SOFT-QOS PROVISIONING USING THE TOKEN BANK ... - CiteSeerX

SOFT-QOS PROVISIONING USING THE TOKEN BANK FAIR QUEUEING SCHEDULING ALGORITHM William K. Wong1,2, Haiying Zhu1, Victor C. M. Leung2 1

Communication Research Centre, Ottawa, ON, Canada and 2 University of British Columbia, Vancouver, BC, Canada

Abstract -- Future generation wireless packet networks will support multimedia applications with diverse quality-of-service (QoS) requirements. Much of the research on scheduling algorithms has been focused on hard QoS provisioning of integrated services. Although these algorithms give hard delay bounds, their stringent requirements sacrifice the potential statistical multiplexing performance and flexibility of the packet-switched network. Furthermore, the complexities of the algorithms often make them impractical for wireless networks. There is a need to develop a packet-scheduling scheme for wireless packet-switched networks that provides soft QoS guarantees for heterogeneous traffic and is also simple to implement and manage. This paper proposes Token Bank Fair Queuing (TBFQ), a soft scheduling algorithm that possesses these qualities. This algorithm is work-conserving and has a complexity of O(1). We focus on packet scheduling on a reservation-based TDMA/TDD wireless channel to service integrated real-time traffic. The TBFQ scheduling mechanism integrates the policing and servicing functions and keeps track of the usage of each connection. We address the impact of TBFQ on mean packet delay, violation probability, and bandwidth utilization. We also demonstrate that due to its soft-provisioning capabilities, the TBFQ performs rather well even when traffic conditions deviate from the established contracts.

I. INTRODUCTION Within the boundaries of a wireless packet network, end-to-end quality-of-service (QoS) guarantees of multimedia services are desirable. This entails securing the QoS at each segregated segment of the entire wireless connection. In this paper we focus on the scheduling of wireless uplink and downlink traffic to guarantee QoS between the wireless terminals (WT) and the base station (BS), using the proposed Token Bank Fair Queuing (TBFQ) scheduling algorithm. Medium access control (MAC) protocols enable WTs at diverse locations within a cell to regulate the sending of their packets over the multiple access wireless links and manage network resources as efficiently and as fairly as possible. The MAC protocols

generally have a dominant effect on the ability of the system to deliver on a QoS contract [1][2]. For reservation-based MAC protocols, scheduling algorithms have a very important role to allocate bandwidth fairly and deliver QoS for each traffic connection. They are one of the key components in the provision of QoS parameters such as delay, delay jitter, packet loss rate or throughput.

Many new applications are driving

requirements for substantial changes in the current network infrastructure. Specifically, the emergence of applications with very different throughput, loss rate, delay or delay jitter requirements underscores the need of a network capable of supporting different levels of service, as opposed to a single, best-effort level of service. To meet the wideranging QoS requirements of these applications, new and more efficient traffic scheduling algorithms are needed. These schedulers operate across different sessions in order to ensure that QoS bounds are met. Schedulers can be classified as work-conserving or non-work-conserving. A work-conserving scheduler is never idle when there is a packet awaiting transmission. Generalized Processor Sharing (GPS), Packet-by-packet GPS, Weighted Fair Queuing (WFQ), Virtual Clock (VC), Self-Clocked Fair Queuing (SCFQ), Round Robin (RR) and it variant Deficit Round Robin (DRR) belong to this category. Contrary to workconserving schedulers, non-working-conserving ones, such as Hierarchical Round Robin (HRR), Stop-and-Go queuing (SGQ), and Jitter Earliest-Due Date (Jitter-EDD), can be idle even if there is a backlogged packet in the system since it may be expecting another higher priority packet to arrive. GPS [also know as fluid fair queueing, i.e., (FFQ)] is an idealized fluid-flow model which services all sessions simultaneously. WFQ is the packet version of the GPS, which simulates in the background to keep track of the virtual time function of each arriving packets. The packets are serviced based on their increasing order of their timestamp. Due to the increasing difference between WFQ and GPS as the number of connections increases, Worst-case Fair Weighted Fair Queueing (WF2Q) was introduced to overcome this problem, and it provides a closer emulation of GPS than WFQ does. TBFQ belongs to the work-conserving category. A good overview and comparison of these schedulers can be found in [3] and their comparison with TBFQ is shown in Table 1.

Table 1: Comparison of the properties of various schedulers.

Delay bound

Fairness

Complexity

PGPS (WFQ)

Small

Good

O(N)

Graceful degradation No

SCFQ

Large

Moderate

N(log N)

No

VC

Small

None

N(log N)

No

DRR

Large

Poor

O(1)

No

WF2Q

Small

Very good

O(N)

No

TBFQ

Small

Good

O(1)

Yes

Wireless links generally possess characteristics that are quite different from those of wired links. They are subject to time and location dependent signal attenuation, fading, interference, and noise, which result in burst errors and time varying channel capacities. Some examples of scheduling algorithms for wireless networks [4] are: Channel State Dependent Packet Scheduling (CSDPS), Idealized Wireless Fair Queuing (IWFQ), Channel-Condition Independent Fair Queuing (CIF-Q), Server Based Fairness Approach (SBFA), and Wireless Fair Service (WFS). For wireless CDMA networks, these scheduling are often applied: Packet-by-packet GPS, Scheduled CDMA, Dynamic Resource Scheduling (DRS), Wireless multimedia access control protocol with BER scheduling (WISPER) [5]. The major differences between scheduling in wired and wireless network are the considerations of adverse channel conditions, distributed channel access, and power consumption.

Many of the aforementioned scheduling algorithms, e.g., PGPS [6],

provide known delay bounds and throughput guarantees to traffic leaving a node by keeping track of the timing information of each arriving packet in a virtual time-stamping system. Packets in the outgoing queue are sorted and arranged, whenever the queue content changes, in such a way that their negotiated QoS are not violated. IWFQ is a realization of PGPS with compensation for channel errors. CSDPS is a wireless scheduling framework that can accommodate different service disciplines. It deals with bursty errors by avoiding them at the link layer − packet transmission is deferred if the channel state of a session is marked ‘bad’. Other algorithms such as WFS and CIF-Q use

similar virtual time-stamping techniques to determine service order of arriving packets. They are differentiated by how they compensate for erroneous transmissions. However, the transmission of time-stamps of packets originating from different WTs over the uplink wireless channel to the BS for centralized scheduling may not be practical due to MAC delays and transmission impairments. Other considerations, such as WT mobility, should also be taken into account if QoS is to be provided in a mobile network. Existing scheduling algorithms are mostly aimed at providing strict QoS guarantees, i.e., hard-QoS, to traffic streams by requiring them to conform strictly to predetermined traffic profiles, often specified by leaky bucket parameters. QoS of out-ofprofile traffic is not guaranteed, which could suffer substantial performance degradations, e.g., being dropped entirely, even though over a period of time the traffic stream may have underutilized its allocated bandwidth. Real-time schedulers such as Earliest Deadline First (EDF) [7] are designed for time-critical multimedia applications and for tasks sets with sophisticated characteristics. They can perform well in an ideal network with precise workload, but in a realistic network with shared bandwidth and unpredictable workloads, their performances may be poor. All schedulers will be challenged by the reality of unpredictable workloads. In practice it is difficult to predetermine profiles of real-time multimedia traffic, and out-of-profile degradations may be detrimental to the overall QoS experienced by the end-user. This motivates us to develop algorithms that have soft-QoS provisioning properties. We define soft-QoS provision of a session to be the graceful acceptance of traffic profile violation when excess bandwidth is available, provided the session does not exceed its bandwidth allocation in the long-term. This prevents sudden degradation of QoS experienced by end-user as a result of traffic profile violations. The significance of the scheme proposed in this paper is in maximizing the effective use of the wireless bandwidth while providing soft-QoS for heterogeneous applications. This is achieved with low computation complexity. We demonstrate these desirable behaviors of the TBFQ as a scheduling algorithm, by evaluating its performance for a generic reservation-based time division multiple access/time division duplex (TDMA/TDD) MAC scheme in next generation wireless networks under errorfree channel condition. This enables the performance evaluations to be focused on the

scheduling behaviors, and assume that proper bit-error rate QoS is provided by the physical layer. The performance of TBFQ is compared with common scheduling methods including PGPS and round robin (RR) to illustrate its effectiveness in meeting the above objectives. The generic TDMA/TDD MAC scheme applies to 2.5G systems such as the General Packet Radio Service (GPRS). Reservation-based MAC schemes are suitable for real-time applications. It reduces the link access delay for real-time data by leaving a minimal reserved signaling capacity available to a real-time connection when it is idle, so that the connection can immediately indicate the resumption of activity to the BS and request reserved data capacity. In this paper, bandwidth utilization and mean packet delay are used as parameters for comparison. The soft-QoS provisioning nature of TBFQ is studied using packet violation probability as a measure.

In section II, we present a wireless system

architecture scenario in which TBFQ is employed for scheduling both uplink and downlink heterogeneous traffic. In section III, the TBFQ algorithm is explained in detail, especially in terms of its soft-QoS provisioning mechanism. Simulation results are presented in section IV. Discussions and conclusions are given in section V.

II. SYSTEM ARCHITECTURE

A. The TDMA Frame Structure and MAC Protocol In this study, we consider packet transfer in both uplink and downlink wireless channels. We assume that the quality of the wireless link is managed by the physical layer, and the link is framed for TDMA/TDD. The TDMA/TDD protocol has a number of attractive features, including the possibility of “on-demand” allocation of bandwidth. Our design is based on the TDMA/TDD frame structure shown in Figure 1. The fixed length TDMA frame is time-duplexed into an uplink and downlink channel, each further divided into control and data transmission periods. Slots assigned for control purposes are divided into control mini-slots each holding a control packet. The BS has absolute control of the numbers of data/control slots in each frame and the WTs assigned to receive or send information during the data slots.

The modem preamble is used for radio physical layer functions. The BS uses downlink control packets to announce slot allocations and assignments to the WTs. Each downlink data slot holds one data packet with link layer and physical layer overheads for error correction and detection, etc. We assume that the packets are received with sufficient signal-to-noise-plus-interference ratio that, in combination with FEC coding, give a wireless data channel that meets the bit-error-rate QoS required by the respective services. Hence packet losses in our analysis are caused only by buffer overflow or by real-time (voice, video) packets being dropped by the system due to violation of their delivery time constraints. We use the following generic MAC protocol to compare the scheduler performance. WT-originated packet transfer is achieved by a two-phase procedure. In the first phase, a WT establishing a new connection with the system accesses the uplink of this TDMA system using a slotted ALOHA based reservation protocol. The WT sends in a control mini-slot a packet_channel_request, containing the number of time slots required for each uplink frame for its connection. The BS responds by a packet_channel_assignment message, with the assigned uplink slots for the WT. Initially (immediately after connection establishment), the effective bandwidth [8] is assigned to the WT. Periodically (e.g., every frame time) the BS checks for packet_channel_request messages, and schedules the necessary time slots to each WT if there are still enough time slots for the connection for the next uplink frame. In the second phase, each WT can communicate through in-band signaling (using compressed control information) to the BS for dynamic allocation of more (or less) time slots as long as there are buffered data in the WT. Sometimes a connection requires no slot at all. In that case, the scheduler assigns no slot to it in the subsequent frames. A 1-bit energy burst with a unique position for each admitted WT in each uplink control period is introduced for in-band signalling, such that when that WT has data in its buffer again, it sends the energy burst to signal the BS in the next uplink control slot period [9]. So, whenever a burst occurs in a specific position, the BS knows which previously idle connection has data to send. In the next downlink period the BS allocates the contracted bandwidth immediately. This allocation

may be modified in subsequent frames depending on the connection’s actual requirement using the above signalling mechanism.

B. System Architecture A simple model of the BS architecture relevant to uplink packet transmission is shown in Figure 1. This model comprises two functional blocks: uplink and downlink schedulers. In both schedulers, the policing mechanism is included. However, for uplink traffic, the policing is enforced at each WT (hence virtual). The uplink scheduler in the BS makes scheduling decisions based on the queue information or status received from WTs. The policing device serves to regulate the traffic source and ensures that the traffic conforms to its traffic parameters negotiated at the connection establishment phase. One commonly used source policing mechanism is the Leaky Bucket (LB).

C. Traffic Models

Voice Traffic The widely used two-state on-off voice activity model with exponentially distributed duration of voice spurts and gaps [10][11][12] is applied to voice sources. A voice source generates a signal that follows a pattern of talk-spurts and silent gaps, which may be represented by negative-exponentially distributed ON and OFF periods. It has been determined experimentally that the average talk-spurt interval falls between 0.4-1.2 sec and the average silent interval falls between 0.6-1.8 sec. Within each talk spurt intervals, 64 Kbps PCM-coded digital voice is assumed. To be compliant with [12] as specified for mixed traffic, the mean duration of a talk-spurt and the subsequent silent interval are set to 3 sec, which means voice activity factor is 50%.

Video Traffic In the literature, video traffic models ranging from classical models based on Poisson arrival processes to sophisticated models like autoregressive processes/Markov chains [13] and self-similar models [14] are used. Most of these models were derived from the data of MPEG video clips, and MPEG streams generally require a high data rate

to sustain the quality of playback; therefore their suitability for the wireless medium is questionable. For these reasons, we chose to use real data streams from videoconference employing the ITU-T H.263 standard. The parameters of the H.263 source are shown in Table 2. Table 2: Parameters of an H.263 video source

Encoding Format Maximum Frame Size Minimum Frame Size Frame Rate Mean Bit Rate Total Number of Frames Encode Frames

352 x 288 66446 bits 1615 bits 15 Hz 90.42 Kbits/sec 500 I:1 P:499 B:0

World Wide Web Data Traffic The World Wide Web traffic model used is an ON-OFF two-state model with the ON period consisting of a sequence of document page transmission request from an individual user [15]. The length of the ON period has a Weibull distribution p( x) =

k x k −1 −( x /θ )k ( ) e θ θ

with k between 0.91 and 0.77, and θ between e4.4 and e4.6 .The length of the OFF period has a Pareto distribution

αk α p ( x) = α +1 x with α between 0.9 and 0.58, and k=60 sec. Within the ON period, the page inter-arrival time also has the Weibull distribution with k=0.5 and θ=e1.5. The size of the document page has the Pareto distribution with α