

increment and K-levels of bandwidth decrement in

the algorithm, and each time only one level of decre-

ment or one level of increment will happen. DD is

directly affected by the choice of one level of incre-

ment or decrement. DR is only affected through DD.

Otherwise the goal would be to remove one degraded

call by increasing more levels at one time as much as

possible. Finally, in the algorithm, the policy of

finding the lowest/highest bandwidth in the chosen

class is to favor DD instead of DR. Therefore, our

implementation favors DD more than DR. In such a

way, the ratios for DD are easier to keep than for DR.

A complexity analysis of the bandwidth adaptation

algorithm is given as follows. In the worst case,

finding the chosen class needs O(m2). Finding the

chosen call takes O(K), where K¼max (K1, K2; . . . ;Km), and the loop needs O(K). The overall time

complexity is O(m2KþK2).

6.2. The 2nd Scenario: Bandwidth DegradationAdaptation Algorithm with DA

For any two classes, define the DA difference as the

absolute value of the difference between their pre-

defined DA ratio in Equation (20) and their actual

ratio. We call the DA difference HIGH if their actual

ratio is larger than their predefined ratio; otherwise we

call it LOW. The pseudo code of the algorithm is

shown in Table V. In this algorithm, there are at the

most K-levels of bandwidth increment and K-levels of

bandwidth decrement.

A complexity analysis of the bandwidth adaptation

algorithm is similar to that in Section 6.1.

7. Performance Evaluation

In this section, we present simulation results de-

monstrating how well the proportional degradation

fairness model works. The performance metrics in-

clude utilization, DR, DD, DA, (New) CBP, HDP and

Number of Calls (NC) in the system. In Section 7.1,

we study the 1st scenario, proportional DD and DR. In

Section 7.2, we study the 2nd scenario, proportional

DA.

Table V. K-level proportional degradation adaptation algorithm withDA.

Accepted Arrivals in class-iif (A� bi,req) Assign_as_most (bi;Ki

Þ;while (A< bi,req){ /* Finding the chosen class according to DA

proportional fairness */27. Find a pair of classes, whose DA difference is maximum

among all pairs of classes;28. if (the chosen pair’s DA difference is HIGH) Choose

the 2nd class in the pair;29. else Choose the 1st class in the pair;30. Find a call in the chosen class who has the highest

bandwidth, which is larger than bi,1;31. if (fail to find the call), break;32. else Decrease this call with one level;33. if (A� bi,req) break;

}if (A< bi,1) Reject the call;else Assign_as_most (bi,req);Departureswhile (A> 0){ /* Finding the chosen class according to DA

proportional fairness */34. Find a pair of classes, whose DA difference is maximum

among all pairs of classes;35. if (the chosen pair’s DA difference is LOW) Choose

the 2nd class in the pair;36. else Choose the 1st class in the pair;37. Find a call in the chosen class who has the

lowest bandwidth;38. Increase this call in one level;39. if (fail to increase because there is not enough

bandwidth to increase one level) break;40. if (A¼ 0) break;

}Assign_at_most (int level){Assign a bandwidth as much as possible but at most level;}Legend: A: available bandwidth.

Table VI. Simulation parameters.

Classes Bandwidth set bi,req 1/hi 1/�i Ri.DR Ri.DD DRi,qos DDi,qos

1 {1,3,5,7,9} 7 5 sec 60 sec 60% 50% 0.1 0.42 {2,4,6,8,10} 8 5 sec 60 sec 30% 30% 0.1 0.43 {3,5,7,9,11} 9 5 sec 60 sec 10% 20% 0.1 0.4

230 Y. XIAO ET AL.


7.1. The 1st Scenario: Proportional Degradationwith DD and DR

Simulation parameters are illustrated in Table VI. 1/hiand 1=�i are the mean interarrival time for handoff

calls and the mean call holding time respectively. The

total number of channels is 1000; simulation time is

20 000 sec; the weight for the admission control is

�¼ 0.3; and�T (¼ 5 sec.) is the measurement period.

From the table, the ratio of R1:DR : R2:DR : R3:DR is

60%:30%:10%¼ 6:3:1, and the ratio of R1:DR :R2:DD : R3:DD is 50%:30%:20%¼ 5:3:2. In our simu-

lation, we let new call arrival rates to be the same, i.e.

�1 ¼ �2 ¼ �3 ¼ �.Figure 3 shows utilization over simulation time

with different � values. As illustrated in the figure,

utilization increases and becomes closer to 1 as �increases. Furthermore, we observe that the variance

of utilization becomes smaller as � increases. The

variance is caused by arrival and departure events.

When a call is finished (departure), the bandwidth

resource is released; the bandwidth resource can be

used by the degraded calls. When a call arrives, some

unused bandwidth, if available, can be allocated to the

newly arrived call. Bandwidth obtained by squeezing

the existing calls to accommodate the newly arrived

call does not affect utilization when the call arrives,

but may affect utilization a little bit when the call

finishes. We also observe that there are more times

when utilization is close to 1 than that far way from 1.

This indicates that the proposed scheme is very

efficient in terms of utilization.

Figure 4 shows average utilization over �. As

illustrated in the figure, utilization increases as �increases. When �> 0.5, utilization becomes almost

constant and close to 1. Figure 3 and 4 indicate that

the proposed adaptive approach efficiently utilizes the

bandwidth resource.

Figure 5 shows DR over simulation time with

different � values. Class 1 (green) is the lowest class,

and class 3 (blue) is the highest class. The goal is to

achieve the ratio of R1.DR:R2.DR:R3.DR¼ 6:3:1. As

illustrated in the figure, for all � values, class 1 has

the highest DR and class 3 has the lowest DR. When �is small (�¼ 0.3), since there are still avai-

lable resource, DRs are zero. As � increases, the

differences among classes become more and more

obvious. This also indicates that proportional degra-

dation is more useful and performs well under heavy

load condition.

Figure 6 shows average DR over �. The goal is to

achieve the ratio of R1.DR:R2.DR:R3.DR¼ 6:3:1. As

illustrated in the figure, when � is small (�< 0.4),

the difference among classes is not obvious. As �increases, the difference becomes more and more

obvious. We observe that when �> 0.5, the propor-

tional ratio (6:3:1) is well achieved, especially under

heavy load condition. This also indicates that propor-

tional degradation is more useful and performs well

under heavy load condition.

Fig. 3. Utilization over simulation time with different � values.



Fig. 4. Utilization over � .

Fig. 5. Degradation ratio (DR) over time with different � values.

232 Y. XIAO ET AL.


Figure 7 shows average DD over �. The goal is to

achieve the ratio of R1.DD:R2.DD:R3.DD¼ 5:3:2. As illu-

strated in the figure, the algorithm does not achieve

good proportional at all. The reason is stated as follows.

DD is defined in such a way that largely depends on the

‘degraded level’ of discrete values. If all existing calls

are degraded at one level, DD should be around 30%.

There are not finer-grained increasing/decreasing DD

values. Furthermore, to achieve proportional DD

is difficult due to the nature of discrete values of

Fig. 6. DR over � .

Fig. 7. Degradation degree (DD) over � .



bandwidth. On the other hand, if the continuous band-

width is adopted, much better proportionality will be

achieved. Another possible way to achieve proportion-

ality in a discrete bandwidth model is to achieve

proportional degraded levels among classes instead

of proportional DD. In the 2nd scenario, we show that

proportional DA is a much better choice.

Figure 8 shows average HDP over �. As illustratedin the figure, HDPs for all the classes are zeros. This

contributes the admission control algorithm and the

degradation adaptation algorithm. As we claimed in

previous sections, HDP can be very small or equal to

zero in this adaptive scheme. Therefore, for adaptive

multimedia services, the problems of handoff drop-

ping and forced termination in traditional wireless

networks are converted to the problem of bandwidth

degradation caused by the adaptation. In the CAC

scheme, handoff calls will be accepted all the time. A

handoff call can be dropped only when there is not

enough available bandwidth (<bi;1) and all existing

calls cannot be degraded.

Figure 9 shows the CBP over �. Although our

proportional model does not aim at CBP, some dif-

ference among CBPs in different classes exists.

Figure 10 shows the NC in the system over �. Asillustrated in the figure, NC increases as � in-

creases when �< 0.5. As � increases and �> 0.5,

NC in class 3 continuously increases whereas NC in

class 1 and NC in class 2 either remain same or

decrease. This may be the side effect of the goal of

proportional DR and DD.

7.2. The 2nd Scenario: Proportional Degradationwith DA

In this subsection, we study the performance of

proportional degradation with DA. Simulation para-

meters are illustrated in Table VII. 1/hi and 1/�i are the

mean interarrival time for handoff calls and the mean

call holding time respectively. The total number of

channels is 1000; simulation time is 20 000 sec; the

weight for the admission control is �¼ 0.3; and �T

(¼ 5 sec.) is the measurement period. From the table,

the ratio of R1.DA:R2.DA:R3.DA is 60%:30%:10%¼6:3:1. In our simulation, we let new call arrival rates

to be the same, i.e. �1 ¼ �2 ¼ �3 ¼ �.Figure 11 shows utilization over simulation time

with different � values. As illustrated in the figure,

utilization increases and becomes closer to 1 as �increases. Furthermore, we observe that the variance

of utilization becomes smaller as � increases. The

variance is caused by arrival and departure events.

When a call is finished (departure), the bandwidth

resource is released; the bandwidth resource can be

used by the degraded calls. When a call arrives, some

unused bandwidth, if available, can be allocated to the

newly arrived call. Bandwidth obtained by squeezing

Fig. 8. Handoff dropping probability (HDP) over � .

234 Y. XIAO ET AL.


the existing calls to accommodate the newly arrived call

does not affect utilization when the call arrives, but may

affect utilization a little bit when the call finishes. We

also observe that there are more times when utilization

is closer to 1 than that far way from 1. This indicates

that the proposed scheme is very efficient in terms of

utilization. Comparing this figure with Figure 3, we

observe that utilization goes close to 1 much quickly. In

other words, utilization of proportional DA is a little

better than that of proportional DD and DR.

Fig. 9. (New) call blocking probability (CBP) over � .

Fig. 10. Number of calls (NC) in the system over � .



Figure 12 shows average utilization over �. As illu-

strated in the figure, utilization increases as � in-

creases. When �> 0.5, utilization becomes almost

constant and close to 1. Figures 11–12 indicate that

the proposed adaptive approach efficiently utilizes the

bandwidth resource.

Figure 13 shows DA over simulation time with

different � values. Class 1 (green) is the lowest class,

and class 3 (blue) is the highest class. The goal is to

achieve the ratio of R1.DA:R2.DA:R3.DA¼ 5:3:2. As

illustrated in the figure, for all � values, class 1 has

the highest DA, and class 3 has the lowest DA. When

� is small (�¼ 0.3), since there are still available

resource, DAs are zero. As � increases, the differences

among classes become more and more obvious. This

also indicates that proportional degradation is more

useful and performs well under heavy load conditions.

It is shown that the proportional ratios are approxi-

mated very well using our algorithm. This indicates

that the proposed algorithm is very effective. Figure 13

also shows that the QoS requirements (DAi�DAi,qos)

are basically satisfied.

Figure 14 shows average DA over �. The goal is toachieve the ratio of R1.DA:R2.DA:R3.DA¼ 5:3:2. As

illustrated in the figure, when � is small (�< 0.5),

the difference among classes is not obvious. As �increases, the difference becomes more and more

obvious. We observe that when �> 0.5, the propor-

tional ratio (5:3:2) is well achieved, especially under

heavy load conditions. This also indicates that propor-

tional degradation is more useful and performs well

under heavy load conditions. Comparing proportional

DA (the 2nd scenario) with proportional DD and DR

(the 1st scenario), much better proportional DA is

achieved. In other words, DA outperforms DD and DR

in terms of proportional degradation.

Figure 15 shows average HDP over �. As illustratedin the figure, HDPs for all the classes are zeros. This is

Table VII. Simulation parameters.

Classes Bandwidth set bi,req 1/hi 1/�i Ri.DA DDi,qos

1 {1,3,5,7,9} 7 8 sec 60 sec 50% 0.32 {2,4,6,8,10} 8 8 sec 60 sec 30% 0.33 {3,5,7,9,11} 9 8 sec 60 sec 20% 0.3

Fig. 11. Utilization over simulation time with different � values.

236 Y. XIAO ET AL.


due to the admission control algorithm and the de-

gradation adaptation algorithm. As we claimed in

previous sections, HDP can be very small or equal

to zero in this adaptive scheme. Therefore, for adap-

tive multimedia services, the problems of handoff

dropping and forced termination in traditional wire-

less networks are converted to the problem of band-

width degradation caused by the adaptation. In the

CAC scheme, handoff calls will be accepted all the

time. A handoff call can be dropped only when there

is not enough available bandwidth (< bi;1) and all

existing calls cannot be degraded.

Fig. 12. Utilization over � .

Fig. 13. DA over time with different � values.



Figure 16 shows the CBP over �. The differences

among classes are minor. The figure shows that CBP is

smaller than 0.35. Compared with Figure 9, the 2nd

scenario has a smaller CBP.

Figure 17 shows the NC in the system over �. Asillustrated in the figure, NC increases as � increases

when �< 0.5. As � increases and �> 0.5, NC in

class 3 and NC in class 2 continuously increase

whereas NC in class 1 decreases a little. This may

be the side effect of the goal of proportional DA.

Compared with Figure 10, NC is a little better in this

scenario.

Fig. 14. DA over � .

Fig. 15. HDP over � .

238 Y. XIAO ET AL.


We also carried out many experiments on variants of

our degradation adaptation algorithms. The idea is to

let the algorithms become less greedy. One approach is

that when the traffic is heavy (but there is still available

bandwidth) and a call arrives, if the call is accepted, we

do not assign the call with the maximum bandwidth,

but a little smaller bandwidth—to be conservative.

The second approach is that when the traffic is heavy

and a call is finished, the available bandwidth is not

immediately assigned to degraded calls. The purpose

Fig. 16. (New) CDP over � .

Fig. 17. NC in the system over � .



of these approaches is to decrease frequent small

changes. Our experiments indicate that DR can be

improved a little whereas other parameters such as

utilization and NC become worse.

Figure 18 shows effects of � on DA. � equals 0 in

the left two sub-figures, and equals 0.9 in the right two

sub-figures. Lower sub-figures are for average values,

and higher sub-figures are for instant values for a fixed

�. From the sub-figures, we observe that � has no

effects average values, but it does have effects on

instant values. We observe that a larger � causes

smaller variances of DA. Please note that such ob-

servations are based on our simulation traffic model,

which is quite ‘regular’. If the simulation traffic model

becomes much more ‘irregular’, we expect that � has

more effects on �. Similar studies about effects of�T

can also be studied. We omit these due to limited

space.

8. Conclusions

In this paper, we proposed and studied proportional

degradation services for multiple classes of adaptive

multimedia services. The proposed proportional de-

gradation fairness model guarantees the proportional

bandwidth degradation among different classes of

services. Two proportional degradation scenarios are

studied in this paper. In the first scenario, we study the

proportional degradation provisioning with two QoS

parameters for adaptive multimedia: the DR and the

DD. In the second scenario, we study the proportional

degradation provisioning with a new QoS parameter

for adaptive multimedia: the DA. For each scenario,

based on the QoS parameters, proportional degrada-

tion adaptation algorithms are proposed to approxi-

mate the proportional degradation model, to fairly

adapt calls’ degradations, to utilize the system re-

source efficiently, as well as to optimize QoS para-

meters. Simulation results show the following

observations.

� Proportional DA has been well achieved.

� Proportional DR has been well achieved, whereas

proportional DD has not been well achieved. The

reason that proportional DD has not been well

achieved is due to the nature of discrete values of

bandwidth. If continuous values of bandwidth are

adopted, proportionality will be achieved much

better. Another possible way to achieve proportion-

ality in a discrete bandwidth model is to achieve

proportional degraded levels among classes instead

of proportional DD. However, we believe that

proportional DA is a better choice.

� DA outperforms DD and DR in terms of achieving

proportional degradation. The proportional DA

scheme is better than the proportional DD and

DR scheme in terms utilization, call blocking

Fig. 18. Effects of � on DA.

240 Y. XIAO ET AL.


probability, numbers of calls in the system, achieved

proportions on QoS parameters (DA or DD and DR)

and QoS bounds on them (DA or DD and DR).

� The proposed degradation adaptation algorithms

fairly adapt calls’ degradations, to utilize the sys-

tem resource efficiently, as well as to optimize QoS

parameters.

� HDPs are almost zero. Therefore, for adaptive

multimedia services, the problems of handoff drop-

ping and forced termination in traditional wireless

networks are converted to the problem of band-

width degradation caused by the adaptation

� Proportional degradation is more useful and per-

forms well under heavy load conditions.

� We observe that a larger � causes smaller variances

of DA.

References

1. Naghshineh M, Willebeek-LeMair W. End-to-end QoS provi-sioning in multimedia wireless/mobile networks using aadaptive framework. IEEE Communications Magazine 1997;35(11): 72–81.

2. Kwon T, Choi Y, Bisdikian C, Naghshineh M. QoS provision-ing for adaptive multimedia in wireless/mobile networks.Wireless Network 2003; 9(1): 51–59.

3. Shacham N. Multipoint Communication by HierarchicallyEncoded Data. In Proceedings of IEEE INFOCOM 1996;2107–2114.

4. McCanne S, Vetterli M, Jacobson V. Low complexity videocoding for receiver-driven layered multicast. IEEE Journal onSelected Areas in Communications 1997; 15(6): 983–1001.

5. Hartung J, Jacquin A, Pawlyk J, Shipley K. A Real-timeScalable Video Codec for Collaborative Applications overPacket Networks. In Proceedings of ACM Multimedia 1998;419–426.

6. Chakrabarti S, Wang R. Adaptive Control for Packet Video. InProceedings of IEEE International Conference on MultimediaComputing and Systems 1994; 56–62.

7. Duffield N, Ramakrishnan K, Reibman A. SAVE: an algorithmfor smoothed adaptive video over explicit rate networks. IEEE/ACM Transaction on Networking 1998; 6(6): 717–728.

8. Yeadon N. Filters: QoS support mechanisms for multipeer com-munications. IEEE Journal on Selected Areas in Communications1996; 14(7): 1245–1262.

9. Xiao Y, Chen CLP, Wang Y. Optimal admission control formulti-class of wireless adaptive multimedia services. IEICETransaction on Communications 2001; E84-B4 (4): 795–804.

10. Talukdar AK, Badrinath BR, Acharya A. Rate AdaptationSchemes in Networks with Mobile Hosts. In Proceedings ofACM/IEE MobiCom 1998.

11. Xiao Y, Chen CLP, Wang B. Bandwidth degradation QoSprovisioning for adaptive multimedia in wireless/mobile net-works. Computer Communications 2002; 25(13): 1153–1161.

12. Alwan A, Bagrodia R, Bambos N, Gerla M, Kleinrock L, ShortJ, Villasenor J. Adaptive mobile multimedia networks. IEEEPersonal Communications 1996; 34(4): 34–51.

13. Bharghavan V, Lee K, Lu S, Ha S, Li J, Dwyer D. The TIMELYadaptive resource management architecture. IEEE PersonalCommunication Magazine 1998; 5(8): 34–49.

14. Lee K. Adaptive Network Support for Mobile Multimedia. InProceedings of ACM MobiCom 1995; 62–74.

15. Cao Y, Li VOK. Scheduling Algorithms for BroadbandWireless Networks. In Proceedings of the IEEE 2001; 89(1):76–87.

16. Bhagwat P, Krishna A, Tripathi S. Enhancing Throughputover Wireless LANs Using Channel State DependentPacket Scheduling. In Proceedings of INFOCOM 1996;1133–1140.

17. Ng STE, Stoica I, Zhang H. Packet Fair Queueing Algorithmsfor Wireless Networks with Location-Dependent Errors. InProceedings of INFOCOM 1998; 1103–1111.

18. Lu S, Bharghavan V. Fair scheduling in wireless packet net-works. IEEE/ACM Transactions on Networking 1999; 7(4):473–489.

19. Fragouli C, Sivaraman V, Srivastava M. Controlled MultimediaWireless Link Sharing via Enhanced Class-Based Queueingwith Channel-State Dependent Packet Scheduling. In Proceed-ings of INFOCOM 1998; 572–580.

20. Ramanathn P, Agrawal P. Adapting Packet Fair QueueingAlgorithms to Wireless Networks. In Proceedings of ACM/IEEE MOBICOM 1998; 1–9.

21. Lu S, Bharghavan V, Srikant R. A Wireless Fair ServiceAlgorithm for Packet Cellular Networks. In Proceedings ofthe ACM SIGCOMM 1997; 63–74.

22. Ckhardt DA, Steenkiste P. Effect-Limited Fair (ELF) Schedul-ing for Wireless Networks. In Proceedings of INFOCOM 2000;1097–1106.

23. Lin P, Benssou B, Ding QL, Chua KC. CS-WFQ: a WirelessFair Scheduling Algorithm for Error-Prone Wireless Channels.In Proceedings of Computer Communications and Networks2000; 276–281.

24. Lou H, Medvedev P, Cheng J, Lu S. A Self-CoordinatingApproach to Distributed Fair Queueing in Ad Hoc WirelessNetwork. In Proceedings of IEEE INFOCOM 2001.

25. Bharghavan V, Lu S, Nandagopal T. Fair queuing in wirelessnetworks: issues and approaches. IEEE Personal Communica-tions 1999; 1: 44–53.

26. Lu S, Nandagopal T, Bharghavan V. A Wireless Fair ServiceAlgorithm For Packet Celluler Networks. In Proceedings ofACM MobiCom 1998; 10–20.

27. Noorman JR, Lockwood JW. Multiclass Priority Fair Queuingfor Hybrid Wired/Wireless Quality of Service Support. InProceedings of ACM WoWMoM 1999; 43–50.

28. Ramanathan P, Agrawal P. Adapting Packet Fair QueueingAlgorithms to Wireless Networks. In Proceedings of MobiCom1998; 1–9.

29. Lu S, Lee KW, Bharghavan P. Adaptive Service in MobileComputing Environments. In Proceedings of IWQoS 1997.

30. Kwon T, Choi Y, Das S. Bandwidth adaptation algorithmsfor adaptive multimedia services in mobile cellularnetworks. Wireless Personal Communications 2002; 22(3):337–357.

31. Noble BD, Satyanarayanan M. Experiences with adaptivemobile applications in Odyssey.Mobile Networks and Applica-tions 1999; 4: 245–254.

32. Corner MD, Noble BD, Wasserman KM. Fugue: Time Scalesof Adaptation in Mobile Video. SPIE Proceedings of Multi-media Computing and Networking 2001; 4312.

33. Margaritidis M, Polyzos GC. MOBIWEB: Enabling AdaptiveContinuous Media Applications over 3G Wireless Links. IEEEPersonal Communications Magazine 2000; 7(6): 36–41.

34. Angin Q, Campbell AT, Kounavis ME, Liao R. TheMOBIWARE toolkit: programmable support for adaptive mo-bile networking. IEEE Personal Communications 1998; 5(4):32–44.

35. Ayyagari D, Ephremides A. Admission control with priorities:approaches for multi-rate wireless systems. Mobile Networksand Applications 1999; 4: 209–218.



36. Kwon T, Kim S, Choi Y, Naghshineh M. Threshold-type calladmission control in wireless/mobile multimedia networksusing a prioritized adaptive framework. IEE Electronics Letters2000; 36(9): 1–2.

37. Dovrolis C, Stiliadis D, Ramanathan P. A case for relativedifferential services and the proportional differentiation model.IEEE Network 1999; 9: 26–34.

38. Demers A, Keshav S, Shenker S. Analysis and simulation of afair queuing algorithm. Internetworking: Research and Experi-ence 1990; 1(1): 3–26.

39. Lin YB, Chlamtac I.Wireless and Mobile Network Architectures.Wiley, 2001.

40. Leung M, Lui J, Yau D. Adaptive proportional delaydifferentiated services: characterization and performance eva-luation. IEEE/ACM Transactions on Networking 2001; 9(6):801–817.

41. Ho C, Lea C. Improving call admission policies in wirelessnetworks. Wireless Networks 1999; 5: 257–265.

42. Sidi M, Starobinski D. New Call Blocking Versus HandoffBlocking in Cellular Networks. In Proceedings of IEEE IN-FOCOM 1996; 35–42.

43. Ross SM. Introduction to probability methods. AcademicPress: London, 1997.

Authors’ Biographies

Yang Xiao received his Ph.D. inComputer Science and Engineeringfrom Wright State University, Day-ton, Ohio, U.S.A. He had been asoftware engineer, a senior softwareengineer and a technical lead, work-ing in the computer industry from1991 to 1996. From 1996 to 2001, hehad been awarded the DAGSI Ph.D.fellowship. From August 2001 toAugust 2002, he worked at Micro

Linear as an MAC architect involving the IEEE 802.11standard enhancement work. Since August 2002, he hasbeen an assistant professor of computer science at TheUniversity of Memphis. He serves as a guest editor for(Wiley) Journal of Wireless Communications and MobileComputing, special Issue on ‘Mobility, paging and quality ofservice management for future wireless networks’ in 2004,and an associate guest editor for special issue on ‘Paralleland distributed computing, applications and technologies’of International Journal of High Performance Computingand Networking, 2003. He serves as a symposium co-chairof IEEE Vehicular Technology Conference (VTC’2003):symposium on Data Base Management in Wireless NetworkEnvironments. He serves as a Technical Program Committee(TPC) member for many conferences, such as IEEE ICDCS2004, IEEE GLOBECOM 2004, IEEE WCNC 2004, IEEEICCCN 2004, IEEE PIMRC 2004, ACM SAC 2004, MWN2004 and MDC 2004. He is a voting member of the IEEE802.11 working group and a member of IEEE and ACM. Hiscurrent research interests include wireless LANs, wirelessPANs and mobile cellular networks.

Haizhon Li is currently a Ph.D.candidate at Computer ScienceDivision, the University of Mem-phis. He received his M.S. degreein Computer Science from the Uni-versity of Memphis in 2003. Hisresearch interests include Qualityof Service and MAC enhancementfor IEEE 802.11 wireless LANs,performance analysis and integra-tion of wireless LANs, wirelessPANs and 3G cellular networks.

Dr. C. L. Philip Chen received hisM.S. degree from the University ofMichigan, Ann Arbor, Michigan, in1985 and his Ph.D. from PurdueUniversity, West Lafayette, Indiana,in December 1988. He has been withthe Computer Science and Engineer-ing Department, Wright State Uni-versity, Dayton, Ohio for 13 years, asan assistant, an associate and a fullprofessor. Since September 2002, hehas joined the Department of Electri-

cal and Computer Engineering, the University of Texas at SanAntonio as a full professor. Dr. Chen has been a visitingresearch scientist, at the Materials Directorate, WrightLaboratory, Wright–Patterson Air Force Base. He has beena senior research fellow sponsored by the National ResearchCouncil. He also has been a research faculty fellow forWright Laboratory and NASA Glenn Research Center forseveral years. His research interests and projects includecomputer networking, neural networks, fuzzy-neural sys-tems, intelligent systems, robotics and CAD/CAM. Hiscurrent research area with NASA Glenn Research Centerincludes data mining on aircraft flight and maintenance data,aircraft engine life prediction and life extending control. Dr.Chen has been a conference co-chairman of the InternationalConference on Artificial Neural Networks in Engineering(ANNIE), 1995 and 1996; a tutorial chairman on IEEE Int’lConference on Neural Networks, 1994; a conference co-chairman of the Adaptive Distributed Parallel Computing,1996; a program/organizing committee of the IEEE Int’lConference on Robotics and Automation, 1996 and 2000–2002, Int’l Conf. on Intelligent Robotics and Systems (IROS),1998–2004. He actively reviews several IEEE TransactionJournals. Dr. Chen is a member of Tau Beta Pi and Eta KappaNu honorary societies and a senior member of the IEEE. Heis the founding faculty advisor of the IEEE Computer SocietyStudent Chapter at Wright State University.

Dr. Bin Wang received his Ph.D. inElectrical Engineering from theOhio State University, in 2000. Hejoined the Department of ComputerScience and Engineering, theWrightState University, Dayton, Ohio, inSeptember 2000 as an assistant pro-fessor. He spent the summer of 1998at Panasonic Information and Net-working Technology Laboratory,Princeton, NJ. His research interests

242 Y. XIAO ET AL.


are communication networks with emphasis on providingquality of service assurance in high-speed networks,DWDM optical networks, wireless and mobile networks,network security, modeling and queuing analysis of systems,simulation optimization and network protocol development.He is a recipient of US Department of Energy Early CareerAward, in 2003.

Yi Pan was born in Jiangsu, China.He entered Tsinghua University inMarch 1978 with the highest collegeentrance examination score amongall 1977 high school graduates inJiangsu Province. Dr. Pan receivedhis B. Eng. and M. Eng. degrees inComputer Engineering from Tsin-ghua University, China, in 1982 and1984 respectively, and his Ph.D.degree in Computer Science fromthe University of Pittsburgh, USA,

in 1991. Currently, he is a Yamacraw professor in theDepartment of Computer Science at Georgia State Univer-sity. His research interests include parallel and distributedcomputing, optical networks, wireless networks and bioin-formatics. His pioneer work on computing, using reconfi-gurable optical buses has inspired extensive subsequentwork by many researchers and his research results havebeen cited by more than 100 researchers worldwide inbooks, theses, journal and conference papers. He is a co-

inventor of three U.S. patents (pending) and five provisionalpatents. He has co-edited 11 books/proceedings, and pub-lished more than 130 research papers including over 60journal papers (more than 20 of which have been publishedin various IEEE journals) and received many awards fromagencies such as NSF, AFOSR, JSPS, IISF and MellonFoundation. His recent research has been supported byNSF, NIH, AFOSR, AFRL, JSPS, IISF and the states ofGeorgia and Ohio. He has served as a reviewer/panelist formany research foundations/agencies such as the U.S.National Science Foundation, the Natural Sciences andEngineering Research Council of Canada, the AustralianResearch Council and the Hong Kong Research GrantsCouncil. Dr. Pan has served as an editor-in-chief or editorialboard member for eight journals including three IEEETransactions and a guest editor for seven special issues.He has organized several international conferences andworkshops and has also served as a program committeemember for several major international conferences such asINFOCOM, GLOBECOM, ICC, IPDPS and ICPP. Dr. Panhas delivered over 40 invited talks, including keynotespeeches and colloquium talks, at conferences and univer-sities worldwide. Dr. Pan is an IEEE distinguished speaker(2000–2002), a Yamacraw distinguished speaker (2002), aShell Oil colloquium speaker (2002), and a senior memberof IEEE. He is listed in Men of Achievement, Who’s Who inMidwest, Who’s Who in America, Who’s Who in AmericanEducation, and Who’s Who in computational science andengineering.



Proportional degradation services in wireless/mobile adaptive multimedia networks

Documents

Transcript of Proportional degradation services in wireless/mobile adaptive multimedia networks