167
`
`TCP Performance in WCDMA-Based Cellular Wireless IP Networks
`Ekram Hossain, Dong In Kim*, Chris %holefield+ and Vijay K. Bhargava
`Department of Electrical and Computer Engineering
`University of Victoria, PO Box 3055 STN CSC
`Victoria, B.C. V8W 3P6, Canada
`Tel: +1-250-72 1-86 17, Fax: + 1-250-721-6048
`Email: {ekram, bhargava}@ece.uvic.ca, dikim@uoscc.uos.ac.kr, chriss@ipsg.mot.com
`
`M e a : Terms- Dynamic rate adaptation, variable
`spreading factor WCDMA system, TCP.
`Abstract- The performance of TCP (Transmis-
`sion Control Protocol) is evaluated for downlink
`data transmission in a cellular WCDMA (Wide-
`band Code Division Multiple Access) network
`where variable rate transmission is supported at
`the RLC (Radio Link Control)/MAC (Medium Ac-
`cess Control) level. A simple rate search proce-
`dure (based on the number of simultaneous TCP
`connections and the channel condition) is proposed
`for the RLC/MAC level transmission rate selection
`for downlink data transmission. The dynamic rate
`variation is assumed to be achieved by using single-
`code transmission with variable spreading factor,
`where the spreading factor varies inversely with the
`transmissiog rate. A novel ‘mean-sense’ approach
`to calculate inter-cell interference in such an envi-
`ronment is developed assuming homogeneous traf-
`fic load (in terms of active TCP connections) in
`different cells. The impact of the different physi-
`cal layer, RLC/MAC layer and TCP parameters on
`the end-to-end throughput performance are inves-
`tigated by simulating the system dynamics under
`multiple concurrent TCP connections.
`
`I. INTRODUCTION
`WCDMA systems (e.g., ETSI WCDMA, cdma2000)
`will be the major radio transmission technologies for
`XMT-2000 (International Mobile Telecommunications
`2000). The core network in an IMT-2000 system will
`be IP (Internet Protocol) based, which will evolve
`through GSM/GPRS-based core network architecture
`[l]. Standardization efforts in 3GPP (Third Gener-
`ation Partnership Project) are now directed in this
`directton. Since the transport layer protocol perfor-
`This work was supported in part by a research grant from
`the Motorola Wireless Data Systems Division, Richmond, B.C.,
`Canada and in part by the Korea Science and Engineering Foun-
`dation (KOSEF) under Grant 97-0101-0501-3.
`Prof. D. I. Kim is with the University of Seoul, Dept of
`Electrical Eng, 90 Jeon Nong-dong, Dong Dae Moon-gu, Seoul,
`Korea 130-743.
`+ Dr. C. Scholefield is with Motorola Wireless Data Systems
`Division, 11411 Number Five Road, Richmond, B.C. V7A 423,
`Canada.
`
`mance is one of the most critical issues in data net-
`working over noisy wireless links ([2]-[3]), the perfor-
`mance of TCP, which is the flagship protocol in to-
`day’s Internet, would be crucial in such an environ-
`ment.
`The performance of TCP for wireless Internet ac-
`cess in a packet-switched cellular WCDMA environ-
`ment would depend on the service provided by the
`underlying RLC/MAC protocol which allows variable
`rate packet data transmission for achieving high ra-
`dio spectrum utilization. Interference calculation for
`TCP performance evaluation is non-trivial in such
`an environment, particularly for heterogeneous traf-
`fic load in different cells.
`In this paper, the performance of TCP is evalu-
`ated for downlink packet data transmission in cellu-
`lar WCDMA networks assuming homogeneous traf-
`fic load (in terms of active TCP connections) in the
`different cells. The RLC/MAC layer frame transmis-
`sions are assumed to be based on a sub-optimal rate
`selection procedure and the transmission rate to an
`MS (Mobile Station) can be controlled on a frame
`by frame basis, depending on the number of concur-
`rent TCP connections passing through the BS (Base
`Station). Variable rate transmission to a mobile sta-
`tion is assumed to be achieved through single-code
`transmission with variable spreading gain ([4]-[6]). A
`novel ‘mean-sense’ approach is used for inter-cell inter-
`ference calculation under such a condition where the
`number of instantaneous TCP connections in a cell
`may vary while the average number of active TCP
`connections over all the cells is assumed to be kept
`fixed’. Performance evaluation is carried out for wide-
`area TCP connections for two different wireless chan-
`nel models ([11]-[7]), namely, multipath channel with
`equal path gain and multipath channel with unequal
`path gain.
`The impacts of channel impairments and different
`physical layer parameters along with the effects of
`variations in the different TCP parameters, e.g., TCP
`
`’This can be done by the RNC (Radio Network Controller)
`in a WCDMA network.
`
`0-7803-5893-7/00/$10.00 0 2000 IEEE
`
`Authorized licensed use limited to: Klarquist Sparkman LLP. Downloaded on May 05,2020 at 18:37:19 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`168
`
`window size ( W ) , transmieeion delay in the wired-
`network or Internet delay (&d)
`on the throughput
`of a target TCP connection are investigated.
`The issue of transmission rate variability at the
`RLC/MAC layer (and hence the issue of interference
`modeling under variable rate transmission) was not
`considered in the earlier works on wireless TCP in a
`DS-CDMA environment (e.g., [8]-[9]). In addition, in
`most of the cases, system dynamics under single TCP
`connection was dealt with.
`
`11. SYSTEM MODEL
`
`A . TCP Model
`A Ccoded event-driven simulator that supports the
`TCP congestion and error control algorithms, includ-
`ing slow-start, congestion avoidapce, fast retransmit
`and fast Tecowery, is used to imitate the TCP behavior.
`The RTO (Retransmission Time Out) value is based
`on RTT (Round Trip Time), which is estimated by ex-
`ponential weighted averaging. The RTT is measured
`for each successfully received unretransmitted packet.
`Based on these measurements, a moving estimate of
`the round trip time (RTTm,,,) is made according to
`the following: RTTi+' = c x RTTmeoszLred + ( 1 - c) x
`RTT'. The value of the low-pass filter gain param-
`eter c is assumed to be 0.1 in this paper. Conven-
`tional TCP implementations update the RTO value
`as the sum of the smoothed average and four times
`the standard deviation of the RTT values (Jacobson's
`Algorithm).
`A sender-based one way traffic scenario is consid-
`ered where the mobile stations act as TCP sinks. It is
`assumed that the ACKs generated by the TCP sinks
`are not dropped in the wired part of the network al-
`though the ACKs may be lost in the wireless link.
`The ACK packets undergo no extra queueing delays
`in addition to the propagation delay. All of the TCP
`connections in a particular simulation scenario are as-
`sumed to experience the same Internet delay.
`B. Radio Link Layer Model
`A TCP data- unit is typically equivalent to several
`RLC/MAC layer frames. For example, in IS-99, the
`CDMA circuit-mode data transmission standards, the
`radio link layer frame size is 24 bytes (with 19 bytes
`of payload) and the TCP segment size is 536 bytes.
`In a GPRS-based UMTS environment, a TCP packet
`will be forwarded to the RLC/MAC layer after IP,
`SNDCP (Subnetwork Dependent Convergence Proto-
`col) and LLC (Logical Link Control) overheads have
`been appended to it.
`
`During each frame-time (or timeslot), the downlink
`transmission rate to a mobile (i.e., the number of ba-
`sic link layer frames) is determined by a sub-optimal
`rate selection procedure (as will be described later)
`based on the number of simultaneous TCP connec-
`tions present at that time. The number of RLC/MAC
`layer frames to be sent in a timeslot is varied accord-
`ing to the transmission rate used.
`Let us assume that the length of a slot, denoted
`by T,, is fixed and that the transmission rates can be
`selected from the set of rates {VI, w2,. . . ,U,+,}.
`If it is
`assumed that 21, = mq (m = 2,. . . , 'p) 2, and that
`only one frame (of fixed length of M bits) correspond-
`ing to the smallest (basic) rate 'u1 can be sent in a slot,
`then for a transmission rate of um, m frames can be
`sent in a timeslot. To achieve this variable rate trans-
`mission, a variable spreading factor (VSF) WCDMA
`system is used where the basic factor is given by N
`chips per bit, and for rate um, the spreading factor is
`reduced to Nlm chips per bit. Note that the signal
`energy per bit (Eb) is kept constant to achieve almost
`the same bit error rate (BER) performance.
`The best performance can be obtained if the si--
`multaneously transmitting users choose the optimum
`combination of transmission rates. The RLC/MAC
`level throughput (p) can be expressed by
`
`~
`
`P(n1,nz,...,n$4 = mnmPc,m
`m= 1
`
`'p
`
`frames/slot
`
`( 1 )
`where the probability of corre'ct reception of a frame
`PC,, is dependent on the physical layer BER (Bit Er-
`ror Rate) performance Pb,m (and hence on the chan-
`nel interference and fading conditions) and on the er-
`ror control protocol employed. Channel interference
`conditions largely depend on the dynamically selected
`transmission rates in the different cells. The proba-
`bility of correct reception of a frame Pc,m for a user
`transmitting at rate 21, can be expressed as
`
`where M is the frame size (in bits) corresponding to
`the smallest (basic) rate V I .
`A one-step rate selection procedure [lo] is used here
`for a specified rate vector v assuming n simultaneous
`TCP connections. The rate selection procedure d e
`terminesthe value of 21,
`for which the RLC/MAC
`2For the rest of the paper, it is assumed that wm = mv1 so
`that the normalized value of U,,, with respect to the basic rate
`VI is m.
`
`Authorized licensed use limited to: Klarquist Sparkman LLP. Downloaded on May 05,2020 at 18:37:19 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`layer throughput (in terms of the number of ba-
`sic RLC/MAC frames) is maximized when the same
`transmission rate 'U,
`is selected for all the TCP con-
`nections.
`An ACK/NAK-based RLC/MAC level error con-
`trol is assumed where the error recovery is similar
`to the conventional SR-ARQ protocol although it is
`more complex due to variable rate transmission. To
`illustrate, flow diagram of the SR-ARQ protocol un-
`der variable rate transmission is depicted in Fig. 1 for
`transmissions from a single user, assuming that the
`acknowledgement delay is Ts (i.e., one timeslot) and
`that the data rate vm can be selected from the set
`{VI, w2, vq}. Here, a heavy trafFic condition is assumed
`so that newly generated frames are always available
`and can be transmitted along with the old frames
`(which are being retransmitted) so that the variable
`rate transmission can be fully exploited from slot to
`slot.
`
`Fig. 1. SR-ARQ based RLC/MAC level error control in
`variable rate transmission systems.
`
`RLC/MAC level error recovery continues until
`kaz transmission failures occur in succession where
`all the frames transmitted an a timeslot are garbled.
`At this point the whole TCP packet is discarded and
`the TCP at the sending host performs end-to-end er-
`ror recovery. The rationale behind this assumption
`is, in the case that all the transmitted frames in a
`timeslot have not failed, the channel condition can be
`assumed to be still not too bad so that persisting on
`retransmission may be beneficial to avoid end-to-end
`error recovery.
`111. DOWNLINK SINR MODELING AND BER
`CALCULATION
`Assume that the transmitted signal from the BS
`is affected by the L-path Rayleigh fading with path
`gains {al} with Cklu; = 1, a; = E{af} (E de-
`notes an expectation). All mobiles are assumed to
`be 'uniformly located in a cell. For synchronous
`transmission in the downlink of a target cell (while
`asynchronous transmission is assumed between cells),
`with rate selection being independent of the loca-
`tion of their corresponding mobiles, the average out-
`put signal-to-interferenceplus-noise-ratio (SINR),,,
`(1)
`
`169
`
`for the lth path of ith mobile can be modeled as (3)
`[lo1 *
`
`p b = [;(l-C)]
`
`L L - l
`
`L - 1 + 1
`
`Here, q is the frequency reuse factor in downlink,
`A
`x = Cg=, mn, with nm = number of mobiles choos-
`ing the rate vm, G = average offered load (i.e., aver-
`age number of active TCP connections per cell) and
`a
`a = E { m } , which accounts for the average rate se-
`lection in other cells. Note that the background noise
`with p.s.d. N,/2 is also included, but the transmis-
`sion power of any of the control channels, e.g., the
`pilot channel, is not included in (3).
`In t,he case of multipath Rayleigh fading channel
`with equal path gain (channel model-A), an L-path
`Rayleigh fading channel with uncorrelated scattering
`and equal average path power is considered. For this
`model, the average output (SINR),,, is given by (3)
`( 1 )
`with 012 = 1/15 and the BER Pt, is given by [?]
`(
`rm-0,i/[2
`+ (SnvR),,J.
`where C =
`In the case of multipath Rayleigh fading channel
`with unequal path gain (channel model-B), the BER
`can be expressed as [7]
`
`1=0
`
`Here, (SINR),,, is as given in (3).
`
`IV. SIMULATION OF T C P
`A. Simulation Assumptions and Parameters
`The buffer size at the BS is assumed to be infinite
`so that packet loss can occur only due to errors in
`the wireless channel. Per-destination queueing with
`FIFO (First In First Out) scheduling (within each
`queue) is assumed for the TCP packets at the BS. The
`RLC/MAC layer frames corresponding to the TCP
`packets are buffered in the link layer queues (Fig. 2).
`From each of the RLC/MAC frame queues, m, frames
`are transmitted simultaneously during each timeslot,
`
`Authorized licensed use limited to: Klarquist Sparkman LLP. Downloaded on May 05,2020 at 18:37:19 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`170
`
`where m, is determined using the rate selection pro-
`cedure previously described. The set of permissible
`rates is assumed to be {VI, v2,. - , vg} with v, = mvl.
`Mobile specific channelization code and the cell spe-
`cific scrambling code are used for downlink transmis-
`sion.
`
`time = 5 ms.
`The performance metric being considered here is
`the TCP throughput ( h c p ) , which is measured as
`the total number of bytes received per unit time at
`the TCP sink. .
`
`...
`
`Fig. 2. Queueing of TCP packets and FUCfMAC layer
`frames at the BS for downlink transmission.
`
`B. Simulation Results and Discussions
`P+--
`The impact of variations in the frequency reuse fac-
`tor (outer-cell interference coefficient) q and the av-
`erage offered load G on TCP throughput are demon-
`started in Figs. 3 - 4. With a TCP packet size of 576
`bgltes and a RLC/MAC frame size of 16 bytes (corre-
`sponding to the basic rate 211 and frame period of 10
`ms) , the maximum per-connection throughput achiev-
`able in an error-free channel would be 92.16 Kbps. Be-
`cause the RLC/MAC level rate selection would allow
`The different physical layer simulation parameters
`8 frames to be transmitted within a frame-time (i.e.,
`are: maximum spreading gain (N), outer-cell inter-
`m, = 8) in this case. For a single TCP connection,
`ference factor (q), FEC (Forward Error Correction)
`with W = 5 and Dwd = 50 ms, here, TCP throughput
`parameter ( t ) , ratio of bit energy and AWGN noise
`h c p M 84, 80 Kbps for channel model-A and channel
`spectral density (&/No), number of resolvable multi-
`model-B, respectively. As the number of simultane-
`paths (L) and tapped-delay-line parameters (for chan-
`ous TCP connections (N,) increases, throughput per
`nel model-B). Parameters for channel model-B are
`connection ( h c p ) falls off rapidly. The observed im-
`based on the vehicular-B model [12] for macro-cell.
`provement in f i c p due to a reduction in the outer-cell
`The different simulation parameters at the TCP
`interference coefficient and/or average system load is
`and RLC/MAC level are: system load G in terms
`not significant.
`of the average number of TCP connections in the net-
`The effects of variations in &/No and t on B c p
`are demonstrated in Figs. 5 - 6 for channel model-B.
`work, TCP segment size (MSS), TCP maximum win-
`dow size (W), TCP fast retransmit threshold (K),
`TCP throughput improves when the ratio of received
`TCP timer backoff parameter (Q), initial RTO value
`bit energy (&) to background noise power spectral
`(RTOinitial), allowable transmission rates (v), wired-
`denqity (N,/2) is increased, but the rate of improve-
`network (or Internet) delay ( D w d ) , number of concur-
`ment diminishes with increasing &,/No. For small
`rent TCP connections (N,J, fast retransmit parame-
`number of connections, an increase in the value of the
`ter (K), RLC/MAC layer frame size M corresponding
`FEC parameter (t) may reduce the per-connection
`to the basic rate 211, frame-length in time (T,), and
`effective TCP throughpup (Fig. 6). This would be-
`the maximum number of retransmissions allowed at
`come more pronounced as the maximum TCP window
`the RLC/MAC layer ( G a Z ) . The assumed values for
`size W decreases and/or Internet delay Dwd increases.
`some of these paraemters are: MSS = 576 bytes, K
`But when the number of concurrent TCP connections
`= 3, Q = 2.0, M = 16 bytes, T, = 10 ms, N = 128,
`is relatively large, the effective throughput increases
`and L = 4 (for channel model-A).
`monotonically with t . The effect of variation in &
`The parameter l?&=
`is to be chosen properly so
`on p T c p is shown in Fig. 7.
`that the RLC/MAC layer retransmissions do not in-
`The effects of the maximum TCP window size (W)
`terfere with the end-to-end retransmission. The trans-
`and the Internet delay (Dwd) on TCP throughput un-
`mission time for a link layer ACK is assumed to to be
`der varying number of simultaneous TCP connections
`one timeslot (2'').
`A similar assumption is made on
`are demonstrated in Figs..8 - 9. As is evident from
`the transmission time of TCP ACK from the mobile
`Fig. 8, for a particular value of Dwd, the value of W
`station. The parameter D,d accounts for the delay
`can be selected such that the throughput is maxi-
`experienced by the TCP packets in the wired part of
`mized. If t ~ c p
`is the transmission delay of a TCP
`the network and hence affects the RTT value of the
`
`Authorized licensed use limited to: Klarquist Sparkman LLP. Downloaded on May 05,2020 at 18:37:19 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`packet (from BS to MS) and t A C K is the correspond-
`ing ACK delay (from MS to BS), then the minimum
`value of the maximum TCP transmission window size
`Wmin4, for which the ‘packet pipe’ corresponding to
`a connection would be full, is given by
`
`As the number of simultaneous connections in-
`creases, t T C p increases and hence Wm,, decreases.
`Again, large values of W may cause more frequent
`invocation of the TCP ‘congestion control’ mecha-
`nism and more timeouts, and consequently, the packet
`transmission rate corresponding to a TCP connection
`may decrease. This causes degradation in the achieved
`throughput performance.
`As Dwd increases, the value of W need to be in-
`creased to keep the transmission ‘packet pipe’ full. For
`this reason, for a particular value of W , throughput
`falls off as D w d increases, especially when the num-
`ber of connections is small (Fig. 8). As the number
`of TCP connections increases, the performance differ-
`ence becomes insignificant, because due to the smaller
`rate selection at the RLC/MAC layer, t T C P increases
`in this case. As a result, for a particular W , even if
`increases, the corresponding values of Wmin (as
`D,d
`given in (6)) remain smaller than W . Under such a
`condition, connections with different round trip delay
`experience the same throughput.
`V. OUTLOOK
`The performance of TCP has been evaluated in
`a WCDMA-based cellular network which supports
`variable rate transmission at the RLC/MAC layer.
`The dynamic rate selection at the RLC/MAC layer
`is based on a one-step sub-optimal rate search pro-
`cedure. A homogeneous traffic scenario has been
`assumed where the mean number of TCP connec-
`tions averaged over different cells is not time-varying,
`although the instantaneous number of connections
`in a cell can vary. The impacts of different phys-
`ical, link and transport layer parameters on the
`achieved throughput performance have been evalu-
`ated. Since $he number of concurrently active TCP
`connections determines the RLC/MAC layer rate se-
`lection and hence the achieved TCP throughput, the
`TCP throughput for wide-area TCP connections can
`be improved by employing ‘intelligent’ RLC/MAC
`layer scheduling policies. We are currently investi-
`gating this issue.
`4For this value of W, the BS transmitter would never remain
`idle.
`
`171
`
`REF E RE N c E s
`[l] R. Pichna, T. Ojanpera, H. Posti and J. Karppinen, “Wue-
`less Internet - IMT-2000/wireless LAN interworking,” Jour-
`nal of Communrcations and Networks, vol. 2, no. 1, Mar.
`2000, pp. 46-57.
`[2] M. Zorzi and R. R. Flao, “The effect of correlated errors on
`the performance of TCP,” ZEEE Communications Letters,
`vol. 1, pp. 127-129, Sept. 1997.
`[3] A. Kumar, “Comparative performance analysis of versions
`of TCP in a local network with a lossy link,” ZEEE “ s a c -
`tions on Networking, vol. 6, no. 4, pp. 485-498, Aug. 1998.
`141 E. Dahlman, P. Beming, J. Knutsson, F. Ovesjo, M. Persson
`and C. Robol, “WCDMA - The radio interface for future
`mobile multimedia communications,” IEEE Zhnsoctions on
`Vehicular Technology, vol. 47, no. 4, Nov. 1998, pp. 1105-
`1118.
`[5] E. Dahlman and K. Jamal, “Wide-band services in a DS-
`CDMA based FPLMTS system,” Proc. ZEEE VTC’96, pp.
`1656-1660.
`[6] M. 0. Sunay and J. Kiihtava, “Provision of variable data
`Rates in third generation wideband DS CDMA systems,”
`P ~ c .
`IEEE WCNC’99, pp. 505-509.
`[7] J. G. Proakis, Digital Communications, 2nd ed. New York:
`McGraw-Hill, 1989.
`[8] A. Chockalingam and G. Bm, “Performance of TCP/RLP
`protocol stack on correlated fading DS-CDMA wireless
`links,” Proc. IEEE VTC’98, pp. 363-367.
`[9] Y. Bai, G. Wu and A. T. Ogielski, “TCP/RLP coordina-
`tion and interprotocol signaling for wireless Internet,” Proc.
`ZEEE VTC’99, pp. 1945-1951.
`[lo] E. Hossain, D. I. Kim and V. K. Bhargava, “Link level
`and transport level throughput performance in variable
`spreading factor WCDMA networks,” to be submitted to
`ACM/Baltzar Wireless Networks.
`[ll] M. B. Pursley, “Performance evaluation for phasecoded
`spread-spectrum multiple-access communication - Part I:
`System analysis,” ZEEE Zhnsactions on Communications,
`vol. 25, pp. 795-799, Aug. 1977.
`[12] Rec. ITU-R M.1225, Guidelines for Evaluation of Radio
`Zhnsmission Technologies for Z q - 2 0 0 0 . 1997.
`[13] S. Lin and D. J. Costello, Error Control Coding: Fbnda-
`mentals and Applrcations, Englewood Cliffs, NJ: Prentice-
`Hall, 1983.
`
`Fig. 3. Variation in h c p with Ne (for different t)).
`
`Authorized licensed use limited to: Klarquist Sparkman LLP. Downloaded on May 05,2020 at 18:37:19 UTC from IEEE Xplore. Restrictions apply.
`
`

`

`1 7.2
`
`m
`m
`
`m
`
`m
`
`JP
`
`d40
`
`50
`
`W
`
`IO
`
`0
`
`Fig. 4:. Variation in PTCP with N, (for different G ) .
`
`Fig. 7. Variation in PTCP with Nc (for channel model-B
`with different RmaZ).
`
`Fig. 5. Variation in PTCP with N, (for channel model-B
`with different Eh f No).
`
`Fig. 8. Variation in PTCP with NE (for channel model-B
`with different W ) .
`
`Fig. 6. Variation in ,&cp with N, (for channel model-B
`with different t ) .
`
`Fig. 9. Variation in PTCP with Ne (for channel model-B
`with different Dwd).
`
`Authorized licensed use limited to: Klarquist Sparkman LLP. Downloaded on May 05,2020 at 18:37:19 UTC from IEEE Xplore. Restrictions apply.
`
`