QoS Scheduling for Multimedia Traffic in Packet Data Cellular Networks

QoS Scheduling for Multimedia Traffic in Packet Data Cellular Networks Young-June Choi and Saewoong Bahk School of Electrical Engineering & Computer Science, Seoul National University, Korea Shillim-dong, Gwanak-gu, Seoul, 151-742, Korea Tel: +82-2-880-8434 Fax: +82-2-880-8198 E-mail: {yjchoi, sbahk}@netlab.snu.ac.kr Abstract— CDMA data networks such as cdma2000 1x EVDO are proposed in the midst of evolving to the 3rd generation wireless networks. Basically they use time division multiplexing and rate control that need a downlink scheduling to increase the system capacity, thereby being able to support high speed data rates. As the systems will eventually support multimedia and data traffic together, we need to have a proper criterion for scheduling that can count various service requirements such as delay and loss. Therefore we visit the concept of utility and opportunity cost considering these together. The opportunity cost is defined as the maximum utility lost among the other users by giving the current turn to a particular user. We design an algorithm to select a job for transmission with the maximum profit that is obtained by subtracting the opportunity cost from its expected utility. The simulation results show that it can support various QoS levels in terms of delay and loss for various traffic scenarios.

I. I NTRODUCTION Recently high data rate systems have been deployed for cellular networks. Examples of such systems are cdma2000 1x EV-DO (EV-DO or HDR) in 3GPP2 [1] and High Speed Downlink Packet Access (HSDPA) in 3GPP [2]. Since these CDMA systems have some similar properties like the fourth generation systems based on all-IP, they can be a good choice for packet cellular networks supporting various traffic. An important issue in these high speed wireless networks is to provide desired QoS for data traffic. Especially the CDMA data network needs to manage network resources deliberately because it shows very different characteristics from the voice network in many aspects. First, data traffic is asymmetric and the traffic volume for the downlink is much higher than that for the uplink. Second, there are so many different kinds of services such as HTTP, WAP, VoIP, real time video traffic etc. Each traffic type has its own requirements of delay and error rate depending on its own traffic source model. Last, data traffic is bursty on the whole. For these reasons, we focus on the downlink of wireless channels, and try to devise an efficient scheduling method to maximally utilize the wireless bandwidth and meet the user’s QoS requirements. Therefore in this paper we examine the downlink scheduling method to provide appropriate QoS for each user. In the CDMA voice network such as IS-95, a sender controls the transmission power to receive the required QoS at the frame level. The quality of a call is influenced by the signal

power from all other users and base stations (BSs) which become interference to its own signal. To support high data rates, CDMA cellular data systems use not the power control but the transmission rate control in the downlink. To maximize the cell throughput the BS transmits at the full power and the rate control substitutes for the power control [5]. Our system uses a set of discrete rates according to the EV-DO specification [1]. To increase the downlink capacity, CDMA data systems also allocate one channel to a BS multiplexed by time division instead of code division [1], [2]. Accordingly we will investigate the downlink scheduling system with time division multiplexing. As a criterion of scheduling, we use the concept of utility which was first applied for the field of networks in [9]. A utility function represents the level of user’s satisfaction with the specific QoS. [6], [10] described the utility functions by using SINR and channel rate respectively. As QoS of multimedia traffic is mostly characterized by the delay bound and the loss rate, we consider both for the design of the utility functions. Our scheduling scheme will choose a user with the maximum profit for transmission by counting the utility of others, thereby supporting the flexible QoS of the mixed traffic under time varying channel conditions. We organize the remainder of the paper as follows. Section II explains the system model and Section III describes the design principle of a utility function considering the QoS of traffic class. Section IV proposes our scheduling scheme using the utility function and the concept of opportunity cost. Then we show the performance of our scheme through simulations in Section V. Finally, Section VI concludes this paper. II. S YSTEM M ODEL We consider the downlink channel of CDMA data networks. The downlink channel is a single broadband link shared by all users in the cell and uses time slots. The choice of transmission rate depends on the physical layer technology and the channel condition. The BS estimates the channel condition by using the measurement feedback from the mobile terminal. Channel coding, hybrid ARQ, 4-slot interlacing, and various diversity schemes are combined to have successful transmission in the error-prone wireless channel [1]. We assume that a transmission rate can be appropriately selected by the well-known relationship between the transmission power and data rate [12],

Class Conversational Streaming Interactive Background

Utility

TABLE I T RAFFIC C LASSES [3]

Attributes of Traffic low delay/loss rates less sensitive to delay, high bandwidth bursty, moderate delay/loss rate highly tolerant to delay/loss rate

S E R = b . I Io W

0.5

(1)

where Eb is the received energy per bit, and Io is the total received interference per hertz, and S , R, and I represent the signal power, transmission rate, and interference respectively. W denotes the system bandwidth of the downlink channel. Eb /Io is determined by the desired frame error rate (FER) which is dependent on the modulation method. Each transmitter controls S to obtain the desired rate in the CDMA voice network, while it simply adjusts R according to the measured SINR in our model. Suppose that Pi is the transmission power at the downlink of BS i, and Gij is the channel gain from BS i to user j . If γ is the minimum required Eb /Io and ηj is the background noise of user j , then the maximum rate for user j , rj , is calculated by (1) as follows. rj =

Gij Pi W
. γ k
=j Gkj Pk + ηj

Class1 Class2 Class3 Class4

1.0

(2)

III. U TILITY D ESIGN A. QoS and Utility Functions The cellular data systems support various kinds of services such as VoIP, WAP and real time video streaming, and nonreal time data. Therefore we need to consider specific QoS requirements for each traffic. 3GPP and 3GPP2 have already classified these as conversational, streaming, interactive and background classes [3], [4]. Their characteristics are shown in Table I. Now we represent the QoS objective by using a utility function which quantifies the performance of the application level according to actually delivered service quality. The level of user satisfaction can be represented by the channel condition, delay, and loss rate [9]. As shown in Table I, multimedia traffic is generally sensitive to delay and loss. So we consider both delay and data loss for the utility function. For example class 2 has the delay bound of DBound in Fig. 1 and packets out of this bound will be dropped. There exists a relationship between the delay bound and the loss rate. In our algorithm, the utility with the offset value gof f set in Fig. 1, influences the loss rate and can sometimes be negative. The service priority is also affected by the utility function. For instance streaming traffic has priority over interactive and background traffics. Considering the priority among the traffic classes, we can have some utility functions as shown in Fig. 1.

0.0

goffset 0

100

200

D Bound

300

Delay(Slots)

Fig. 1.

Design of utility functions

We will simulate the scheduler for the mixed traffic according to these utility functions. B. Delay The delay for a user consists of access and propagation delay as well as queueing delay at the scheduler. As the propagation delay is fixed and the access delay is irrelevent to the downlink scheduling scheme, we leave both delays out of our concern. We also consider the loss only at the wireless link which often becomes congested. Now the issue is how we incorporate the delay into the utility function. It will be a function of the requested batch size from a user [7]. denote the delay required to serve residual packets Let dres j of job j under the assumption that job j is allowed to transmit all packets successively and that the average data rate remains constant during transmission [8]. If Jjres is the residual job is given by size of user j and Tslot is one slot time, dres j dres =
j

Jjres Bj

Sj Tslot

(3)

where Sj is the number of consecutive slots to transmit Bj bits. Sj and Bj are determined by the feedback information about the channel condition. That is, they determine the data rate. For example, if Sj , Bj , and Tslot equal 16, 1024bits, and 1.67ms respectively, the data rate is 38.4kbps. This allows the modulation scheme to easily use the hybrid ARQ [1]. When the scheduler adopts the policy of consecutive transmission which will be validated in IV, it can obtain the can be calculated estimated delay for each job because dres j from the modulation parameters. Let dj be the total estimated delay of job j . If job j has been already served for dcur j , dj can be written as dj = dcur + dres (4) j j . As job j will be able to get the utility for the current job at dj , the utility is not available at present. Therefore we call uj (dj ) as the potential utility because it is effective at time dj .

This algorithm does not guarantee the long-term optimality since it is impossible to find a long-term optimal value. This is because the scheduler cannot exactly forecast new arrivals and future channel conditions.

Termination of service

Job1 Job2

Fig. 2.

Job1

Job2

C. Delay Bound

The effect of successive transmission

IV. S CHEDULING S TRATEGY A. Scheduling Sequences When the average transmission rate for a job is fixed, it is efficient to transmit all the packets from that job continuously without being interrupted. This is because undoubtedly the utility in terms of delay is monotonically decreasing. We can see the effect of successive transmission in Fig. 1 and simply states it in the following theorem. Theorem 1: Assume that the average data rate for a job is constant and the utility function is monotonically decreasing. To maximize the sum of the potential utilities, the scheduler needs to assign slots consecutively for that job. Proof: To simplify our proof, we only consider the case of two jobs. Assume that job i is currently being serviced. Then we can express the sum of potential utilities as ui (di ) + uj (dj + dres i ) when there exists another job j . If job j uses one slot during i’s service time, the utility sum is given by ui (di + Tslot ) + uj (dj + dres i ) as job i completes its service one slot later. Obviously this is smaller than the previous one since the utility function is monotonically decreasing. Afterwards we can prove the theorem by induction. B. Algorithm Now we determine who gets the next turn at the scheduling. Generally it is good for the scheme to select a job with the maximum utility. As this scheme, however, does not reflect the utility of delay, it may result in the starvation of some other jobs with a low utility value. Therefore we introduce the concept of opportunity cost. It is defined as the maximum utility lost among the other users by giving the current turn to a particular user. Consider a case of three jobs. Let job j get the potential utility of 10 at the current scheduling, and jobs k and l lose the utilities of 6 and 8 respectively at the cost of choosing j first. Then the opportunity cost is 8. As the time division multiplexing system gives the transmission right to only a job, the other jobs should wait for the current job to finish its transmission. Therefore the opportunity cost caused by j can be expressed as follows. Cj = maxi
=j {ui (di ) − ui (di + dres j )}

for all i.

(5)

Fig. 1 introduced gof f set in the utility function to indicate the delay bound. It is desirable for the job approaching the delay bound to have high priority. Then the question is how we set gof f set at a proper value considering the delay bound appropriately. Our finding is that we can obtain a lower bound of gof f set by using the following theorem. Theorem 2: Let Aj be the set of jobs except job j . If job approaching the delay bound has priority over the other jobs, the condition of gjof f set larger than 2{supi∈Aj ui (0) − Bound )} is sufficient to guarantee the delay bound inf k∈Aj uk (Dk Bound . of job j Dj Proof: Assume that job j is the closest to the delay bound. Regarding another job i ∈ Aj with the maximum utility, gof f set should satisfy the following inequality condition according to the design principle of utility functions. j

of f set uj (dj ) − uj (dj + dres i )≥g

In order to give priority to j , the profit of j should be larger than that of any other job. If j has the opportunity cost caused by job k, the relationship between the profits of j and i is given by uj (dj ) − {uk (dk ) − uk (dk + dres j )}

≥ ui (di ) − {uj (dj ) − uj (dj + dres i )}

(6)

In this manner our scheme can choose a job with the maximum profit. (7) maxj Φj

(9)

Then we have a need to find the minimum value of gof f set satisfying the following condition. of f set

gj

≥ ui (di ) − uj (dj ) + uk (dk ) − uk (dk + dres j )

(10)

As the utility function naturally decreases with the increase of delay in the range of [0, DBound ], the maximal value of the right-hand term in (10) is given by Bound )} 2{ sup ui (0) − inf uk (Dk k∈Aj

i∈Aj

]

(11)

≥ ui (di ) − uj (dj ) + uk (dk ) − uk (dk + dres j )

Hence we can simply obtain the following inequality. of f set

Then we can denote the profit Φj by subtracting the opportunity cost from the utility of job j . Φj = uj (dj ) − maxi
=j {ui (di ) − ui (di + dres j )}

(8)

gj

≥ 2G

(12)

where Bound )} G
{ sup ui (0) − inf uk (Dk i∈Aj

k∈Aj

(13)

2000

groupA groupB groupC groupD groupE

0.8

1500

0.6 1000

group1 group2 group3 group4

500

Utility

Number of Jobs (delay