Siemens Mobile Phones, Haidenauplatz 1, München, Germany

Transcription

1 +<%5,'$54$1'$'$3,9(2'8/$,21$1'&2',1*6&+((6)25 +,*+63((''2:1/,1.3$&.($&&(66 D' O- HFHO%HDG5DDI iemens Mobile Phones, Haidenauplatz 1, 8166 München, Germany $DF - A major evolution of the MT standard is the so-called High peed Downlink Packet Access (HDPA) mode, which provides peak data rates of 1.8 Mbps. ey enabling technologies include fast scheduling, adaptive modulation and coding (AMC), as well as hybrid automatic repeat request (HA). This paper gives a general overview of the current HDPA standard. pecial focus is put on the implementation and performance of the AMC and HA functions, as well as on their interworking. The performance of different HA schemes are compared for P and 16-AM and varying code rates. ubsequently link-level throughput is simulated. It is shown that the throughput gains offered by advanced HA schemes depend on the performance of the AMC function. In particular, the gain offered by incremental redundancy techniques increases with increasing errors in selecting the appropriate modulation and code rate..h\zg - HDPA, MT, hybrid A, adaptive modulation and coding, incremental redundancy. I. ITODCTIO A major innovation within the recent release of the MT standard is the High peed Downlink Packet Access (HDPA) channel. It provides packet-oriented downlink connections for streaming, interactive and background services at high data rates. HDPA is designed to reuse as much of the existing MT functionality as possible and intended primarily for urban/indoor scenarios and low to medium user speed. The use of multiple codes in conjunction with adaptive modulation and coding (AMC) allows peak data rates of 1.8 Mbps. Additionally, fast scheduling and fast hybrid automatic repeat request (HA) are used 1]. HA combines forward error coding using the MT turbo code with a 1-channel top-and-ait protocol for automatic repeat request. Dependent on the network resources, the code rate and the terminal capabilities either Chase Combining, Partial or Full Incremental edundancy (I) are used in retransmissions. A brief description of general physical layer aspects of HDPA is given in section II together with a detailed discussion of the AMC and HA functions. Frame error rate and link-level throughput of different HA schemes are shown in section III for perfect and imperfect AMC. Concluding remarks follow in section I. II. TH MT HIGH PD DOI PACT A CHA The HDPA data channel is basically an enhancement of the existing MT downlink shared channel 2]. HDPA allows to code multiplex different users on up to 15 codes with spreading factor of 16. The primary multiple access however is in the time domain, where different users can be scheduled every transmission time interval (TTI), which corresponds to 3 MT slots, i.e., 2 ms. ote, that also the number of codes allocated to one user can change from TTI to TTI. An example for the use of code and time domain resources by HDPA is illustrated in Fig. 1. In this example a maximum of 1 FPD = 9 codes are allocated to HDPA. codes 1 FPD recipient of data: TTI: Transmission Time Interval user 1 user 2 user 3 time Fig. 1: Time and code domain resources of HDPA $$GDYHGXODDG&G Depending on the system load and channel conditions, the base station adapts modulation and code rate for each user. HDPA uses P and 16-AM. The transport block size can be adapted with fine granularity, resulting in code rates that range from 1/4 to 1. Five modulation and coding scheme (MC) levels are chosen out of this large set for simulation purposes, see Table 1. The different code rates are obtained from the 5 = 1/3 mother code of the turbo encoder by puncturing or repetition. The MC level may change every TTI. It is determined by the base station based on feedback information from the user terminal, which is send with a periodicity ranging from one to 8 TTIs /2/$1. 22 I PIMC 22

2 The user terminal measures the downlink channel quality from the common pilot channel and decides according to its specific receiver performance which MC level it would be able to receive in the current channel conditions. The feedback information given to the base station is a request for a specific transport format, which contains amongst other information the MC level. The base station decides on the MC level of the next transmission based on this request, but also based on further information, like network configuration and status, load and available resources 3]. Therefore the MC level actually used in the next transmission might deviate from the level requested by the user. Table 1: eference Modulation and Coding cheme &6 OHYHO PGXOD FGHDH 5 IDH HFGH 3/4 2 kbps 16-AM 1/2 48 kbps 3/4 36 kbps P 1/2 24 kbps 1/4 12 kbps %+\G$54 If a HDPA packet cannot be decoded retransmissions are sent according to a 1-Channel top-and-ait (A) Protocol, which has been chosen since it provides a good tradeoff between achievable throughput, memory requirements and complexity 4, 5]. In a single A protocol, the transmitter waits after each transmission until an acknowledgement from the receiver is received. In case of positive acknowledgement a new packet is sent, a retransmission is sent otherwise. The 1Channel A protocol makes use of the waiting time by transmitting another A channel in the meantime. In case of retransmissions, the multiple transmissions of one packet are combined by adding the corresponding soft values at bit level. Different types of HA combining exist, and in general, a trade-off between implementation effort and buffer requirements on the one hand and decoding gain on the other hand can be achieved. Chase Combining 6] has been proposed, since it requires minimum complexity and buffer. Initial transmission and retransmission are identical. Chase Combining offers time diversity and soft combining gain (i.e., the energy accumulation effect for each bit). Additional decoding gain is achieved if the retransmissions consist partly or entirely of new parity bits. These techniques are called Partial Incremental edundancy (I) and Full I, respectively. I reduces the effective code rate with each retransmission. Partial I maintains self-decodability in retransmissions, since the systematic bits are always repeated, while Full I retransmissions consist only of parity bits and are not self-decodable. Full I offers maximum decoding gain at the expense of highest soft combining buffer requirements. In summary, apart from running the A protocols, the HA functionality must create the appropriate redundancy versions for retransmissions (i.e., choose different sets of coded bits) and prevent terminal buffer overflow. For HDPA the HA block is combined with the rate matching function, which adapts number of coded bits (with 5 1/3) to the number of available channel bits. A twostage rate matching has been agreed, see Fig. 2 3]. ystematic bits Parity 1 bits Parity 2 bits First ate Matching M P1_1 M P2_1 irtual I Buffer 1 \ 1 1 econd ate Matching M M P1_2 M P2_2 Parameters s and r Fig. 2: The HDPA HA functionality 1 \ In the first rate matching stage the number of coded bits is adapted to the available buffer in the user terminal, i.e., regardless of the subsequent processing it is ensured that all bits can be stored in the terminal. This is achieved by puncturing both parity bit streams of the turbo encoder. ote, that without buffer limitations the first rate matching stage will be transparent. The second rate matching stage further adapts the number of bits to the number of available channel bits per TTI by puncturing or repetition. It also implements different types of HA, depending on the redundancy version parameters and : The parameter distinguishes self-decodable transmissions ( 1) from non selfdecodable ones ( ), whereas allows to select different set of bits to be punctured or repeated. uppose the initial transmission used the parameter {} = {1, }. Then Chase Combining, Partial I and Full I can be achieved using the parameter settings of Table 2 for retransmissions. Table 2: Implementation of different HA schemes +$54FHPH DDPHH DDPHH &DH&P 1 3DDO,5 1 1,..., PD )XOO,5,..., PD ote, that Chase Combining might be required, if the number of channel bits is greater or equal than the available soft combining buffer in the terminal. This implementation provides high flexibility at little additional complexity, since all major function blocks are inherited from already existing MT features. 1 1

3 The simplest solution, however, would be to implement only Chase Combining, where only one rate matching stage would be required. Therefore the performance gain offered by the two-stage rate matching is investigated. III. T AD DICIO The following results compare Chase Combining (labelled as in the figures) and the HA implementation detailed above, which enables the use of I. Fig. 4 to Fig. 8 show the results for all five MC levels (cf. Table 1). The curves show the frame error rate (F) after one, two, three and four transmissions. The curves for different number of transmissions are distinguished by different line types according to Fig. 3, whereas different colours and markers are used for Chase Combining (blue, "x" marker) and Itype HA (red, "o" marker). 1 transm. 2 transm. 3 transm. 4 transm. Fig. 3: egend for subsequent plots One frame corresponds to 48 symbols. The abscissa,, F is the ratio of the total power spectral density of the downlink to one user terminal to the power spectral density of the band limited noise and interference.,, F is defined as, ( 5 1, =, (1) 1 5 F F G where ( 1 is the bit energy to spectral noise density, 5 is the bit rate, and 5 F is the chip rate. The factor G is fraction of radiated power devoted to the HDPA data channel. An total overhead of 2% is assumed (e.g., for pilot channels and signalling) leading to G =.8. Idealised AG simulations where chosen to assess the performance of the scheme itself by eliminating all other factors (like the effect of varying,, F values for different transmissions in fading simulations, or scheduling effects). Thus an isolated impact analysis of the HA scheme performance can be performed. Fig. 4 compares Chase Combining to I for P and code rate 5 1/4. ince the code rate is lower than the MT turbo mother code (5 1/3), repetition is used. Thus the additional gain from I techniques is small, since all parity bits have already been sent once in the first transmission and no additional decoding gain is available. The small advantages (around.2 db) of I result from the fact that the parameter allows to select different bits for repetition in each transmission, thus leading to more homogeneous distribution of the cumulative bit energy over the whole frame. Fig. 5 shows the same comparison for P, 5 = 1/2. ow the code rate is greater than the mother code rate and true I decoding gain is achieved. This gain increases from.9 db after two transmissions to 1.3 db after four transmissions. For P, 5 = 3/4 the performance gains are even more distinct (cf. Fig. 6). Here, I outperforms Chase Combining by 2 db to 2.2 db. Fig. and 8 show the results for 16- AM, 5 = 1/2 and 5 = 3/4, respectively. Again, significant performance gain can be seen for the I scheme (about.8 db for 5 = 1/2 and from 2.3 db to 2. db for 5 = 3/4). F F F F I Fig. 4: F in AG for P, 5 1/4 I Fig. 5: F in AG for P, 5 1/2 I Fig. 6: F in AG for P, 5 3/4 I Fig. : F in AG for 16-AM, 5 1/2

4 F I Fig. 8: F in AG for16-am, 5 3/4 Based on the above results link-level throughput per code has been calculated for both HA schemes. It has been assumed that sufficient soft combining buffer is available in all cases. ince a maximum of four transmissions per packet are considered, the valid range for further evaluation is limited to,, F -1 db, -1 db]. Fig. 9 shows the achievable throughput per code for the five MC levels using Chase Combining, Fig. 1 the same graph for the I scheme X4 X X4 X3 2 1 MC1 MC2 MC3 MC4 MC ,,FG% Fig. 9: Throughput curves for Chase Combining MC1 MC2 MC3 MC4 MC ,,FG% Fig. 1: Throughput curves for I scheme ach colour corresponds to one particular modulation and code rate. For each curve plateau regions alternate with steep sections. Following one curve from the right to the left hand side, the plateaux correspond to,, F values where exactly one, two, three or four transmissions are required. The steep parts are the regions where the number of required transmissions is changing. Comparing Fig. 9 to Fig. 1 it is evident that I allows to shift these transitions regions to significantly lower,, F values. e now consider a system that sets the MC level switching points such that always the MC level with the highest throughput is chosen, i.e., the hull curves in Fig. 9 and Fig. 1 result. The corresponding MC level switching points are shown as circles in Fig. 9 and Fig. 1. If the same throughput would be offered by several MC, the lowest (most robust) MC is chosen and the switching point is set to the highest possible,, F value. The resulting hull curves are compared in Fig X4 X3 2 1 I ,,FG% Fig. 11: Throughput curves for I cheme, σ = db Due to the I gain, MC level 3 and 4 contribute to the hull also after two transmissions in the range of -12 db and -15 db, respectively and increase the mean throughput by around 4 % from 265 kbps to 25 kbps. A maximum throughput gain of 5% is obtained at,, F = -12 db. In a real system the downlink quality indicator received at the base station might not represent the actual state at the user terminal due to various reasons, such as measurement inaccuracy, signalling errors, limited signalling bandwidth and reporting delay. Additionally the base station scheduler may deviate from the requested MC level due to reasons not transparent to the user. A straightforward way to account for this imperfect mapping between used MC level and true,, F at the user terminal is to introduce a random error variable between assumed and true,, F that accounts for all deviation in summary. Due to the multiple causes for this error it might be reasonably large, especially in scenarios with high time variance, e.g., caused by user velocity or bursty interference. Fig. 12 shows the hull curves for a standard deviation of σ= 6 db. Comparing Fig. 11 to Fig. 12, it is evident that the imperfect MC selection process reduces the achievable throughput for a given,, F. The performance gain of I over Chase Combining however increases: I allows to increase the mean throughput from 196 kbps to 21 kbps,

5 i.e., by around 11%. ote, that the,, F ranges where significant I gain is obtained have also changed. hile for σ = db 5% gain is achieved around,, F = -12 db, only 14% gain occurs for σ = 6 db. This is due to the fact that around,, F = -12 db narrow-spaced MC switching points are used for the I scheme (cf. Fig. 1), leading to more frequent selection of a sub-optimal MC level in case of imperfect AMC X 4 X I ,,FG% Fig. 12: Throughput curves for I cheme, σ = 6 db Further results on the relation between MC selection errors and I gain are given in Table 3. ince the operational,, F distribution in a cell depends strongly on the scheduling algorithm used, a base-line evaluation is performed assuming equal distribution. By increasing MC selection error σ to 9 db, the mean throughput decreases by 34% for Chase Combining and by 24% for I. At the same time the mean gain offered by I with respect to Chase Combining increases from 4% to 18%. The,, F values were I provides improvement increases from 55% up to 8%. In general, the benefits of I become more distinct for imperfect adaptive modulation and coding, in particular if the standard deviation of the error reaches the order of magnitude of the MC switching point intervals. Table 3: Impact of MC selection error DGDGGHYDσ G% G% G% G% PHDXX&& PHDXX, PHDD 4% 3% 11% 18% PDD 5% 64% 88% 11% PD % -31% -38% -48% PYHPH 55% 65% 8% 8% GHHD % 35% 13% 13% The link-level performance increase shown above clearly justifies the implementation complexity of the HA functionality based on two-stage rate matching. In an operational system, several additional influences, like the effective,, F distribution of the scheduled users, the impact of fading (e.g., retransmissions at different MC levels), variation of allocated codes per transmission and the choice of the MC switching points will further affect the overall system-level performance 2, ] I. COCIO Physical layer aspects of HDPA, the most recent enhancement of the MT standard, have been discussed. A simple, backward compatible, but versatile implementation of different HA techniques, like Chase Combining, Partial and Full I, has been outlined and evaluated in connection with adaptive modulation and coding. The gain offered by incremental redundancy compared to Chase Combining increases with increasing code rate. Furthermore, the performance improvements obtained from incremental redundancy and adaptive modulation and coding are interdependent. ith increasing probability that the actually used modulation and code rate is not the optimum one, the relative gain achieved by I increases. Thus the advanced HA functionality that implements Chase Combining, Partial I, as well as Full I has been adopted for HDPA. FC 1]. Parkvall,. Dahlman, P. Frenger, P. Beming, and M. Persson, "The volution of CDMA Towards Higher peed Downlink Packet Data Access," 3F,(((9& 6, hodes, Greece, May 21, pp , 2]. Malkamäki, D. Mathew,. Hämäläinen, "Performance of Hybrid A Techniques for CDMA High Data ates," 3F,((( 9& 6, hodes, Greece, May 21, pp , 3] 3GPP TG A, "High peed Downlink Packet Access: Physical ayer Aspects," 6*5 HFFDO 5H, T , 4] 3GPP TG A, "Physical ayer Aspects of High peed Downlink Packet Access (elease 2)," 6* 5HFFDO5H T , 5]. in, D.. Costello, r., and M.. Miller, "Automatic- epeat-equest rror-control chemes,",((( &PPX D vol. 22, no. 12, pp. 5-1, December 1984, 6] D. Chase, "Code Combining A Maximum-ikelihood Decoding Approach for Combining an Arbitrary umber of oisy Packets,",(((D&PPX, vol. 33, no. 5, pp , May 1985, ]. ove, A. Ghosh,. ikides,. alloul, M. Cudak, and B. Classon, "High peed Downlink Packet Access Performance," 3F,((( 9& 6, hodes, Greece, May 21, pp