A Concept for Hybrid Random/Dynamic Radio Resource Management

Transcription

1 A Concept for Hybrid Random/Dynamic Radio Resource Management Miguel Berg Radio Communication Systems Lab. Dept. of Signals, Sensors and Systems Royal Institute of Technology SE Stockholm, Sweden ABSTRACT Third generation mobile communication systems will allow many different services with varying demands on data rate and delay. For those systems, the fixed radio resource management (RRM) techniques used in today s low rate systems will not be appropriate to ensure efficient utilization of the radio spectrum. Dynamic RRM is necessary in order to cope with the large variations in data rate. In this paper we study the bunch concept which is a hybrid between random and dynamic RRM. Within a small synchronized cluster (bunch) we use dynamic radio resource allocation combined with SIRbased power control and between the bunches we use time/frequency hopping to average the interference. We propose an improved RRM algorithm and evaluate the performance of the bunch concept for a Manhattan scenario with mixed traffic. The performance is improved while the computational complexity is reduced compared to our previously published results for the bunch concept. We also show that the performance of a single bunch system is close to the theoretical upper bound. I. INTRODUCTION The interest for wireless multimedia services is growing rapidly. Designers of wireless systems will face difficulties such as highly varying user concentrations and vastly different demands on quality of service. Some services will require very high data rates while others will have very strict delay limits. The current second generation mobile communication systems are mainly intended for speech and low rate data and typically use Fixed Channel Allocation (FCA) and simple constant received power control strategies which are not suitable for large data rate and population density variations. FCA schemes are not feasible when one user temporarily needs a large part of the total system bandwidth since they split the bandwidth in advance between base stations [1]. Random channel allocation (spread spectrum) systems can solve some of the problems and give good performance but even better performance is possible with dynamic channel allocation (DCA) altough the signaling load can be a problem in wide area systems. The bunch concept is a hybrid between dynamic and random RRM. It uses fast dynamic RRM in a locally centralized and burst-synchronized bunch of base stations; and random channel allocation between the bunches where synchronization and high signaling load prohibits the use of dynamic RRM. Our version of the bunch concept was presented in detail in [2] and extends the work in [3-5], in the sense that a novel global (inter-bunch) solution is offered. We also proposed a new RRM architecture using Signal-to-Interference-Ratio (SIR) based channel allocation and transmitter power control. In this paper we have improved the RRM algorithms in [2] to obtain higher performance when the power control has a limited dynamic range as well as to lower the computational complexity. Furthermore, we have studied the performance with mixed traffic. II. THE BUNCH CONCEPT Our wireless system as presented in [2] is based on the concepts discussed in [3-5] where a limited number of Remote Antenna Units (s) are connected to a hub or Central Unit (CU) as in Fig. 1. This is similar to GSM where a number of Base Transceiver Stations (BTSs) are connected to a Base Station Controller (BSC) but we want to centralize further by moving most of the intelligence and signal processing into the CU. Fig. 1 The bunch concept. Bunch of s CU The s in the bunch are connected to the CU with high speed communication links which makes communication cheap and enables burst synchronization within the bunch. The s could in some cases even be fed an RF signal directly through optical fibers. All information like transmitter powers, receiver noise factors, channels used, channel measurements, etc. is known by the CU. A bunch is ideal for hotspots like central parts of a city or a building floor. Preferably we would like the bunch to cover the whole hot-spot and thus most of the interference would be generated within the bunch. Network planning could aid in this task but in some cases we will need several bunches to cover an area effectively. That means that we need a method to combat the inter-bunch in-

2 terference. Earlier studies [3-5] assumed that dedicated frequency bands where used for different bunches. Segregation strategies could be used but are not considered here. We have chosen to use interference averaging [6] by time and frequency hopping (TH/FH) to handle the inter-bunch interference. The hopping sequence is the same for all users within a bunch but different for different bunches. This preserves the intra-bunch interference and averages the inter-bunch interference. A macro-cell with slow FH can thus be seen as a special case of a bunch with only one. III. A RADIO RESOURCE MANAGEMENT ARCHITECTURE FOR THE BUNCH CONCEPT A. Overview The bunch concept is very flexible and allows several different approaches for radio resource management. In [3] an interference-matrix based approach with a maximum-packing algorithm is used for resource allocation. Only two power levels are considered. In [4,5] channels are selected based on measurements of cochannel interference but power control is not discussed. Our version of the bunch concept as presented in [2] uses a SIR-based dynamic channel allocation and power control scheme. Two different multiple access modes have been proposed by the European ACTS project FRAMES for UMTS (Universal Mobile Telecommunication System). We mainly consider FRAMES multiple access mode 1 (FMA1) which uses wideband TDMA with and without spreading [7]. Different sized slots are possible and thus the number of slots within one TDMA frame (4.615 ms) can vary between 8 and 64. The smallest possible resource that can be assigned within a TDMA frame is here denoted a Resource Unit (RU). For FMA1 without spreading this is one slot in the frequency/time matrix. We use a distributed slow power control scheme (e.g. DCPC [8]) to ensure that all assigned RUs achieve their SIRtarget, γ T. Admission of new RUs is assisted by interactive admission power control schemes such as Active Link Protection (ALP [9]) or Soft-and-Safe (SAS [10]). These schemes limit the power for new users until the admission is considered safe. For resource allocation we have the following four sub-algorithms: Selection finds the most suitable, e.g. lowest path loss. RU Selection chooses the resources to try according to some heuristic or on a random basis. Feasibility Check calculates if all cochannel users within the bunch can achieve their γ T if the new user is admitted. Interactive Admission and Initial Power ensures a smooth admission of new users without disturbing the existing links. RU selection is done by picking one RU at random and then testing it with the feasibility check. The advantage with random selection is that it gives almost as high capacity as the best known methods [11] while still keeping the computational complexity down. B. Inter-Bunch RRM As described earlier, Inter-Bunch interference is handled by TH/FH in combination with the power control (DCPC). After each frame, the time and frequency matrix for a bunch is scrambled according to a (pseudo-)random sequence. Every bunch has its own sequence and the long term correlation between sequences is low or zero. C. Intra-Bunch RRM The intra-bunch RRM is based on knowledge of all transmitter powers and the link gain matrix, G [12] which is constructed from downlink measurements on the beacon channel broadcasted by all s. By proper use of the training sequences in each traffic burst it would also be possible to do uplink gain measurements. For those beacons that can not be reached by a mobile, we can use the average or median value of the gain from that beacon to our cell. The powers and gain values are then used in the feasibility check as follows: In a system with M cochannel users, the SIR for mobile i is given by i ii γ i = M j = 1 j i PG PG j ij + N where G ij is the link gain from mobile i to the used by mobile j, P i is the transmitter power and N i is the thermal noise in receiver i plus any measured inter-bunch interference. The objective of the bunch RRM is that all admitted users should achieve the SIR target, γ T. In our earlier work [2], we solved the matrix equation system resulting from γ i T i (1) γ, i (2) for the power vector P by Gaussian Elimination in order to see if admission was feasible and to get the initial power levels. The problems with this approach is that the computational complexity is rather high, O(M 3 ), and there is no straightforward way of handling minimum power constraints. If some of the calculated powers are below the minimum (P min ) we can not be sure that simply raising them to P min still gives a feasible solution, especially when the PC dynamic range is small, e.g. 20 db as could be the case in the downlink. To cope with the problems mentioned, we present an iterative method to find the power levels. The main idea is to run the distributed power control algorithm (DCPC [8]) as a simulation within the CU. We assume that new users arrive one at a time. Before the new arrival, all users achieve their SIR target. The new user is given a starting power level high enough to overcome the current interference, i.e. M P0 = max( Pmin, ( G0j Pj + N0) γ / G00), (3) j= 1 T

3 and then the iteration starts. After every iteration we bound the transmitter powers between P min and P max. If all the links achieve their γ T within maxiter (about five) steps then the new connection is supported. The full algorithm looks as follows: Set P 0 according to (3) feasible=false if P 0 <P max P i (0)=P i, i for n=1:maxiter, Calculate γ i using equation (1), i Update the power levels P i (n) using DCPC, i if any (P i ==P max, i) then break else if γ i >γ T tol, i then feasible=true, break end for if (feasible==true) then Admit new user else Try another channel or block the user One DCPC update is O(M) but since we run non-interactively, the SIR-values can not be measured and thus the central unit in the bunch has to calculate these values from (1) for each iteration. This results in a co mplexity of O(M 2 ). D. Service Requests Every user requesting service in our system has a resource vector RV={R W R M γ T γ L } where R W is the wanted number of RUs, R M is the minimum acceptable number of RUs, γ T is the SIR target and γ L is the SIR limit for outage. IV. PERFORMANCE EVALUATION A. Models and Parameters We have evaluated the performance of our system with snapshot simulations on a Manhattan grid as shown in Fig. 2. The simulation area is finite i.e. no wraparound is used. In order to investigate the border effects we have studied the difference when performance statistics are collected from only a few s in the center (filled circles in Fig. 2) compared to when we collect statistics from all s (stat=center versus stat=all). Perfect estimates of SIR and link gains are assumed in the simulations. We have not performed full system simulations and thus our results can not be compared directly to other results published by the FRAMES project. The block size is 200 by 200 meters and the street width is 30 meters. The models used are similar or equal to the ones recommended by ETSI for UMTS radio transmission technology selection in [13]. Some of our simulations are extended from the snapshots to short dynamic simulations with TH/FH. In this case we use error correction coding and interleaving with depth I depth. A user with code rate R needs 1/R slots per TDMA frame. We assume perfect erasure decoding which means that we consider a block of I depth bursts as error-free if at least R I depth of those blocks have a SIR larger than γ L and totally destroyed otherwise. North south distance (m) Manhattan scenario, 12 by 12 blocks and 72 s East west distance (m) Fig. 2 A Manhattan scenario, 12 by 12 blocks with 72 s (circles). The filled circles near the center are the s used for collecting statistics in order to avoid border effects. All results are for the downlink since this is the capacity limiting direction. Mobiles are distributed randomly at the street centers with a uniform distribution over the street length. Mobility is not modeled. Base station antennas are assumed to be placed below rooftops. The propagation model is a recursive model where the path loss is calculated as a sum of lineof-sight (LOS) and non-los segments [14]. A dual slope behavior is included with a breakpoint at 300m. The path loss is proportional to x 2 and x 4 before and after the breakpoint respectively (x is the distance between transmitter and receiver). Furthermore we calculate the path loss over the rooftops according to the COST Walfish-Ikegami model: L = log(d+20) where d is the shortest geographical distance. The path loss we use in the simulations is the minimum between the recursive and the COST model. Log-normal shadow fading with zero mean and 10dB standard deviation is included. We have studied three different mobile classes: Class A users with RV={ }, class B with RV={ } and class C with RV={ } When two or more classes are mixed in one simulation we set the number of users in each class so that they generate an equal number of RU requests, i.e. there are 16 class A users for every class B or C user. One of our performance measures is the user assignment failure rate,, which we define as the fraction of arriving users that either do not get their minimum amount of channels, or the channels they get are unusable due to insufficient quality (SIR). The total number of available channels or RUs in our system is C. We define the relative traffic load, ω C, as the fraction of C requested for in each cell. Thus, we have on average ω C C RU requests in every cell. Slots needed for coding are not counted here. Also of interest is the computational complexity which is the number of floating point operations (flops) needed to allocate one RU. Our system consists of B cells and s. Since the number of users in a cell has a Binomial distribution and all users are active, we can deduce a lower bound, U k C U * = ( ) k k B B k= C+ 1 k U k, (4)

4 for the assignment failure rate based on an interference free system. U is the number of users in the whole system (U = ω C C B). The bound is valid as long as we select based on lowest path loss (or shortest distance). For the bunch DCA (BDCA) system, all channels are available in every cell but for the FCA system we use N channel groups, i.e. C/N channels can be used in every cell. We denote this as FCA reuse N. For the pattern used here, the lowest possible reuse for pure FCA is three. It is also the best performing FCA in this environment but it was not included in our earlier results [2]. B. Numerical Results Fig. 3 shows a comparison between the BDCA and FCA reuse three. All users are from class A. The capacity is calculated as the value of ω C for = 2%. We see that BDCA has more than three (0.81 versus 0.25) times higher capacity here. Our previous algorithm [2] would only allow a load of 0.6 here which means that our new algorithms improve performance by 35% User perf. vs. load. RUs=64,PCU=1 TDMA frames) to compare the performance between different bunch sizes covering a fixed size area. The result is shown in Fig. 4. All users are from class A but due to the coding rate (R=0.5) used, they need two RUs each. In this way we can make a reasonably fair comparison with the results in Fig. 3 where R=1. We see that the larger bunch sizes (fewer s) give better performance than the small ones. With 2 bunches (36 s per bunch) we can allow 30% higher load than with 72 bunches. The performance for the large bunches is limited by lack of channels due to the coding rate. This is unfair to the larger bunches since in reality not all mobiles would need the same coding rate. Mobiles in the middle of a large bunch will not need much coding at all. In small bunches on the other hand, almost all mobiles will need much coding. Optimization of the code rate will giver higher capacities for all bunch sizes. Anyhow the results show that the performance with four bunches is at least 30% better than the FCA system in Fig. 3. User perf. vs. load. PCU=1,T sim =1000 frames,t PC =40 fr,r=0.5,i depth =8,prct= Rand BDCA,2 bunches Rand BDCA,4 bunches Rand BDCA,72 bunches Rand BDCA,stat=all Rand BDCA,stat=center Rand FCA,reuse=3,stat=all Rand FCA,reuse=3,stat=center Fig. 3 Comparison between BDCA and FCA for stat=all and stat=center. Both systems use DCPC power control with one update (PCU=1). The lines without markers are the calculated theoretical bounds from (4). We also compare the two cases when statistics are collected from the whole area and only from the eight s in the center (shown in Fig. 2). As can be seen in Fig. 3, the border effect is negligible for FCA since the curves for stat=center and stat=all almost coincide. For BDCA, the performance is actually slightly better in the center compared to the average over the whole bunch. The reason is that some of the border cells are larger than the standard size which leads to a lack of channels. We also see that both the FCA and the BDCA are close to their respective bounds which indicates that the systems are mostly limited by blocking and not by interference. In the following, we use statistics from the whole area. Due to computational complexity, signaling load and synchronization problems, one single bunch might not be possible if we want to cover a large area. However, our improved RRM algorithms will allow larger bunches due to the decreased complexity. For instance, with a 12 by 12 bunch and ϖ C = 0.6. our old algorithm required on average flops per allocation while our new algorithm requires only flops. Anyhow, there will be some kind of upper limit for a bunch and therefore we ran a short dynamic simulation ( Fig. 4 Three different bunch sizes covering the 12 by 12 blocks. The power control period (T PC) is 40 frames. In order to see the effects of heterogeneous traffic we compared BDCA and FCA with the traffic mix {A,B}.To be fair to the FCA we increased the number of RUs in the system to 192 (three carriers). As can be seen in Fig. 5, the BDCA performs much better than FCA. Since class B (16 RUs each) is limiting the capacity we calculate the performance at B = 2%. In this case BDCA has three times higher capacity (0.5 vs. 0.17) than FCA User perf. vs. load. T mix =AB,RUs=192,PCU=1 ClassA:Rand BDCA ClassB:Rand BDCA ClassA:Rand FCA,reuse=3 ClassB:Rand FCA,reuse= Fig. 5 Comparison between BDCA and FCA reuse 3 with traffic mix {A,B}. We have also compared BDCA and FCA with traffic mix {A,C}. Class C users are more flexible than class B since the former can be admitted when only eight of the sixteen re-

5 quested RUs are available. In Fig. 6 we can see that BDCA once again performs almost three (0.52 vs. 0.18) times better than FCA at C = 2%. Interesting to note here is that class A performs worse here compared to Fig. 5. The reason is that when class B users fail there is more RUs left for class A than when class C users fail User perf. vs. load. T mix =AC,RUs=192,PCU=1 ClassA:Rand BDCA ClassC:Rand BDCA ClassA:Rand FCA,reuse=3 ClassC:Rand FCA,reuse= Fig. 6 BDCA and FCA with reuse 3. The traffic mix is {A,C}. V. CONCLUSIONS In this paper we have studied the performance for the bunch concept with an improved hybrid dynamic/random resource management scheme. The computational complexity for a mobile allocation in a 12 by 12 bunch is improved by approximately two orders of magnitude compared to our old algorithms. The capacity for a single bunch is not far from the upper bound, especially when border effects are excluded. A capacity advantage by a factor of more than three over FCA is obtained with homogeneous traffic. The situation is the same with heterogeneous traffic even though the number of channels is tripled which should favor FCA. With a lower number of channels, the difference would be even larger. For the case with several small bunches and homogeneous traffic we obtained at least 30% higher capacity than FCA. Slightly larger bunches and optimized code rates will increase that figure substantially. Perfect estimates of SIR and link gains were assumed in this paper which means that the results are a bit optimistic in that sense. Further work include optimizing the performance for multiple bunch systems and study of performance degradation due to measurement errors. VI. ACKNOWLEDGMENT This work has been performed in the framework of the ACTS project FRAMES, partly funded by the European Commission. The author would like to acknowledge the contributions of his colleagues from Siemens AG, Roke Manor Research Limited, Ericsson Radio Systems AB, Nokia Corporation, Technical University of Delft, University of Oulu, France Telecom CNET, Centre Suisse d Electronique et de Microtechnique SA, ETHZ, University of Kaiserslautern, Chalmers University of Technology, The Royal Institute of Technology, Instituto Superior Técnico and Integracion y Sistema. REFERENCES [1] E. Anderlind, Resource Allocation for Heterogeneous Traffic in a Wireless Network, Proc. 6th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Toronto, Sept [2] M. Berg, S. Pettersson, J. Zander, A Radio Resource Management Concept for Bunched Personal Communication Systems, Proc. Multiaccess, Mobility and Teletraffic for Personal Communications Workshop, MMT 97, Melbourne, Dec [3] U. Dropmann, X. Lagrange, P. Godlewski, "Architecture of a Multi-Cell Centralized Packet Access System", Proc. PIMRC'95, Toronto, Sept 1995, pp [4] Kronestedt, Frodigh, Wallstedt, Radio Network Performance for Indoor Cellular Radio Systems, Proc. ICUPC 96. [5] Broddner, Lilliestråhle, Wallstedt, "Evolution of Cellular Technology for Indoor Coverage", Proc. ISSLS 96, Melbourne, Feb [6] H. Olofsson, "Interference Diversity as means for increasing capacity in GSM", Proc. EPMCC 95, Bologna, Italy, Nov [7] A. Klein, R. Pirhonen, J. Sköld, and R. Suoranta, FRAMES multiple access Mode 1 - wideband TDMA with and without spreading, Proc. PIMRC 97, Helsinki, Sept. 1997, pp [8] S.A. Grandhi, J. Zander, R. Yates, "Constrained power control", Wireless Personal Communications, (Kluwer) vol 2, no3, Aug [9] N. Bambos, S.C. Chen, D. Mitra, "Channel probing for distributed access in wireless communication networks", Proc. GLOBECOM 95, Singapore, Nov [10] M. Andersin, Z. Rosberg, J. Zander, "Soft Admission in Cellular PCS with constrained power control and noise", Proc. 5th WINLAB Workshop on Third Generation Wireless Information Networks, New Brunswick, NJ, [11] J. Whitehead, "Performance and Capacity of Distributed Dynamic Channel assignment and Power Control in Shadow Fading", Proc. International Conference on Communication, [12] J. Zander, "Radio Resource Management - an overview", Proc. VTC 96, Atlanta, GA, May [13] Universal Mobile Telecommunications System (UMTS); Selection procedures for the choice of radio transmission technologies of the UMTS, UMTS Technical Report 30.03, [14] J.-E. Berg, "A Recursive Method for Street Microcell Path Loss Calculations", Proc. PIMRC 95, Vol 1, pp