Dynamic bandwidth allocation for multimedia traffic with rate guarantee and fair access in WCDMA systems

Dynamic Bandwidth Allocation for

Multimedia Traffic with Rate Guarantee

and Fair Access in WCDMA Systems

Chih-Min Chao, Yu-Chee Tseng, Senior Member, IEEE Computer Society, and

Li-Chun Wang, Member, IEEE

Abstract—Packet scheduling in a WCDMA system poses a new challenge due to its nature of variable bit rates and location-dependent, time-varying channel conditions. In this work, three new downlink scheduling algorithms for a WCDMA base station are proposed to support multimedia transmissions. Using a credit management and a compensation mechanism, our algorithms provide rate guarantee and fair access to mobile terminals. In particular, we propose to allow a user to simultaneously use multiple OVSF codes in a time-sharing manner, which we call a multicode, shared model. Using multiple codes allows us to compensate those users suffering from bad communication quality or even errors. The proposed schemes can tolerate a multistate link condition (compared to the typically assumed two-state, or good-or-bad, link condition) by adjusting the number of OVSF codes and the spreading factor of each code. Simulation results show that the proposed schemes do achieve higher bandwidth utilization while keeping transmission delay low.

Index Terms—Mobile computing, OVSF, personal communication services, WCDMA, wireless communication, 3G.

æ

1

I

S

service (QoS) requirements is one of the ultimate goals in next-generation wireless systems. The current second-generation (2G) systems, such as IS-95 (CDMA) and GSM (TDMA), are circuit-switched, fixed-rate, and voice-traffic-oriented and, thus, not appropriate to support multimedia services. The third-generation (3G) wireless standards [1], [2], [12] will be based on the WCDMA technologies and can flexibly support mixed and variable-rate services. Two transmission schemes are proposed in the 3G wireless standard: multicode-CDMA (MC-CDMA) and Orthogonal Variable Spreading Factor CDMA (OVSF-CDMA). In MC-CDMA, multiple Orthogonal Constant Spreading Factor (OCSF) codes can be assigned to a user [10], [13]. The maximum data rate a user can receive depends on the number of transceivers in the device. In OVSF-CDMA, a single OVSF code can provide a data rate that is several times than that of an OCSF code, depending on its spreading factor [2].

In CDMA systems, multiple connections are allowed to receive packets simultaneously. This is opposite to TDMA systems where only one connection can be active at any moment. Thus, the scheduling problem for WCDMA systems is harder than that for TDMA systems. It can

neither be solved directly by wireline scheduling disci-plines, such as WFQ [19], virtual clock [27], and EDD [15], since they do not consider the variability of wireless connections, nor be solved by wireless scheduling strate-gies, such as IWFQ [16], CIF-Q [18], and SBFA [21], since they only consider a single-server, two-state link model.

This paper considers the bandwidth allocation problem in an OVSF WCDMA system. In the literature, solutions to this problem can be classified depending on two factors:

. single-code/multicode: Whether a user can

simulta-neously utilize multiple OVSF codes or not.

. dedicated/shared: Whether a code can be time-shared

by multiple users or not.

In Table 1, we categorize existing solutions and the solutions proposed in this paper based on such classification.

The OVSF code assignment strategy, though playing an important role in system performance, is not explicitly addressed in the current 3G WCDMA standard. Most works in the literature [3], [6], [9], [17], [23], [26] fall into the dedicated, single-code class. One OVSF code is exclusively assigned to one client until the call is terminated or reallocated. Intelligently assigning codes to calls can reduce code blocking [23], [26]. Solutions [5], [7], [22], [25] allow a user to occupy multiple but dedicated OVSF codes. Such solutions are more flexible and, thus, can reduce code blocking.

Since wireless bandwidths are limited resources, it is usually desirable that a code can be time-shared by multiple users. The advantage is higher flexibility in utilizing bandwidth, especially in handling bursty traffic. Allowing sharing means that we need to decide “which code can be used by which user at what time.” So, a solution should comprise a code assignment strategy and a scheduling

. C.-M. Chao is with the Department of Computer Science, National Taiwan Ocean University, 20224, Taiwan. E-mail: cmchao@axp1.csie.ncu.edu.tw. . Y.-C. Tseng is with the Department of Computer Science and Information

Engineering, National Chiao-Tung University, 30050, Taiwan. E-mail: yctseng@csie.nctu.edu.tw.

. L.-C. Wang is with the Department of Communication Engineering, National Chiao-Tung University, 30050, Taiwan.

E-mail: lichun@mail.nctu..edu.tw.

Manuscript received 20 Aug. 2003; revised 26 Jan. 2004; accepted 9 Mar. 2004; published online 27 July 2005.

For information on obtaining reprints of this article, please send e-mail to: tmc@computer.org, and reference IEEECS Log Number TMC-0131-0803.

scheme. Recently, two solutions in the single-code, shared category are proposed [14], [24]. In [24], a packet scheduling scheme is provided but the code allocation issue is not addressed. A credit-based scheduling scheme is proposed in [14]. However, this scheme does not consider channel quality, which may impact the selected spreading factor. A more comprehensive review is in Section 1.1.

In this paper, we propose three new credit-based bandwidth allocation schemes that allow a user to utilize multiple time-shared codes. This is more flexible than what was proposed in existing works. With our credit manage-ment mechanism and the compensation mechanism pro-vided by additional codes, the schemes can provide fair access and data rate guarantee. Using multiple OVSF codes has two purposes. First, it is for compensation purpose when a terminal encounters errors. Second, it can adapt to an environment with multiple link states, in the sense that a higher spreading factor can be used when the link quality is bad; however, the user can still enjoy the guaranteed bandwidth supported by multiple codes. Simulation results show that the proposed schemes can achieve higher bandwidth utilization and keep average delay low.

The rest of this paper is organized as follows: In Section 2, we introduce the system model. Section 3 presents the proposed dynamic bandwidth allocation algorithms. Per-formance evaluations are in Section 4. Concluding remarks are given in Section 5.

1.1 Related Works

In the dedicated, single-code class, [23] shows that significant improvement can be obtained by using a leftmost-first or a crowded-first strategy, compared to a random assignment. When a code tree is used for long time, it is possible that the tree may become too fragmented, due to calls arriving and leaving the system. Code replacement can be conducted to improve bandwidth utilization. The dynamic code assignment (DCA) scheme [17] is designed for this purpose. In this case, code assignment is semidedicated, in the sense that a call can be moved to a different code when relocation is conducted.

Allowing a user to occupy multiple dedicated codes provides more adaptability. The works in [7], [22] address the relationship between a requested data rate and the number of codes to be assigned to the request. However, the placement of codes in the OVSF code tree is left unspecified. A multicode scheduling scheme to be run on the base station is proposed in [25], where fair scheduling for users to share the resources is addressed. Again, the important code placement issue is left unaddressed. In [5], both the code placement and replacement issues are addressed for such an environment.

The scheme proposed in [14] is a member in the single-code, shared class. Initially, each user is assigned a leaf code (of the lowest rate). However, the scheduling algorithm would allow a user to utilize any ancestor code of his/her assigned code. In each time frame, the algorithm will repeatedly pick the user with the highest credit, say A, and try to move A’s code one level up from its current stand if A still has backlog packets to be sent. This increase also means sacrificing other users’ assigned, but unused, codes. That is, the latter will be inhibited from transmission in this round. The process is repeated until all capacities in the code tree are consumed or no packet is pending. Inhibited users’ credits will be increased in the next frame. This scheme fails to consider the possible higher bit error rate (BER) that is caused by using a lower spreading factor. Depending on the channel condition, the spreading factor should not be unlimitedly reduced. Otherwise, a lot of retransmissions may be incurred, resulting in even worse performance. The scheme also includes an exception handling mechanism to solve the starvation problem that may occur during scheduling.

2

S

M

In the WCDMA system [20], data transmission involves two steps. The first step is channelization, which transforms every data bit into a code sequence. The length of the code sequence per data bit is called the spreading factor (SF), which is typically a power of two. The second step is scrambling, which applies a scrambling code to the spread signal. Scrambling codes are used to separate signals from different sources, while channelization codes multiplex data to different users from the same source.

The OVSF codes are used as the channelization codes in the WCDMA system. The possible OVSF codes can be represented by a code tree as shown in Fig. 1. Each OVSF

code is denoted as CSF ;k, where SF is the spreading factor

and k is the branch number, 1
k
SF . The number of codes at each level is equal to the value of SF . All codes in the same layer are orthogonal, while codes in different layers are orthogonal only if they do not have ancestor-descendant relationship. Leaf codes have the minimum data rate, which

is denoted by 1Rb. The data rate is doubled whenever we go

one level up the tree. For example, in Fig. 1, C4;1has rate 2Rb,

and C2;1has rate 4Rb. The resource units in a code tree are

codes. Fig. 1 shows an example where codes C4;2, C8;1, and

C8;5are occupied by users. The remaining capacity is 4Rb. If

the dedicated, single-code model is assumed, a new call

requesting a rate of 4Rb will be rejected because there is no

TABLE 1

Classification of Bandwidth Allocation Solutions for WCDMA Systems

such code available. Such a situation is called code blocking. However, the problem can be solved if up to three codes can be used by a user under the multicode model.

A main characteristic of wireless communications is link variability, which could be time-dependent and location-dependent. It is time-dependent because interference can come to hit at any time. One example frequently seen is bursty errors. The capacity of a wireless link is location-dependent because a longer distance typically has to suffer lower transmission rate. These properties necessitate the design of dynamic bandwidth allocation mechanisms.

Since the quality of a wireless link is both time and location-dependent, we assume that its capacity follows a multistate model. Specifically, a wireless link’s symbol error rate is inversely proportional to its received signal

strength, which is given by Eb

N0¼ SNR  SF , where Eb is

the energy per bit, N0 is two times the power spectral

density of additive white Gaussian noise, SNR is the signal-to-noise ratio, and SF is the spreading factor. To

achieve a target Eb=N0, one can either increase SNR

through power control or increase the spreading factor

SF. When the transmission power is up to a limit,

increasing SF is the only way to reduce bit error rate. In this work, we assume that there exists a target BER for all users. As a result, at any moment, there is always a maximum data rate (and, thus, minimum spreading factor) for each user. In such an environment, the scheduling problem is still an open question [4]. Most existing scheduling strategies [8], [11], [16], [18], [21] consider a simple two-state channel model where a channel can be either good or bad, receiving full or none capacity, respectively.

This paper considers the downlink bandwidth allocation problem for a base station. The resource at the base station is an OVSF code tree. Each user i, when entering the system,

needs to specify its peak data rate piRband guaranteed data

rate aiRb. Both piand aiare powers of two and pi ai. The

base station maintains a queue Qifor user i’s packets to be

delivered. So, the inputs to our scheduling algorithms are pi, ai, and Qi, for all users i ¼ 1::n. The base station is

responsible for utilizing its code tree to serve users efficiently and fairly. According to the current 3G standard, each frame is 10 ms, which consists of 15 slots. So, in every 10 ms, we can reschedule codes for users. The purpose of bandwidth allocation is to achieve high utilization while providing guarantee rates for individual users.

We assume the shared, multicode model, where a user

can simultaneously use up to Nmax codes. This model has

several advantages: 1) when a user suffers severe inter-ference, we can increase the spreading factor to improve reliability but maintain the promised data rate by using multiple codes, and 2) when a user suffers temporary degradation, compensation can be made up to it by using multiple codes for fairness.

3

D

B

A

S

This section presents our dynamic bandwidth allocation strategies to provide QoS-guaranteed services. Our solu-tions consist of three parts: code assignment, credit management, and packet scheduling.

3.1 Code Assignment Part

This part decides which codes should be assigned to each user to satisfy his/her demand. We assume that there is an admission control mechanism such that a user is accepted only if the sum of guaranteed data rates over all users does not exceed the capacity of the code tree. For each code k in

the code tree, the base station maintains a variable ðSCkÞ

(read as shared count) to keep track of the number of users who are currently sharing k, either partially or completely. Specifically, a user is said to share code k if the user is using code k, a descendant code of k, or an ancestor code of k. For example, in Fig. 1, the number in each node is the corresponding code’s shared count. It is easy to maintain the shared counts. Whenever a code k is allocated to a user, the shared counts of k and k’s ancestor and descendant codes are all increased by one. These counts are decreased by one when this user leaves.

Note that, in contrast to the code assignment in [14], which assigns a leaf code to each user initially, we allocate codes to users according to their maximal data rates. During transmission, a user can use any of the descendent codes of the initially assigned one, which is compatible to the 3gpp standard [1].

3.1.1 Scheme 1: Multiple Fixed Codes

In this scheme, we allocate Nmax fixed codes, each of rate

pi Rb, to user i. Under normal situations, only one code is

sufficient. The additional codes can be used for interference reduction or bandwidth compensation. This will be further elaborated on in Section 3.3.

Given user i’s request, the scheme sequentially picks

Nmax codes. To allocate each code, the following rules are

applied:

1. Scan all codes in the code tree with rate piRb. Pick

the one(s) with the least shared count. If there is only one candidate, assign this code to user i. Otherwise, go to step 2.

2. If the shared counts of these candidate codes are 0s,

we follow the crowded-first rule to do the allocation. Specifically, if x and y are two codes such that

SCx¼ SCy¼ 0, we compare x’s and y’s parent

codes. The one with a larger SC is selected. If they tie again, we further compare their parents. This is repeated until the tie breaks. One special case is that x’s and y’s parents could be the same code. If so, we simply select the one on the left-hand side.

3. If the shared counts of the candidate codes are

nonzero, we follow the fairness rule. Specifically, if x

and y are two codes such that SCx¼ SCy6¼ 0, we

compare x’s and y’s parent codes similar to the recursive procedure in step 2. However, we select the code with a smaller shared count, instead of a larger one.

Intuitively, when there are still free codes, we try to place users’ requests in the code tree as compact as possible (this is what suggested in [23]). Otherwise, some degree of sharing among users is necessary, and we try to assign codes as fairly as possible.

Take the scenario in Fig. 2a as an example. Suppose

Nmax¼ 2 and a new user requests for a 2Rb rate. After

searching for all 2Rbcodes, we find that C8;2, C8;5, C8;6, and

C8;8 all have zero shared counts. So, we further compare

their ancestors, C4;1, C4;3, and C4;4. Among them, C4;1 and

C4;4, which have the same shared counts of two, tie again.

After going one level up, we see that C2;1has a larger shared

count than C2;2. So, C8;2is allocated to the user. The same

procedure is applied to allocate the second code. This time

C8;8will be picked. Another example where the code tree is

fully occupied is shown in Fig. 2b. Again, assuming the same situation, step 1 will pick C8;3, C8;4, C8;7, and C8;8, all of

which have the same shared count ¼ 1. After going one

level up, we see that C4;2has a smaller shared count. Since

C8;3 and C8;4 tie, following the leftmost-first rule, we will

select C8;3. To assign the second code, we see that C8;4, C8;7,

and C8;8 all have the same least SC value. Their ancestors,

C4;2 and C4;4, tie again with the same SC value of two.

Another tie is found when coming to codes C2;1 and C2;2.

Since C2;1 and C2;2 have the same ancestor, the leftmost

candidate, C8;4, will be assigned to the user.

3.1.2 Scheme 2: Single Fixed Code with Multiple Dynamic Codes

This scheme only assigns one fixed code to each user. The

other Nmax 1 codes are all assigned in a dynamic manner.

Specifically, the OVSF code tree is partitioned into two areas: partially shared area on the left and fully shared area on the right (how to select a good partition point will be evaluated in Section 4). Given a user requesting a rate piRb,

it is assigned a code of rate piRbin the partially shared area.

The assignment follows the same rule in the previous scheme except that only the partially shared area is

assignable. The remaining Nmax 1 codes will be assigned

to the fully shared area, but the assignment will not be done in this stage and will be decided at the scheduling stage. 3.1.3 Scheme 3: No Fixed Codes

In contrast to the previous schemes, this scheme does not

assign any fixed code to users. All Nmax codes will be

allocated dynamically at the scheduling stage. This scheme provides the maximal flexibility (alternatively, one can consider the whole OVSF code tree as a fully shared area in Scheme 2).

3.1.4 Remark on Code Notification

The codes used in our schemes can be directly mapped to the transport channels in 3gpp. To schedule users, the

Forward Access Channel (FACH) can be used to notify them that their transmission data rates as well as codes. An additional ID field is needed in the header to distinguish users (we will analyze the signaling overhead in Section 3.4). In schemes 1 and 3, all codes allocated to a user can be mapped to the Dedicated Transport Channel (DCH). In scheme 2, the single fixed code can be mapped to DCH, while the remaining dynamic codes to Downlink Share Channel (DSCH). According to 3gpp, a DSCH is always associated with a DCH, which provides necessary signaling for the DSCH.

3.2 Credit Management Part

In the above part, we already assigned each user multiple codes which exceed the user’s guaranteed data rate. These codes are not necessarily always used by the user. We employ a credit management mechanism to dynamically allocate bandwidths to users. The base station maintains a credit Cifor each user i. Initially, Ci¼ 0. After every 10 ms,

aicredits, i.e., user i’s guaranteed rate, are granted to user i

and added to Ci. However, for every code of rate 2kRb that

is consumed by user i, 2k _{credits are paid by decreasing 2}k

from Ci. Hence, as a user’s perceived data rate is below its

guaranteed rate, its credits can be accumulated for later use.

The value of Ci can also be negative, if the user is

over-served (which happens, for example, if other users under-use their capacities due to poor channel condition). Long-term fairness and rate guarantee are provided if the base station honors all users’ credits.

Care should be taken if a user’s perceived data rate is below his/her request rate for long time, due to reasons such as poor channel condition, sudden congestion, or simply low data arrival rate. In this case, a lot of credits may be accumulated for the user. Later on, if a large amount of traffic arrives for the user, the user may take up a lot of bandwidth, thus blocking the opportunity of other users. This is alright if only long-term rate guarantee is required, but it does cause a problem viewed for a short term. Short-term fairness is not guaranteed if we allow a user to accumulate credits unlimitedly. To resolve this problem, we

propose to maintain a credit limit Li for each user i such

that Ci
Li is always true. The value of Li can be

negotiated with the base station in a per-user basis when the connection was first established according to the user’s priorities, degree of burstiness, or even the system loading. Note that we do not distinguish the situation that a user has lower traffic than expected from that it suffers from bad channel condition. In both situations, it can accumulate credits. An alternative is to apply the credit limit only in the

Fig. 2. A code assignment example when the code tree has (a) free 2Rbcodes and (b) no free 2Rbcode. The number in each node is its shared

former case [14]. However, we believe that it is more reasonable to apply the limit to both cases because wireless channels may suffer from burst errors more frequently.

3.3 Packet Scheduling Part

The scheduling algorithm is executed once every 10 ms (frame length). It examines all users with backlog data at the base station and determines which users are to be served by which codes. The goal is to provide a fair and rate-guaranteed service to each user.

3.3.1 Scheduling for Scheme 1 (Multiple Fixed Codes) To facilitate the scheduling scheme, we introduce the term normalized credit of user i, which is defined to be the ratio

Ci=ai. The normalized credit represents the amount of

services a user i should receive relative to his/her

guaranteed rate ai. A user with the greatest normalized

credit, instead of credit, should be scheduled first. The scheduling algorithm for scheme 1 has six steps:

. STEP 1. Sort all users according to their normalized

credits.

. STEP 2. Pick the user, say i, with the greatest

normalized credit such that user i has backlog packets and at least one of user i’s codes is free. If there is no such user, the procedure terminates.

. STEP 3. Let Mi be the maximum bit rate for user i

such that the BER is satisfied (this can be obtained by monitoring the channel condition).

. STEP 4. Let Ti ¼ minfMi; pig, which represents the

feasible transmission rate considering user i’s channel quality. We search all free codes of user i. If there exists at least one free code, we allocate the leftmost one for user i. Otherwise, we go one level down by searching all descendant codes (of rate

Ti

2Rb) of user i’s codes. If there exist at least one free

code, we allocate the leftmost one for user i. Otherwise, we repeat the same procedure and search

the descendant codes of ratesTi

4; Ri

8;   , etc.

. STEP 5. For the code allocated in STEP 4, we

decrease user i’s credit Ci by the corresponding

amount. Then, user i is put in to the sorted list.

. STEP 6. Go back to STEP 2.

Take Fig. 3 as an example. Suppose that user i has the greatest normalized credits and C8;1, C8;2, and C8;3are user

i’s codes. When user i is selected for the first time, only C8;2

is free. So, C8;2is allocated to serve user i in the next frame

and two credits are subtracted from Ci. If user i is selected

for the second time, C16;2will be allocated to it.

3.3.2 Scheduling for Scheme 2 (Single Fixed Codes with Multiple Dynamic Codes)

The scheduling algorithm is similar to the previous case. The difference lies on how a user’s codes are used. All steps are the same as the previous case, except Step 4, which is described below:

. STEP 40_{. If this is the first time user i is selected in}

this frame, go to (a). Otherwise, go to (b).

- (a) Let Ti¼ minfMi; pig. We search the single

fixed code of user i. We allocate it to user i if it is free. Otherwise, we go one level down by

searching all descendant codes (of rateTi

2Rb) of

user i’s code. If there exists at least one free code, we allocate the leftmost one to user i. Otherwise, we repeat the same procedure and

search the descendant codes of rates Ti

4; Ti

8;   ,

etc. If we cannot find a free code, user i is skipped for further scheduling in the current time frame.

- (b) Allocate a free code at the fully shared area

(this is similar to case (a) but we have more freedom because any free code can be allocated). Again, take Fig. 3 as an example. Let the quarter of the code tree on the right-hand side be the fully shared area. Suppose that user i has the highest normalized credit and

C8;1 is his/her initial code. The scheduling algorithm will

allocate C16;2to user i in the first round. If user i is selected

again for next two rounds, C16;15 and C16;16 in the fully

shared area will be allocated to it. On the contrary, if C8;3is

user i’s initial code, it cannot be scheduled in the current

frame because C8;3is already occupied by another user.

3.3.3 Scheduling for Scheme 3 (No Fixed Codes) The algorithm is similar to the previous two schemes. It follows scheme 1, except the following modifications:

. STEP 200. The same as STEP 2 except that we only

need to make sure that user i has not exhausted its

Nmax codes.

. STEP 400_{. The same as STEP 4 except that any free}

space in the code tree can be used by user i.

3.4 Time Complexity and Signaling Overhead

Analysis

Next, we analyze the time complexity of our schemes. Consider the code assignment part. There are SF codes in the code tree with a spreading factor of SF . So, the cost is

OðSF Þ to search for such a code. For the fixed code

assignment part of scheme 1, we may need to compare the shared counts of candidate codes’ ancestors. Each time when we go one level up the code tree, the number of ancestors reduces by half. It is easy to see that the cost to allocate a code with a specific SF is

SFþSF

2 þ    þ 1 ¼ OðSF Þ:

Thus, to assign Nmaxcodes, the searching cost is OðNmax

SFÞ ¼ OðSF Þ because Nmax is a constant. The time

complexity for the fixed code assignment part of scheme 2 is similar to that of scheme 1, except that the searching domain is restricted to the partially shared area. So, the time complexity is the same.

For the credit assignment part, it is easy to see that the time complexity is OðnÞ for all three schemes, where N is the number of users currently in the system. For the packet scheduling part, all three schemes need to sort all users, giving a cost of Oðn log nÞ. The STEP 4 of scheme 1 may spend 1 þ 2 þ    þSFmax

SF ¼ Oð SFmax

SF Þ time to find a free code,

where SFmax is the maximum allowable SF . So, the total

searching cost is Oðn0_SFmax

SF Þ, where n0is the actual number

of transmissions that are scheduled in this time frame. STEP 5 takes OðnÞ to put the scheduled user back to the sorted list. For scheme 2, the searching cost of STEP 40_{(a) is}

OðnpSF_SFmaxÞ, where np is the number of scheduled

transmissions in the partially shared area. In STEP 40_(b),

each code assignment takes OðSF Þ time, so the total cost is

Oðnf SF Þ, where nfis the number of scheduled

transmis-sions in the fully shared area. Similarly, for scheme 3, the searching cost of STEP 400 is Oðn0 SF Þ.

Overall, the time complexity for scheme 1 is

O SFþ n log n þ n0SFmax

SF



: For scheme 2, the time complexity is

O SF þ n log n þ np

SFmax

SF þ nf SF



: For scheme 3, the time complexity is

OðSF þ n log n þ n0_{ SF Þ:}

The complexities are typically low. For example, in our

simulations, with SFmax¼ 256, the value of n is around 68,

which should take less than tens of microseconds to complete these calculations in modern processors.

The signaling overhead for scheme 1 is n0_{ lg S}

1

d e,

where S1 is the maximum number of users that use the

same code. For scheme 2, the signaling overhead is np lg Sd 2e þ nf lg nd 2e, where S2 is the maximum

num-ber of users that use the same code in the partially shared

area and n2is the number of codes in the fully shared area.

The signaling overhead for scheme 3 is n0_{ lg n}

3

d e, where

n3is the number of codes in the code tree.

4

P

E

We have implemented a simulator to evaluate the perfor-mance of the proposed strategies. The max SF is set to 256. We control the call generation such that the overall guaranteed traffic load falls between 10 percent and 90 percent. Three traffic models are tested [14]:

. Model I (constantly backlogged model): Each user

has queued packets all the time. Each user’s peak transmission rate is uniformly distributed between

4Rband 32Rb, and guaranteed rate one fourth of the

peak rate (i.e., ai ¼14pi).

. Model II (highly bursty traffic model): Calls are

generated into the system with a peak rate uniformly

distributed between 4Rband 32Rband a guaranteed

rate equal to one fourth of the peak rate. Packet arrival of each connection follows a 2-state Markov model. A connection can transit between an idle state or an active state. No packet is generated during the idle state, while P packets per 10 ms are generated during the active state, where P is uniformly

distributed between 2ai and 4ai. A state transition

can be made every 10 ms, with probabilities of 1/3 and 2/3 from idle to active and from active to idle, respectively.

. Model III (lowly bursty traffic model): Calls are

generated into the system with a peak rate uniformly

distributed between 2Rband 16Rband a guaranteed

rate equal to one half of the peak rate. Packet arrival also follows the above 2-state Markov model, but P

is uniformly distributed between ai and 2ai, while

state transition probabilities are 2/3 and 1/3 from idle to active and from active to idle, respectively. Each transport block is set to 150 bits, which means a

1Rb code can transmit 150 bits every 10 ms. A transport

block will be retransmitted if an error is detected by the cyclic redundancy check (we assume no channel coding is

applied). In schemes 1, 2, and 3, Nmaxranges from 2 to 4. In

scheme 2, the first three quarters of the code tree is the partially shared area, and the last quarter is the fully shared area.

Using the QPSK modulation, we model the symbol error probability by PE  2Q ffiffiffiffiffiffiffiffi 2Eb N0 s ! ; ð1Þ where QðxÞ ¼ Z 1 x ey2=2 ffiffiffiffiffiffi 2
p dy and Eb N0 ¼ SNR  SF :

Intuitively, the transmission rate is inversely proportional to the SF being applied. A lower SF has a higher transmission rate but the potential penalty, according to (1), is a higher

PE. To achieve a better throughput, we should choose the

smallest SF that meets the required PE. A scheme without

considering channel condition, such as [14] (denoted by

KMS below), is impractical because it may choose an

improper SF and, thus, suffer from severe performance

degradation. In this paper, the upper bound of PE is set to

105. We assume a Rayleigh fading channel, so signal

power is modeled by an exponential random variable X

with mean , i.e., fðxÞ ¼1

e

x

 _{for x  0 [20].}

We first experiment on the KMS scheme on traffic model II by varying  and evaluating the unsatisfactory transmission, which is defined to be the percentage of frames

threshold 105_{. As shown in Fig. 4, KMS performs highly}

depending on the channel condition. At low , a lot of packets may experience failures. So, the scheme is not channel-sensitive. To keep the failure rate low, say under 5 percent,  should be maintained 25 or higher. Note that traffic load would increase the error rate because the KMS scheme favors users with more backlogged packets. This would aggravate error transmission.

Below, we further experiment on four aspects. Each result below is from an average of 50 simulation runs, where each run contains at least 3,000 frames:

1. Impact of Code Assignment Strategies: This experiment

tests different code assignment strategies. We eval-uate three metrics: code utilization, effective code utilization, and average delay. Code utilization is the average capacity of the code tree that is actually used for transmission. However, considering those trans-missions experiencing failure, effective code utiliza-tion counts only those successfully transmitted bits. In addition to the KMS scheme, we also simulate the code, shared” scheme and the “single-code, dedicated” scheme.

Fig. 5 shows the code utilization and effective code utilization at various traffic loads. For our

schemes, Nmax is set to 2. For code utilization, our

scheme 3 performs the best, which is sequentially followed by the KMS scheme, our scheme 2, our scheme 1, and then the single-code schemes. After

taking failure transmissions into account, effective code utilization of KMS scheme reduces signifi-cantly when  ¼ 1. For example, when the code tree is 90 percent fully loaded, KMS’s effective utiliza-tion reduces to around 51 percent, while our schemes can still maintain an effective utilization over 85 percent. Only when the channel condition is extremely good (such as  ¼ 25), can KMS maintain high effective code utilization. Since our schemes have considered channel condition, the code utiliza-tion and effective code utilizautiliza-tion are very close.

Fig. 6 shows the average delay, which is defined as the average time, measured in time frames, a transport block is queued at the base station. With

¼ 1, our scheme 3 experiences the least delay,

which is subsequently followed by our scheme 2, scheme 1, single-code, shared scheme, and KMS. Our scheme 3 produces low delay since there is no code blocking. The delays for scheme 1 and single-code, shared scheme increase significantly when the code tree is above 80 percent loaded, which means the system is unstable then. The KMS scheme makes the system unstable after traffic load is over 30 percent. These results reveal that our scheme 3 and scheme 2 can accept more time-bounded services. The delays of KMS reduces significantly when  ¼ 25, which is reasonable since the channel condition is pretty good.

2. Fairness Properties: Next, we verify how our schemes

support rate guarantee and fairness. We set Nmax¼ 4

since it achieves better performance than Nmax¼ 2

and 3 (performance comparisons of different Nmaxare

not reported here due to space limit). To see rate guarantee, we randomly choose a user i with request

ai¼ 1; 2; 4 or 8Rb and observe the number of

transmitted transport blocks. As shown in Fig. 7, both schemes 2 and 3 support rate guarantee well. For scheme 1, there is some lag for ai¼ 4 and 8Rb, which

stems from the higher failure probabilities when allocating codes for such calls.

To verify access fairness, we introduce a metric called interservice time, which is the interval that a

Fig. 4. Unsatisfactory transmission versus traffic load at different  for the KMS scheme [14].

backlogged user experiences, measured in time frames, between two successive transmissions. We observe the maximum difference of interservice times of two backlogged users who are waiting for scheduling. The result is in Table 2. In all three traffic models, scheme 3 has the least difference, which is followed by scheme 2 and then scheme 1. For the constantly backlogged model, the differences are low for all schemes and are proportional to traffic load. For example, at traffic load = 80 percent, the most unfortunate user will experience at most 15.1, 8.1, and 3.9 time frames of delay for schemes 1, 2, and 3, respectively. For both bursty traffic models, the difference is relatively independent to traffic load. Scheme 3 performs the best, and scheme 1 the worst. In general, the differences are larger for the highly bursty traffic model, which is reasonable because traffics with higher variation are more difficult to handle.

We also investigate the impact of different traffic models on average delay. The results are in Fig. 8. Scheme 3 always has the least delay and is quite independent of traffic models. Scheme 2 also has very low delay (e.g., the average delays are only 8 and 5 time frames at load = 90 percent for traffic models II and III, respectively). Scheme 1 has a longer delay for lowly bursty traffics because bursty

traffics arrive more frequently (thus causing more backlogged users).

3. Signaling Overhead: Since different values of Nmax

have similar results, here we use Nmax¼ 4 as a

representative case. We plot the signaling overhead of different schemes according to the formulation in Section 3.4. As shown in Fig. 9, in all traffic models, scheme 2 incurs the least signaling overhead while scheme 3 the highest. Note that we should also take into account the utilization issue while evaluating these schemes. When this factor is considered, scheme 2 will be the best for traffic models II and III. For example, for model II under load = 90 percent, the signaling overheads are 152 and 720 bits for schemes 2 and 3, respectively. The effective utilization of scheme 3 is only 1 percent higher than scheme 2, which means around 1% 

256 150 ¼ 384 bits of reward at the expense of

720 152 ¼ 568 bits of overhead when comparing

scheme 3 and scheme 2. However, for model I, scheme 3 performs the best when the utilization factor is considered.

4. Impact of Fully Shared Code Space for Scheme 2: Last,

we investigate how the ratio of fully shared code space in scheme 2 affects its performance. Four ratios,1 8, 2 8, 3 8, and 4

8, are tested with different Nmax

values. Due to the space limit, we report the results

without figures. When Nmax¼ 2, the one with2₈ fully

shared code space performs the best. As Nmax

increases, larger ratios would be more beneficial. It is because calls can more freely utilize the fully shared space, causing less blocking problem. When

Fig. 6. Average delay versus traffic load under traffic model II with (a) ¼ 1 and (b)  ¼ 25.

Fig. 7. Transmitted transport blocks versus time under traffic model II.

TABLE 2

The Maximum Difference of Interservice Times under Different Traffic Models

Nmaxis large enough to better utilize the fully shared

space, the utilization would be the best. (However,

the advantage cannot be seen when Nmax is too

small.) When Nmax¼ 3 and 4, the one with38 fully

shared code space performs the best.

To conclude, we believe that a scheme that does not consider channel condition, such as the KMS scheme, may perform poorly. In our proposal, scheme 3 has the best behavior in terms of utilization, delay, rate guarantee, and fair access. However, it also imposes the highest signaling overhead. Scheme 2 falls behind scheme 3, but has the least signaling overhead. Taking all these issues into account, we recommend using scheme 3 under constantly backlogged traffic model (Model I) and scheme 2 under bursty traffic models (Model II and III) by properly tuning the fully

shared code space depending on the given Nmax.

5

C

Wireless bandwidth is a precious resource. Thus, resource management is an important issue for WCDMA. Existing mechanisms supporting multimedia traffic either do not consider channel condition or fail to address the exact code position in the code tree, which may result in inefficiency in resource utilization. In this paper, we have proposed three algorithms to provide fair and rate guaranteed service for multimedia traffic in a WCDMA system. The essence of our protocols is a credit-based scheduler which considers channel condition and explores the concept of compensation codes. With our channel-sensitive scheduling algorithms, a user with more credits will have more chance to transmit without compromising to the transmission quality. Simula-tion results do justify that schemes 2 and 3 work well.

A

Y.C. Tseng’s research is cosponsored by the MOE Program for Promoting Academic Excellence of Universities under grant number 89-E-FA04-1-4, by the NSC of Taiwan under grant numbers NSC92-2213-E009-076 and NSC92-2219-E009-013, by the Institute for Information Industry and MOEA, R.O.C, under the Handheld Device Embedded System Software Technology Development Project, by the Lee and MTI Center of NCTU, and by Computer and Communications Research Labs., ITRI, Taiwan.

R

[1] Third Generation Partnership Project; Technical Specification Group Radio Access Network, Spreading and Modulation (FDD), http://www.3gpp.org, 1999.

[2] F. Adachi, M. Sawahashi, and H. Suda, “Wideband DS-CDMA for Next-Generation Mobile Communications Systems,” IEEE Comm. Magazine, vol. 36, pp. 56-69, Sept. 1998.

[3] R. Assarut, K. Kawanishi, U. Yamamoto, Y. Onozato, and M. Masahiko, “Region Division Assignment of Orthogonal Variable-Spreading-Factor Codes in W-CDMA,” Proc. IEEE Vehicular Technology Conf., pp. 1184-1898, 2001.

[4] Y. Cao and V.O. Li, “Scheduling Algorithms in Broad-Band Wireless Networks,” Proc. IEEE INFOCOM, vol. 89, no. 1, pp. 76-87, Jan. 2001.

[5] C.-M. Chao, Y.-C. Tseng, and L.-C. Wang, “Reducing Internal and External Fragmentations of OVSF Codes in WCDMA Systems with Multiple Codes,” Proc. IEEE Wireless Comm. and Networking Conf. (WCNC), pp. 693-698, 2003.

[6] W.-T. Chen, Y.-P. Wu, and H.-C. Hsiao, “A Novel Code Assign-ment Scheme for W-CDMA Systems,” Proc. IEEE Vehicular Technology Conf., pp. 1182-1186, 2001.

[7] R.-G. Cheng and P. Lin, “OVSF Code Channel Assignment for IMT-2000,” Proc. IEEE Vehicular Technology Conf., pp. 2188-2192, 2000.

[8] D. Eckhardt and P. Steenkiste, “Effort-Limited Fair (ELF) Scheduling for Wireless Networks,” Proc. IEEE INFOCOM, pp. 1097-1106, 2000.

[9] R. Fantacci and S. Nannicini, “Multiple Access Protocol for Integration of Variable Bit Rate Multimedia Traffic in UMTS/IMT-2000 Based on Wideband CDMA,” IEEE J. Selected Areas in Comm., vol. 18, no. 8, pp. 1441-1454, Aug. 2000.

[10] V.K. Garg, IS-95 CDMA and cdma2000. Prentice Hall, 2000. [11] J. Gomez, A.T. Campbell, and H. Morikawa, “The Havana

Framework for Supporting Application and Channel Dependent QoS in Wireless Networks,” Proc. Int’l Conf. Network Protocols, pp. 235-244, 1999.

[12] H. Holma and A. Toskala, WCDMA for UMTS. John Wiley & Sons, 2000.

[13] C.-L. I et al., “IS-95 Enhancements for Multimedia Services,” Bell Labs. Tech. J., pp. 60-87, Autumn 1996.

[14] A.Z. Kam, T. Minn, and K.-Y. Siu, “Supporting Rate Guarantee and Fair Access for Bursty Data in W-CDMA,” IEEE J. Selected Areas in Comm., vol. 19, no. 12, pp. 2121-2130, Nov. 2001. [15] D. Kandlur, K. Shin, and D. Ferrari, “Real-Time Communication

in Multihop Networks,” Proc. ACM SIGCOMM, pp. 300-307, 1991. [16] S. Lu and V. Bharghavan, “Fair Scheduling in Wireless Packet Networks,” IEEE/ACM Trans. Networking, vol. 7, no. 4, pp. 473-489, 1999.

[17] T. Minn and K.-Y. Siu, “Dynamic Assignment of Orthogonal Variable-Spreading-Factor Codes in W-CDMA,” IEEE J. Selected Areas in Comm., vol. 18, no. 8, pp. 1429-1440, Aug. 2000. [18] T.S.E. Ng, I. Stoica, and H. Zhang, “Packet Fair Queueing

Algorithms for Wireless Networks with Location-Dependent Errors,” Proc. INFOCOM, pp. 1103-1111, 1998.

[19] A. Parekh and R.G. Gallager, “A Generalized Processor Sharing Approach to Flow Control in Integrated Services Networks: The Single-Node Case,” IEEE/ACM Trans. Networking, vol. 1, pp. 334-357, June 1993.

[20] J. G. Proakis, Digital Communications, fouth ed. McGraw-Hill, 2001. [21] P. Ramanathan and P. Agrawal, “Adapting Packet Fair Queueing Algorithms to Wireless Networks,” Proc. ACM/IEEE MOBICOM, pp. 1-9, 1998.

Fig. 8. The impact of traffic models on average delay.

[22] F. Shueh, Z.-E.P. Liu, and W.-S.E. Chen, “A Fair, Efficient, and Exchangeable Channelization Code Assignment Scheme for IMT-2000,” Proc. IEEE Int’l Conf. Personal Wireless Comm. (ICPWC), pp. 429-436, 2000.

[23] Y.-C. Tseng and C.-M. Chao, “Code Placement and Replacement Strategies for Wideband CDMA OVSF Code Tree Management,” IEEE Trans. Mobile Computing, vol. 1, no. 4, pp. 293-302, Oct.-Dec. 2002.

[24] J. Wigard, N.A.H. Madsen, P.A. Gutierrez, I.L. Sepulveda, and P. Mogensen, “Packet Scheduling with QoS Differentiation,” Wire-less Personal Comm., vol. 23, pp. 147-160, 2002.

[25] L. Xu, X. Shen, and J.W. Mark, “Dynamic Bandwidth Allocation with Fair Scheduling for WCDMA Systems,” IEEE Wireless Comm., vol. 9, no. 2, pp. 26-32, Apr. 2002.

[26] Y. Yang and T.-S.P. Yum, “Nonrearrangeable Compact Assign-ment of Orthogonal Variable-Spreading-Factor Codes for Multi-Rate Traffic,” Proc. IEEE Vehicular Technology Conf. (VTC), pp. 938-942, 2001.

[27] L. Zhang, “Virtual Clock: A New Traffic Control Algorithm for Packet Switching Networks,” Proc. ACM SIGCOMM, pp. 19-29, 1990.

Chih-Min Chao received the BS and MS degrees in computer science from Fu-Jen Catholic University and National Tsing-Hua University in 1992 and 1996, respectively. He was with SENAO International in 1996. He received the PhD degree in computer science and information engineering from National Cen-tral University in January of 2004. He now works in the Department of Computer Science at National Taiwan Ocean University. His research interests include mobile computing and wireless communication, with a current focus on resource management in WCDMA systems.

Yu-Chee Tseng received the BS and MS degrees in computer science from the National Taiwan University and the National Tsing-Hua University in 1985 and 1987, respectively. He worked for the D-LINK Inc. as an engineer in 1990. He received the PhD degree in computer and information science from the Ohio State University in January of 1994. He was an associate professor at the Chung-Hua University (1994 to 1996) and at the National Central University (1996 to 1999), and a full professor at the National Central University (1999 to 2000). Since 2000, he has been a full professor in the Department of Computer Science and Information Engineering, National Chiao-Tung University, Taiwan. Dr. Tseng served as a program chair in the Wireless Networks and Mobile Computing Workshop, 2000 and 2001, as a vice program chair of the International Conference on Distributed Computing Systems (ICDCS), 2004, as a vice program chair of the IEEE International Conference on Mobile Ad hoc and Sensor Systems (MASS), 2004, as an associate editor for The Computer Journal, and as guest editor of special issues for several journals including ACM Wireless Networks, IEEE Transactions on Computers, Journal of Internet Technology, etc. He is a two-time recipient of the Outstanding Research Award, National Science Council, ROC, in 2001-2002 and 2003-2005, and a recipient of the Best Paper Award at the International Conference on Parallel Processing, 2003. Several of his papers have been chosen as selected/distinguished papers in confer-ences and been included for publications in journals. He has guided students to participate in several national programming contests and received several awards. His research interests include mobile comput-ing, wireless communication, network security, and parallel and distributed computing. Dr. Tseng is a member of the ACM and a senior member of the IEEE Computer Society.

Li-Chun Wang received the BS degree from National Chiao Tung University, Taiwan, in 1986, the MS degree from National Taiwan University in 1988, and the MSci and PhD degrees in electrical engineering from the Georgia Institute of Technology, Atlanta, in 1995, and 1996, respectively. From 1990 to 1992, he was with the Telecommunications Laboratories at the Ministry of Transportations and Communications in Taiwan (currently the Telecom Labs of Chunghwa Telecom Co.). In 1995, he was affiliated with Bell Northern Research of Northern Telecom, Inc., Richardson, Texas. From 1996 to 2000, he was with AT&T Laboratories, where he was a senior technical staff member in the Wireless Communications Research Department. Since August 2000, he has been an associate professor in the Department of Communication Engineering at National Chiao Tung University in Taiwan. His current research interests are in the areas of cellular architectures, radio network resource management, and crosslayer optimization for high-speed wireless networks. Dr. Wang was a corecipient of the Jack Neubauer Memorial Award in 1997 recognizing the best systems paper published in the IEEE Transactions on Vehicular Technology. He holds three US patents and one more pending. Currently, he is the editor of the IEEE Transactions on Wireless Communications. He is a member of the IEEE.

. For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib.