A Fast Scheduling Algorithm Considering Buffer Occupancy and Channel

Document Sample
A Fast Scheduling Algorithm Considering Buffer Occupancy and Channel Powered By Docstoc
					      A Fast Scheduling Algorithm Considering Buffer Occupancy and Channel
                Condition for High Speed Downlink Packet Access

                                  Hoang Nam Nguyen, Riaz Esmailzadeh, Iwao Sasase
                          Department of Information and Computer Science, Keio University
                                3-14-1 Hiyoshi, Kohoku-ku, Yokohama 223-8522, Japan
                        Email: nguyen@sasase.ics.keio.ac.jp, riaz@ieee.org, sasase@ics.keio.ac.jp

                         Abstract                            algorithms have been proposed based on the radio link
     High speed downlink packet access (HSDPA) has           condition of users. They include maximum carrier-to-
been introduced for beyond 3G mobile systems for             interference power ratio (MCIR), round robin (RR) and
providing high transmission rates by using adaptive          the proportional fairness (PF) and its variants
modulation and coding scheme (AMC) and fast                  scheduling algorithms [3, 4, 5]. In [5], performance
scheduling. In this paper, a fast downlink scheduling        comparison of these algorithms is given focusing on the
algorithm for HSDPA is proposed, which selects users         evaluation of user throughput. However, previous
for transmission based on their buffer occupancies as        studies do not focus on the issues of packet loss due to
well as channel conditions. Performance results              full queues or queuing delay expiration.
obtained by computer simulations show that the                    In this paper, we propose a new fast scheduling
proposed scheduling scheme can provide superior gains        algorithm based on queuing occupancy and radio link
in terms of low packet loss probability and high             condition, which is aiming at providing low packet loss
downlink throughput while at the same time providing         probability as well as high downlink, and good fairness
good fairness among users in terms of user throughput        in terms of user throughput and packet delay. The
and packet delay.                                            scheduling algorithm aims at choosing a user who has a
Keywords: HSDPA, fast scheduling, mobility                   higher probability of full queue occupancy based on the
                                                             information of queue‘s size and channel conditions.
                                                                  The paper is organized as follows. The system
                     1. Introduction                         model including basic concepts and principles of
     Current third generation (3G) mobile networks,          HSDPA is presented in Section 2. In the next section, a
UMTS (Universal Mobile Telecommunication System)             new fast scheduling algorithm is introduced and
or IMT-2000 (International Mobile Telephony) are able        described. Section 4 includes simulation results and
to provide downlink shared channels with transmission        discussions of performance comparison between the
bit rates of up to 2Mbps [WCDMA]. Wideband CDMA              proposed algorithm and others. Conclusion remarks are
(WCDMA) is the air interface for UMTS which                  drawn in the last section.
provides high downlink data transmission rates by using
time division shared channels. In order to offer higher                                  2. System model
data transmission rates, HSDPA is proposed for the new
release of WCDMA 3G networks [1]. In HSDPA, data                External Networks          Core Network                Radio Access Network
users share in a time division manner a common data
channel so-called high-speed downlink shared channel                Internet           GGSN                 SGSN        RNC
(HS-DSCH). With HSDPA, two fundamental features
of WCDMA, variable spreading factor (SF) and fast
power control, are disabled and replaced by other
                                                                 PLMN, PSTN            GMSC            MSC/VLR              RNC
advanced technologies. They include adaptive data
modulation and channel coding (AMC) with fast link             PLMN:   Public Land Mobile Networks
adaptation according to the channel condition, hybrid          PSTN:   Public Switched Telephone Networks
                                                               GPRS:   General Packet Radio Service
automatic repeat request (HARQ) and fast scheduling.           GGSN:   Gateway GPRS Support Node
The features are implemented in base stations (BSs)            SGSN:
                                                                       Serving GPRS Support Node
                                                                       Gateway MSC
                                                                                                                   Node B

instead of radio network controllers (RNC) to minimize         MSC:    Mobile Switching Center
                                                               RNC:    Radio Network Controller
the signaling and processing delay [2].                        HLR:    Home Location Register
     In HSDPA, fast scheduling, which is performed             VLR     Visitor Location Register

every 2ms, is a very important function of resource                       Figure 1: UMTS network architecture
management. It has to provide not only high system
throughput as well as high user throughput but also QoS         An UMTS mobile network consists of a core
of data services in terms of packet loss probability. A      network (CN) and a radio access network (RAN) [6] as
suitable downlink fast scheduling algorithms should          shown in Fig. 1. The CN connects the UMTS network
have a low computation load. Several scheduling              with other external networks while the RAN provides
radio access and transmission. The UMTS network is                                   Several AMC schemes are proposed including QPSK,
connected with other external networks such as public                            16QAM and 64QAM with coding rates of 1/3, 1/2 and
land mobile networks (PLMN), public switched                                     3/4. Depending on the channel qualities of a user, the
telephone system (PSTN) and the Internet by different                            base station will select a suitable AMCS configuration
technologies. A circuit-switched (CS) technology is                              for packet transmission. If the user has a good channel
used when the UMTS network connects with other                                   quality, it can receive packets at very high transmission
circuit-switched networks like exiting PLMN, PSTN                                rate by using 64 QAM with coding rate 3/4. In order to
whereas a packet switch technology is used for                                   increase transmission rate, multi-code can be used. H-
connecting with other packet-based networks like                                 ARQ consists of two schemes: chase combining (CH)
Internet. Therefore, the CN is divided into two separated                        and incremental redundancy (IR). In the CH scheme,
parts including Mobile Switching Center/Visitor                                  when a received packet has errors, the original packet is
Location Register-Gateway MSC (MSC/VLR-GMSC)                                     retransmitted and then combined with the received
for circuit-switched connections and Serving GPRS                                packet symbol-by-symbol before FEC decoding.
Support Node- Gateway GPRS Support Node (SGSN-                                   Meanwhile, in IR scheme, the retransmitted packet only
GGSN) for packet-switched connections.                                           carries redundancy information such as parity bits and
    The RAN with WCDMA air interface consists of two                             then combined with the received packet.
components: Radio Network Controller (RNC) and                                       At base stations, packets arriving from RNC are
Node-B (base station). Base stations perform the                                 classified and stored in appropriated queues. Each
functions of the air interface processing such as channel                        physical queue is assigned for a data flow. Each
coding and interleaving, rate adaptation and spreading.                          realtime (RT) flow is assigned a dedicated channel
It also performs some basic radio resource management                            whereas several non-realtime (NRT) flows consume a
operation as the inner loop power control. The RNC                               shared channel. To allocate downlink radio resource to
performs functions of mobility management and radio                              NRT users, base stations use fast scheduling methods to
resource management such as scheduling and admission                             select NRT packets for transmission with the radio
control. Because RNC is separated from Node-B, packet                            frame of 2 ms.
processing delay will occur.                                                            Two existing scheduling algorithms, which are the
                                                                                 maximum carrier-to-interference (MCIR) and the
    Fast Link Adaptation                                                         proportional fairness (PF) algorithms, are considered for
                                                      Channel quality feedback
                                                      HS-DPCCH, DCH
                                                                                 the performance comparison. In the MCIR scheduling
                                                                                 algorithm, the user, who has maximum SIR, is selected
     Fast Cell Selection                           Data channel
                                                   HS-DSCH                       for transmission. This might cause the situation that
                                                                                 users in worse channel conditions are not transmitted
        Hybrid-ARQ                                                               thus their queue are filled up by arriving packets. In this
                                                                                 case, queues shall become full thus the packet dropping
      Fast Scheduling                                                            probability will increase and the system throughput
                                                                                 might be decreased.
                                                                                     In the (PF) scheduling algorithm [4], the user, who
                                                                                 has the highest ratio of the requested rate ri and the

                                                   channels                      average_rate Ri(t), is selected for transmission. The
     From RNC                                                                    average_rate Ri(t) is defined as following:
                                                                         Tx       Ri (t + 1) = (1 − 1 )Ri (t ) + 1 * current_ tranmissio _ rate
                                                                                                     tc           tc
                                                                                 where tc is set to 1000 time slots.
                                                           channel               The average_rate Ri(t) is updated every transmission
                                                                                 frame. If a user does not get transmission, the
   Figure 2: Concepts of HSDPA and queuing model                                 current_transmission_rate is set to zero. The
                                                                                 performance of this heuristic approach depends on the
   An illustration of HSDPA is provided in Fig. 2.                               value of tc. The efficient value of this parameter is
Several data users share a data channel so-called high                           selected equal 1000 time slots according to recent
speed downlink shared channel (HS-DSCH) associated                               studies [5]
by a control channel for providing channel control and
retransmission information. In HSDPA, several                                           3. The proposed scheduling algorithm
advanced technologies including fast link adaptation,                               For a flow i, several parameters are taken into
fast cell selection, hybrid automatic repeat request (H-                         account while performing a scheduling algorithm: the
ARQ) and fast scheduling are provided at base stations                           transmission rate requested by the user (ri) which
in order to provide high downlink transmission rate and                          depends on the user‘s SIR, the queue threshold (qi-th)
reduce the inherent large delay of RNC-based                                     and the amount of packets stored in the queue (qi).
processing. The key idea of the HSDPA is to increase                             Packets arriving to queues will be dropped if the queues
packet data throughput by using adaptive modulation                              have no free spaces. Aiming at providing a high
and coding schemes (AMCS), an efficient                                          downlink throughput while at the same time
retransmission and a short transmission frame.
guaranteeing the lowest packet dropping probability, a       is observed whereas other cells are interference sources
fast scheduling algorithm namely low dropping                to the center cell.
probability (LDROP) is introduced. The originality of            Propagation model: The relative propagation loss is
this method is that it collects the information of buffer    as follows:
occupancies and channel conditions for not only                               ξ
minimizing packet losses of each user but also                     L = 10         10
                                                                                       r −α                 (1)
providing high speed access rates to users. At the time      where r is the distance (in km) between a mobile user
before a new frame starts, the algorithm is performed to     and a base station, α is the path loss factor with the
choose a user for packet transmission as follows:            typical value is selected in the range (2.7 – 4.0), ξ in dB
• Step 1: Eliminate users whose queues are empty.            is an Gaussian random variable with zero mean and a
Only active users, who have packets for transmission,        standard deviation σ represented for shadowing effects.
are taken into the selection procedure. Based on SIR            Mobility model: At the initiation of simulation
values updated by users, requested rates of users are        experiments, a number of mobile users (N) are
calculated.                                                  uniformly distributed in the center cell. A user moves
• Step 2: Check queue occupancies of the active users        randomly with a speed ν and a direction angle θ. Every
then classify them into two groups: over_threshold and       5 seconds, a mobile user changes its speed, which is
non_over_threshold. In the over_threshold group, the         following the uniform distribution of the range from 10
queue sizes of these users exceed their thresholds.          km/h to 36 km/h. The direction angle changes every 20s
• Step 3: Select a user for transmission                     randomly with the angle difference ∆θ is uniformly
                                                             distributed within [-90o, 90o]. In this paper, impacts of
IF            There are several users belonging to the       handover issues to scheduling are not considered i.e. the
              over_threshold group,                          system load is controlled stably. Assuming that
                                                             whenever a mobile user crosses the cell border, it will
THEN          Choose the user who has the maximum            change the direction angle with the absolute value of ∆θ
              value of qi.                                   of 180o in such a way as to go inside the center cell. It
                                                             means that at any given time, there is the same number
ELSE          Among     users    belonging    to    the      of active mobile users consuming the high speed
              non_over_threshold group, choose the user      downlink thus the offered load is controlled.
              having maximum value of qi/ri.                    Traffic model: Data connections (flows) are
     The step 1&2 can be carried out offline. Only step 3    simulated where the mobile users are data receivers
is performed online while scheduling. This scheduling        connected with external data sources. A data source
scheme aims at choosing within the over_threshold            generates traffic following an on-off model with the
group a user who has higher probability of the full          mean on-time and off-time durations are 4s and 2s
queue occupancy. Whereas among users belonging to            respectively. For the simplification, packets have same
the non_over_threshold group, the algorithm preferably       size and the packet size after coding is fit to a time slot
selects the user having higher probability of exceeding      of 2Mbps downlink.
queue’s threshold.                                              Buffer allocation: A fixed buffer allocation strategy
                                                             is exploited which has the queuing threshold is 90% of
                  4. Performance results                     total buffer spaces allocated to each flow. Each flow is
                                                             allocated 96 Kbytes buffer space.
4.1 Simulation configuration                                    Performance results: For every time interval T = 10s,
                                                             performance results are updated and at the end of
                                                             simulation program, the mean value is calculated.
                                                                • The mean downlink throughput (DL_throughput)
                                                                is the ratio of the total transmitted data during the
                                                                duration T and the time interval value.
                         C0                                                                   Total _ transmitted _ data
                                                                DL _ throughput =                                        (2)
                                               ν                • The packet dropping probability (Pdrop) is
                                                                measured as the number of dropped packets divided
                                           θ                    to the number of arrived packets.
                                                                          Number _ of _ dropped _ packet              (3)
                                                                Pdrop =
                                                                          Number _ of _ arrived _ packet
     Figure 3: The cell layout of simulation experiments
                                                                • The user throughput (User_throughput) is the
   Cell layout: Omni-cells are used in the simulation           ratio of the total received data of a user during the
experiments. A cluster of a mobile network is simulated         duration T and the time interval value.
consisting 25 cells as shown in Fig. 3. The center cell C0                                     Total _ received _ data
                                                                User _ throughput =                                    (4)
  • The average last hop delay (LH_delay): The last        4.2 Performance comparison
  hop delay of a packet is the sum of the queuing time         In the first simulation, the downlink throughput and
  and the transmission time. For each user, the average    packet dropping probability are evaluated and shown in
  last hop delay is calculated by summing up the last      Fig. 4 and Fig. 5, respectively. Each mobile user
  hop delay of all received packets those have arrived     maintains a data connection with another external end
  to the user during interval T and then dividing it by    user, which generates packets as the described on-off
  the number of arrived packets.                           model. The downlink load is varied by varying the
                  last_ hop_ delay of _ packet
                                  _                        number of active mobile users. The mobile users are
   LH_ delay=                                        (5)   moving according to the described mobility model.
                     _            _
                Number of _ arrived packet                     As shown in Fig. 4, when the system load is
                                                           increased, the proposed LDROP scheduling algorithm is
    Simulation parameters are shown in Table 1 and         able to provide a higher downlink throughput than other
Table 2, which are taken from [Kol01]. HSDPA uses          algorithms. The MCIR algorithm shows the worst
ten codes with the spreading factor of 16. The minimum     performance, especially its performance is decreased
transmission per code is 0.237 Mbps in the case of using   when the system load exceeds a particular value. As
QPSK wit coding rate ½. The maximum transmission           shown in Fig. 5, the LDROP algorithm provides a
per code is1.076 Mbps in the case of using 64QAM wit       superior performance of packet dropping probability
coding rate ¾. The central cell C0 allocates 80% of        where as the PF and MCIR algorithms show a high
maximum power for the high speed downlink shared           packet dropping probability when the number of mobile
channel. Other cells are assumed to transmit at the        users increases. This is because the MCIR algorithm
maximum power. Interference at the center cell C0 is       prefers to allocate radio resource to those users who
caused by 2-tier cells. In the simulation program,         have high quality channels. Therefore, the users, who
retransmission is not implemented i.e. the packets are     are located far from the base station, might not receive
assumed to be transmitted without errors. In the           transmission assignment that causes their queues
simulation, the channel loss caused by multi-path fading   become full resulting in the arriving packets are dropped
is also neglected. Five AMC schemes are used with the      and low downlink throughput. The PF algorithm
required SIR and the transmission rate per code as         allocates radio resource fairly to mobile users. However,
shown in Table 2.                                          it does not deal with the issues of overload queues and it
                                                           causes a high packet dropping probability. Therefore its
Table 1: Simulation parameters                             downlink throughput becomes smaller than that of
                   System parameters                       LDROP algorithm.
Number of cells            25                                  In the second simulation, we evaluate the impacts of
Cell radius                500 m                           channel quality to the performance of the user
Frame period               2 ms                            throughput and mean last hop delay. Here the system
Time slot per frame        3                               load is kept constant i.e. the number of mobile user is
Spreading factor           16 (with QPSK ½, peak data      fixed with 90 mobile users. Users are uniformly located
                           rate is 0.237 Mbps/code

                                                           in the center cell. Because the aim of this experiment is
Number of multi 10                                         to show how the proposed scheduling algorithm provide
codes                                                      fairness between mobile users, the mobile users are
Power allocated to 80%                                     fixed for maintaining stable channel quality..
DSCH                                                           As shown in Fig. 6, the PF and LDROP scheduling
                                                           algorithms can provide good user throughput to all
Path loss factor (α)       3.8
                                                           mobile users who have different channel quality. Good
Standard deviation of 8 dB
                                                           fairness can be achieved by using LDROP algorithm.
shadowing (σ)                                              The MCIR algorithm does not assign transmission to
                     Traffic generator                     users who have low channel quality so that their user
Non-realtime          (NRT) ON-time: 4s                    throughput can be zero. As shown in Fig. 7, the last hop
traffic using on-off model OFF-time: 2s                    delay provided by PF and LDROP algorithms is very
with fixed packet size          Peak data rate: 40 Kbps    similar. Mobile users experience nearly equal packet
                                                           delay. That means the LDROP algorithm is able to
Table 2: Selected MCS and their parameters                 provide very good fairness among mobile users in terms
Modulation and coding schemes for SF = 16                  of delay. For the MCIR algorithm, when the Eb/N0 value
MCS                Peak data rate Minimum                  is smaller than 1dB, no packets are transmitted and
                   per code          required SIR          therefore we have seen the packet delay is equal -1. The
QPSK ½             0.237 Mbps        -20 dB                mobile users having Eb/N0 between 1dB and 6dB have
QPSK ¾             0.356 Mbps        -16 dB                very low user throughput as shown in Fig. 6 because
16QAM ½            0.477 Mbps        -9 dB                 many of their packets have been dropped. For such
16QAM ¾            0.716 Mbps        -4 dB                 users, only few packets, which have arrived when other
64QAM ¾            1.076 Mbps        6 dB                  queues are empty, are transmitted. That is the reason
why the lower packet delay is shown for the mobile                                                                                                     5. Conclusions
users having such Eb/N0 values.                                                                                                       In this paper, a fast scheduling algorithm for
                                                                                                                                   providing high system performance while at the same
                                                                                                                                   time maintaining a low packet dropping probability has
                                                                                                                                   been proposed for HSDPA in the next generation of
                        HS-DSCH throughput (bits/s)

                                                      2.00E+06                                                                     WCDMA mobile networks. This scheduling scheme
                                                                                                                                   aims at choosing for transmission the user, who has a
                                                      1.50E+06                                                                     high probability of queue size exceeding the buffer
                                                                                                                                   threshold by using the information of buffer
                                                      1.00E+06                             LDROP                                   occupancies and channel conditions. The scheduling
                                                                                           MCIR                                    algorithm can provide superior gains in terms of high
                                                      5.00E+05                                                                     downlink throughput and very low packet dropping
                                                                                                                                   probability. It is also able to provide good fairness
                                                                                                                                   between mobile users in terms of user throughput and
                                                                               44        50     56        62       68     74
                                                                                                                                   mean last hop delay. Future work will extend the
                                                                                         Number of mobile users
                                                                                                                                   algorithm for handover scenarios to evaluate the
                                                        Figure 4: HS-DSCH mean throughput                                          impacts of handover to packet dropping probability. A
                                                                                                                                   dynamic buffer allocation will be investigated to
                Packet dropping probability

                                                                                    LDROP                                          optimize system performance.
                                                      0.15                          MCIR
                                                                                    PF                                                             Acknowledgement
                                                                                                                                     This work is partly supported by the grant of Japan
                                                                                                                                   Society for Promotion Science (JSPS) and the Keio
                                                                                                                                   University‘s COE program on “Optical and Electronic
                                                                                                                                   Device in Advanced Networks”.
                                                                     44         50         56        62        68        74
                                                                                Number of mobile users                                                   6. References
                                                      Figure 5: Mean packet dropping probability
                                                                                                                                   [1] 3GPP, “Physical layer aspects of UTRA high speed
                               30000                                                                                               downlink packet access (Release 2000)”, 3G TR 25.848
                                                                               LDROP                  MCIR                    PF   V0.6.0, March 2001.
 User throughput (bits/s)

                                                                                                                                   [2] A. Furuskar, S. Parkvall, M. Persson and M. Samuelsson,
                                                                                                                                   “Performance of WCDMA high speed packet data”, in the the
                                                                                                                                   Proc. of IEEE Vehicular Technology Conference, 2002 (VTC
                               15000                                                                                               Spring 2002), Birmingham, AL, USA, May 2002, Vol. 3, pp.
                               10000                                                                                               [3] Y. Choi, Y. Han, “A channel-based scheduling algorithm
                                                                                                                                   for cdma2000 1xEV-DO system”, Proc. Of IEEE Personal,
                                              5000                                                                                 Indoor and Mobile Radio Communications 2002
                                                                                                                                   (PIMRC02), Lisboa, Portugal, Sept. 2002, Vol. 5, pp. 2259-
                                                        0                                                                          2263.
                                                             -5       -2        -1        1    3    6    7          10    11       [4] A Jalali, R. Padovani and R. Pankaj, “Data throughput of
                                                                                          Eb/N0 value (dB)                         CDMA-HDR a high efficiency-high data rate personal
                                                                                                                                   communication wireless systems”, Proc. of IEEE Vehicular
                                                                      Figure 6: User throughput                                    Technology Conference Spring 2000 (VTC2000-Spring),
                                                        25                                                                         Tokyo, Japan, May 2000, Vol. 2, pp. 1854-1858.
                                                                                                                                   [5] Y. Ofuji, S. Abeta and M. Sawahashi, “Comparison of
                                                        20                      LDROP                     MCIR                PF   packet scheduling algorithms focusing on user throughput in
                 Mean last hop delay

                                                                                                                                   high speed downlink packet access”, IEICE Transaction on
                                                        15                                                                         Communications, Vol. E86-B, No. 1, Jan. 2003, pp. 132-139.

                                                                                                                                   [6] H. Holma and A. Toskala, “WCDMA for UMTS”, John
                                                        10                                                                         Wiley &Son Publisher, 2001, ISBN 0471486876
                                                                                                                                   [7] T. Kolding, F. Frederiksen and P. Mogensen,
                                                                                                                                   “Performance Evaluation of Modulation and Coding Schemes
                                                            0                                                                      Proposed for HSDPA in 3.5G UMTS Networks”, in the Proc.
                                                                -5        -2        -1    1     3     6        7    10    11       of the Symposium on Wireless personal Multimedia
                                                         -5                                                                        Communications (WPMC01), September 2001, Aalborg,
                                                                                          Eb/N0 value (dB)                         Denmark, Vol. 1, pp. 307-312.

                                                                Figure 7: Average packet delay