1936
`
`Adaptive loading strategy for a high speed OFDM-based WLAN
`
`Liesbet Van der Perre, Steven Thoen, Patrick Vandenameele,
`Bert Gyselinckx, Marc Engels
`
`Abstract
`Wireless LANs based on OFDM are limited in their
`performance by the dips in the frequency spectrum of the
`indoor multi-path channel. To simulate perfonnance, we
`first derive a simplified channel model from a ray-tracing
`analysis of the indoor channel. The performance of
`OFDM-based WLANs can be augmented by adaptive
`loading of the subcarriers, which has shown to be very
`valuable for DMT modems on wired media (for example
`in ADSL). In this paper, we investigate the application of
`adaptive loading to a 100 Mbit/s OFDM-based wireless
`LAN. From the different adaptive loading algorithms, the
`Fischer et al. algorithm is selected because it offers
`computational advantages for similar performance. This
`algorithm is then integrated in a complete system by
`implementing a channel and noise power estimation.
`Simulations show that the proposed adaptive loading
`strategy improves the system performance considerably
`(6 dB gain for a BER=lO.*). Moreover, the simulations
`confirm that the simplified channel model is a reliable
`description of the indoor channel.
`
`1 Introduction
`Wireless LANs show the clear advantages over their
`wired counterparts that they reduce the initial wiring
`effort and allow much more flexibility. However, to
`become really competitive with traditional LANs, they
`require higher data rates than what they currently offer,
`and thus a very efficient usage of the available capacity
`is needed. For the topology of such a WLAN we
`envision a star-configuration consisting of relatively
`large cells, e.g. a comdor in an office environment. In
`order to limit propagation losses for such a topology, the
`target frequency bands should not be very high (probably
`in the 2-3 GHz or 5-6 GHz hands). As a modulation
`technique for the WLAN, we chose OFDM (Orthogonal
`Frequency Division Multiplexing), because of its ability
`to mitigate IS1 by introducing a cyclic prefix, its
`promising multiple-access possibilities and its flexibility.
`A prototype OF’DM-modem ASIC is under development
`at IMEC at the moment 111. The simulations presented in
`this paper are based on the concept of this OFDM-
`modem. Applying adaptive loading on top of OFDM-
`modulation is one means to increase the capacity usage
`of the WLAN and mitigate the effects of Rayleigh
`fading. This paper presents such a system and analyzes
`its performance.
`The organization of this paper is as follows. The second
`section proposes an indoor channel model following a
`
`0-7803-4984-9/98/$10.00 01998 IEEE.
`
`geographical approach based on ray-tracing. Therefrom,
`we derive time and frequency domain responses for a
`typical office environment. A corresponding simplified
`channel model is deduced for simulation efficiency.
`Section 3 describes several adaptive loading algorithms,
`which are evaluated in terms of their performance,
`implementation complexity and practical implications.
`The Fisher et al. algorithm [6] is selected over the
`alternatives. Section 4 describes the integration of the
`selected algorithm
`in
`a
`complete
`system, by
`implementing a practical channel and noise power
`estimation which provide the input for the adaptive
`loading. Simulation results are given for the different
`the resulting BER-
`in the algorithm, and
`steps
`is compared with the results without
`performance
`adaptive loading. Finally, section 5 summarizes the most
`important conclusions and gives a view on the. future
`work.
`
`2 Channel model
`It is of primordial importance to first dispose of a reliable
`model for the wireless indoor channel, in order to he able
`to give a realistic estimate of the system’s performance,
`on
`simulation
`results. Especially when
`based
`investigating adaptive loading, the results are heavily
`dependent on the specific channel characteristics, and a
`reliable model for the channel response is a necessity.
`The following aspects are important for the indoor
`propagation modeling approach:
`on the one hand, the indoor channel should be
`modeled as close
`the actual propagation
`to
`phenomena as possible,
`on the other hand, it is not necessary to obtain an
`exact channel response which takes all geometrical
`details and propagation effects into account.
`Taking into account the above considerations, a ray-
`tracing technique on a ‘higher’ level (only modeling of
`the walls and not of different obstacles) was considered
`to he the best approach for a sufficiently accurate (yet
`time-consuming) model of the channel
`A typical cumdor in an office environment is taken as
`the geomemcal input for the ray-tracing algorithm. This
`input corresponds with the topology we propose for a
`W A N . Figure 1 shows a typical output of the
`implemented algorithm and displays the different
`propagating waves. We take multi-path components into
`account up to a certain threshold for the power loss
`(which was 30dB lower than the power of the first ray for
`the presented example). The transmitter (T) is located in
`
`APPLE 1032
`
`1
`
`

`

`1937
`
`the middle of the corridor, while the receiver (R) is
`situated in one of the office rooms. Solid lines display
`the walls, while all dotted lines represent propagation
`paths.
`
`the transmission in these frequency bands will he
`degraded dramatically.
`
`d
`
`.,ID
`
`-,om
`
`Figure 1: Typical ray-tracing output
`Tbe above picture shows that a lot of reflections take
`place in the indoor environment and consequently a large
`number of multi-path components arrive in the receiver.
`The channel impulse response in the time domain will
`clearly consist of a number of discrete pulses. The
`resulting power delay profile is given in figure 2.
`
`- 0
`
`smwmy
`
`,a**
`
`,m
`
`I-
`
`m
`Figure 3: Frequency response
`The ray-tracing results also confvm that the channel
`responses for receivers (or transmitters) separated over
`distances of at least U2, are completely uncorrelated. U2
`is 6 cm for a frequencie of 2.4 GHz which means that
`this constraint is met in all practical multi-user situations.
`This result is important in the view of applying adaptive
`loading in a multiple user situation, as we plan to do in a
`next step.
`From the analysis of the indoor propagation channel, a
`simplified channel model was deduced which can
`generate equivalent power delay profiles (and thus also
`channel transfer functions) without going through the
`rather extensive ray-tracing step. The time domain
`response h(t) is approximated by a sum of impulses with
`random phases and with amplitudes that are on the
`average exponentially decaying. The exact mathematical
`expression is given by:
`
`Bigure 2: Power delay profile
`Simulation results show that the delay spread in this
`environment varies between 10 and 40 ns. The delay
`differences between subsequent multi-path components
`are in the order of 0.1 ns. From the simulation r e d < the
`channel transfer function in the frequency domain was
`calculated by simply taking the FFT transform. The
`result is presented in figure 3. This frequency response
`shows a very good agreement with measurement data in a
`similar environment [2]. It is clear that the indoor
`propagation channel is frequency selective, with a
`coherence bandwidth in the order of 5 to 25 MHz. From
`this observation, we can conclude that using a flat fading
`model is not reliable for the envisaged hit rate of 100
`Mbps, and it should be possible to improve the system’s
`performance by exploiting frequency diversity. Dips in
`the spectrum of up to 30 dB are perceived, meaning that
`
`A, = exp(-a.i.Az).(l+ pi)
`with,
`pi = v.si andsi = uniformly distributed[-1 ... 11
`$I = uniformly distributed [O ... 2 . ~ 1
`
`(2)
`
`where a = I/aRMs with aRMS the delay spread on the
`channel, v is a parameter for the variation on the
`amplitudes of the multi-path components, AT gives the
`resolution between different multi-path components, and
`T m = M . A ~ is taken to he 4 times the delay spread.
`Specific values and typical ranges for the parameters of
`the model are derived from the geometrical analysis,
`giving the following results:
`urns - 10 ... 40 ns
`v - O...l
`AT - 0 . 1 . 3 ns
`Channel responses generated by this simplified model,
`show a very good correspondence with the results from
`the ray-tracing algorithm, as will be shown further in
`section 4.
`3 Adaptive loading algorithms
`
`(3)
`(4)
`(5)
`
`2
`
`

`

`1938
`
`Generally speaking, about 1% of the carriers have an
`attenuation that is 20dB below the average and about
`0.1% of the caniers aftennate 30 dB below the average.
`Since transmission over this carriers is severely degraded
`(BER = 0.5 for the 30dB dips at moderate SNR), these
`caniers determine
`the performance, which can be
`achieved with OFDM. Therefore, adaptively loading the
`carriers seems an interesting means for increasing the
`capacity usage on the channel. By measuring the channel
`and identifying dips and peaks in the spectrum, the
`transmission can be optimized. In particular, adaptive
`loading adjusts the number of bits assigned to a specific
`carrier, to the SNR of this carrier. Capacity can be
`achieved by distributing the energy according to the
`waterfiiling distribution [9]. However, this distribution is
`difficult to compute and assumes infinite granularity in
`constellation size. Different practical (but suh-optimal)
`adaptive loading algorithms have been proposed in the
`ADSL-context.
`The Hughes-Hartogs [3,4] algorithm allocates bits one
`by one, each time selecting that carrier for which
`transmission of an additional bit can be done with the
`smallest additional transmit power for a requested BER.
`Unfortunately, this strategy is too complex for a practical
`implementation in the envisaged high speed WLAN,
`since it asks for comprehensive sorting. The algorithmic
`complexity for Hughes-Hastogs is proportional to O(R,d
`x NJ, with Rmd the total number of bits to be assigned
`and N, the numher of carriers.
`A second loading algorithm, proposed by Chow et al.
`[5], focuses on a rate distribution according to the
`capacity of the subcbannels. The hits are assigned as
`follows:
`
`where Si is the signal power, Ni the equivalent noise
`power and R, the transmission rate of carrier i. The
`algorithm is based on the ‘gap approximation’. The noise
`margin yme is determined iteratively to make sure that
`the total bitrate equals the targeted hitrate. The SNR gap
`r is a constant which estimates the difference between
`channel capacity and the actual capacity usage by the
`transmission scheme. The complete transmission system
`is characterized by this single parameter. This algorithm
`does not rely on comprehensive sorting and is therefore
`computationally less expensive than the Hughes-Hartogs
`algorithm. The worst case algorithmic complexity here is
`only proportional to 0(1 x N, + ZNJ, where I equals the
`maximum number of iteration steps which is allowed.
`While the algorithm of Chow et al. mes to maximize the
`usage that is made of the channel capacity, in onr
`practical case, we try to transmit at a given data rate
`(determined by
`the application) with a minimal
`achievable bit error rate. This reasoning forms the basis
`of the algorithm that was proposed by Fischer et al [6].
`
`Rate and power are distributed over the different carriers
`in order to minimize the BER at a constant total bitrate
`RT and transmit power. The bits in this algorithm are
`assigned according to the formula shown in (8).
`
`This bit assignment is done iteratively since R, as
`calculated by (8) may become negative and those carriers
`have to be excluded. D equals the number of carriers that
`are still included in a specific iteration. The algorithm
`iterates until the target rate is reached and all rates of the
`remaining channels are positive. The worst case
`algorithmic complexity is comparable to the previous
`algorithm, but the computations required for the Fischer
`strategy are a lot simpler. Computationally, it is the least
`expensive, making it especially suitable for high-speed
`data transmission. Furthermore, the performance of this
`algorithm is shown in [6] to be superior 01 at least
`equivalent to the other ones in terms of achievable BER.
`The Fiscber et al. algorithm was therefore chosen as
`being the most appropriate for the high speed WLAN
`case.
`
`4 An adaptive loading system for WLAN
`In order to investigate the effect of adaptive loading on
`the performance of a high speed WAN, we simulated
`the algorithm of Fiscber et al. for both the ray-tracing
`and the simplified channel model. All simulations are
`based on an OFDM modem, which uses 256 carriers and
`a 8 bit cyclic prefix for coping with IS1 and ICI.
`The first step that is required in any loadmg technique, is
`an estimation of the noise powers and the attenuations
`on all carriers. We accomplish this by sending a BPSK-
`modulated PN-sequence 4 times over the channel. Both
`the attenuation and the noise power estimation are based
`on these reference symbols. Let r(i) be the signal
`received on the i-th Carrier. Formula (9) gives the output
`of the OFDM demodulator (i.e. the FFT), provided the
`length of the guard interval was chosen to he longer than
`the delay spread of the channel:
`r(i)=PN(i). H(i)+N(i)
`19)
`where H(i) stands for the complex channel attenuation
`factor (i.e. the FFT of the channel impulse response h(t))
`and N(i) stands for the complex additive white Gaussian
`noise on carrier i with noise power No@) on carrier i. We
`take the average of the channel estimations obtained for
`each PN-sequence. This reduces the noise power of the
`estimation for each carrier by a factor 4. We deduce the
`channel attenuation estimation
`from:
`B(i) =avg{r(i)}.~~(i)~
`(10)
`= H(i)+avg{N(i).PN(i)} (11)
`Here, avgI.1 means taking the average over the 4 PN-
`sequences that were sent. Figure 2 shows the spectrum of
`
`3
`
`

`

`1939
`
`a channel generated by the simplified model of section 2,
`together with the unfiltered channel estimation.
`I
`
`i
`
`I
`I
`
`,ol
`
`C a m S E
`
`,*
`
`jmo
`
`1
`
`1o
`
`Figure 4: Unfiltered channel estimation
`To further diminish the estimation noise power and
`obtain a better estimation H I , we apply a lowpass filter
`to the channel estimation. Following a property of the
`FFT for finite time signals, the spectnun of H(i) is equal
`to h(-t) + n(-t), which only has signal components up to
`Tm. The rest of the spectnun of H(i) (up to NT) only
`contains noise contributions and can he fitered away.
`This decreases, in the case of an ideal filter, the noise
`power of the carrier attenuation estimation by a factor
`TmiNT. However, a simple lowpass filter, based on the
`mean filter, was used for reasons of complexity. Figure 5
`shows a typical result for the filtered channel estimation.
`
`1
`>o~.!
`
`EbNo
`-
`Figure 6:Performance of the estimators
`The noise power estimation N is obtained in a similar
`manner
`
`R(i)=uvg{Ir(i) -ii, (i). PN(~)I* 1 , (11)
`Using these noise power and attenuation estimations, the
`bits are assigned in fmt instance with infinite granularity
`to the carriers using the Fischer et al. algorithm. In the
`next step, the assigned number of hits is quantized to an
`even integer. Only even integers are allowed in order to
`balance the BER on the I and Q components. The third
`part of the algorithm adapts the hitassignment in order to
`make sure that Rtow equals the targeted bitrate. Finally,
`the power of each carrier is adjusted so that all carriers
`yield the same BER. The bit assignment for the channel
`estimation in figure 5 is shown in figure 6.
`
`lol
`
`camem
`
`mm I
`
`Figure 5: Filtered channel estimation
`Figure 6 shows the performance of the different steps in
`the estimation in terms of MSE. The final (filtered and
`averaged) attenuation estimation achieves a MSE of 10.'
`for 6dB of Eh/No. A further reduction of MSE is
`possible by using better filtering at the expense of
`complexity. Alternative channel estimation strategies are
`also possible [71[81.
`
`Figure 6: Bit assignment
`This figure indicates that the dips in the specmm are not
`carrying any hits while more hits are transmitted on the
`carriers with low attenuation. Furthermore, the order of
`the constellation on the good carriers is low (QPSK or
`16-QAM) compared to the order in ADSL, since in
`general the Eb/No is relatively low for the indoor
`channel.
`
`4
`
`

`

`1940
`
`response and noise estimation phase, and uses the
`Fischer approach [6] to assign the bits to the carriers.
`Simulations show a 6 dB improvement for a BER of 10.'
`to the normal OF'DM-modem. In other words, adaptive
`loading is certainly a good method to improve the
`capacity usage in WLANs.
`In our ongoing work, we are analyzing a multi-user
`situation with Orthogonal Frequency Division Multiple
`Access (OFDMA). The assignment of the carriers to the
`different users adds an extra degree of freedom. The
`optimum solution is based on multi-user waterfiling.
`However, complexity
`issues may necessitate
`the
`development of heuristics for this optimization problem.
`Preliminary results confirm the expectation that in a
`multi-user situation, we can benefit from the fact that the
`channels of the different users are uncorrelated. The
`advantages of adaptively assigning bits and carriers will
`increase with the numher of users. Thus, the usage of the
`total network capacity can be greatly improved in
`comparison to the single user case.
`
`Figure 7: Performance for ray-tracing model
`
`E: Io- m
`
`m 0
`0
`
`2
`
`4
`
`6
`
`
`
`EWNbO (dBf
`Figure 8: Performance for simulation model
`Figure 7 compares the BER for adaptive loading and for
`plain QPSK on all the carriers for the ray-tracing channel
`model when the data rate is lOOMbps and a cyclic prefix
`of 8 hits is used. A 6 dB gain is observed for BER = 10 '.
`Figure 8 shows the BER for the simplified simulahon
`channel model. As can be concluded from these two
`figures, the difference hetween the simplified channel
`model and the ray-tracing model is minimal. From the
`figures, it is clear that a substantial advantage can he
`aclneved by adapting the bit allocation to the spectnun of
`the fadmg channel. Tbe main benefit arises from the fact
`that the carriers which form an errorfloor in QPSK, are
`now avoided. To accomplish tlns, more bits have to be
`sent on the peaks hut since even on the good carriers the
`SNR is relauvely low, this does not result in high order
`constellations as in ADSL.
`5 Conclusions
`The simphfied stanshcal model in this study, which was
`denved from a ray-tracing model, forms an efficient
`representanon of the indoor channel. It was used to
`simulate the performance of the adaptive loading strategy
`that we propose for a high speed OFDM-based WLAN.
`This adaptive loading strategy comprises of a channel
`
`2.
`
`6 References
`1. W. Eberle, L. Van der Perre, B. Gyselinch, M. Engels, I.
`Bolsens. Design Aspects of an OFDM Based Wireless
`LAN with Respect to ASIC Integration. Proc. of the 2-nd
`Braunschweig,
`Germany,
`OFDM-Frachgesprach,
`September1 997.
`G.J.M. Janssen, P.A. Stigter, and R. Prasad. Wideband
`Indoor Channel Measurements and BER Analysis of
`Frequency Selective Multi-path Channels at 2.4,4.75, and
`11.5 GHz. IEEE Trans. on Communications, Vol. 44, No
`10, Oct 1996, pp. 1272-1281.
`3. Hughes Hartogs. Ensemble Modem Structure for
`Imperfect Transmission Media. U S Patents Nos.
`4,679,227 (July 1987), 4,731,816 (Mmh 1988) and
`4,833,706 (May 1989).
`J.A.C. Bingham. Multicarrier Modulation for Data
`Transmission : An Idea Whose Time has Come. IEEE
`Communications Magazine, 28(5):5-14, May 1990.
`P.S. Chow, J.M. Cioffi, and J.C. Bingham. A Practical
`Discrete Multitone Transceiver Loading Algorithm for
`Data Transmission over Spectrally Shaped Channels.
`IEEE Transactions on Communications, 43(2):773-775.
`J.B. Huber. A New Loading
`6. R.F.H. Fischer, and
`Algorithm for Discrete Multitone Transmission. IEEE
`Proc. GLOBECOM '96, London, England, November
`1996, pp, 724-728.
`J-J. van de Beek, 0. Edfors, M. Sandell, S.K. Wilson,
`P.O. Borjesson. On Channel Estimation in OFDM
`Systems IEEE Proc.of VTC '95, Chicago, USA, July
`1995, pp.815-819.
`J-J. van de Beck, 0. Edfors, M. Sandell, S.K. Wilson,
`P.O. Borjesson. OFDM Channel Estimation by Singular
`Value Decomposition. Researcb Report TULEA 1996:18,
`Div. Of Signal Processing, Lulea University of
`Technology, September 1996.
`9. R.G. Gallager.
`Information Theory and Reliable
`Communication. Wiley, New York, NY, 1968.
`
`4.
`
`5.
`
`7.
`
`8.
`
`5
`
`