SPATIAL SCHEDULING ALGORITHMS FOR WIRELESS SYSTEMS

Diego Bartolome, Antonio Pascual-Iserte, Ana I. Perez-Neira *

{diego, tonip, anuska} @gps.tsc.upc.es
Universitat Politbcnica de Catalunya (UPC) Centre Tecnolhgic de Telecomunicacions de Catalunya (CTTC)
c/ Jordi Girona 1-3, Mhdul D5, Campus Nord UPC c/ Gran Capita 2-4, Edifici Nexus 1, Planta 2
08034 Barcelona (Spain) 08034 Barcelona (Spain)

ABSTRACT As reported in [31, the performance of a Spatial Division Mul-

tiple Access (SDMA) system is better when each antenna at the BS
This paper addresses the problem of the spatial scheduling of users is assigned to a user rather than using all the antennas for transmis-
in a cell for simultaneous downlink transmission from a Base Sta- sion to a single user. On the other hand, several papers in the liter-
tion (BS) having multiple antennas under a perspective of joint ature have dealt with the problem of spatial scheduling, but they do
Physical and Medium Access Control (PHY-MAC) design. First of not take into account the spatial signature of the users. Therefore,
all, wecompute the transmit beamvectors for each group according the overall efficiency of the system is reduced because co-linear
to a Zero Forcing (ZF) criterion, which gives a simple closed-form spatial signatures may be allocated in the same group. In this pa-
solution. We show first that it is equivalent to the minimization of per, the spatial signatures of the users are taken into account in the
the maximum Bit Error Rate (BER). In this paper, the main con- optimal group assignment for the scheduling.
tribution lies on the resolution of the NP-complete combinatorial In [4], the authors maximize the capacity of a SDMAITDMA
problem that comes up as cost function if we want to minimize the system. As in our work, their problem also resides on the best
total transmit power. The solution of the NP-complete problem is combination of users that optimizes a certain criterion, in their case
performed by the stochastic technique Simulated Annealing (SA). the maximization of capacity. They propose a solution based on
Additionally, we present two heuristic algorithms that may enable graph-theory, which is a NP-complete problem, so they point out
a real-time implementation of this scheduling approach. the need for efficient heuristic algorithms.
In our paper, instead of the maximization of the capacity, we
have chosen the minimization of transmit power as a cost function.
1. INTRODUCTION In fact, we vote for a realistic physical parameter that yields to a
reduction in the power consumption and that can be easily used
Multiple Element Antenna systems may provide a large increase in in higher layers. On the other hand, we are taking into consider-
capacity for future wireless communications standards. However, ation a spatio-temporal scheduler that could be developed in real
in realistic scenarios and systems, the number of antennas might time, so exhaustive search methods or the solution in [4] based on
be limited by the high cost of the RF front-ends. Therefore, we graph-theory have been disregarded. Firstly, we implement SA,
focus on the use of a small set of antennas at the BS, or at the which obtains good near-optimal solutions with reasonable com-
Access Point (AP) in Wireless Local Area Networks (WLANs). puting time. In fact, simulations have shown to obtain the opti-
In this paper, we address mainly the problem of clustering a set mum when the number of users is low. As a main contribution, we
of users for simultaneous transmission in the downlink of a TDMA propose simple heuristic solutions that may enable the real-time
system. We assume that the BS has to allocate the K users in the operation of the scheduler with only a slight power degradation.
cell in groups of Q, which is the number of transmit antennas. This paper is organized as follows. First, in Section 2 we de-
After solving this combinatonal problem, G groups are formed scribe our problem and find the global cost function to be min-
and the scheduler at the BS will choose how to allocate them in imized, together with the above-mentioned equivalence. Then,
different time slots, but this is out of the scope of this paper. Section 3 deals with the proposed solutions to the problem, i.e. SA
To begin with, we have to choose a transmit beamforming de- and the heuristic algorithms. After that, we give some simulations
sign. Optimal downlink beamforming is addressed in [ I ] or in results in Section 4 and the conclusions.
[2]. In the former, the transmit power is minimized while assur-
ing a certain Signal to Interference and Noise Ratio (SINR) at the
2. PROBLEM STATEMENT
receivers. However, this kind of techniques are difficult to imple-
ment on a real system. As practical issues play an important role, In the following, boldface capital (lowercase) letters refer to mat%
we have selected the minimization of the transmit power subject ces (vectors). The operator (.)* denotes conjugation, ( . ) T transpo-
to a ZF criterion. This criterion is suboptimal, but the key point is
sition,and (.)" = ((.)*)=. Theelement at row ithandcolumnjth
that it provides a simple closed-form solution and guarantees that
of matrix A is denoted by [A],,. We consider the downlink of a
there is no interference among users belonging to the same group.
communications system, where the BS is provided with Q transmit
antennas, although the notation and solutions presented henceforth
'This work is partially supported by the Telecommunications Tech-
nological Center of Catalonia (CTTC), by the Spanish govemment FIT- .. to the uDlink. Let K -
can be directly. applied > Q. be the number of
070000-2000-649 (Medea+ A105 UniLan). TlC99-0849 and the Integrated active users in the cell, each having a single antenna. The users are
Action HF2WI-0055 and by the Catalan govemment thanks to grants then distributed into G = [K/Q1 groups. Each group is sched-
2001SGR 00268.2003FI 00190 and 2001FI 00714. uled for transmission in a different time slot, whereas the Q users

0-7803-7663-3/03/$17.0002003 IEEE IV - 185 ICASSP 2003

in one group are served simultaneously by a SDMA scheme, i.e. a
different beamformer for each user.
Motivated within the context of Orthogonal Frequency Divi-
son Multiplexing (OFDM) systems, such as WLANS', the channel
is assumed Rat fading. In those systems, the frequency-selective
channel in the time domain is converted into a flat fading channel
at each of the useful subcaniers, see [5] for details.
Our problem is the minimization of the total transmit power,
m ...-
subject to a ZF design criterion for each group. ZF implies that at Fig. 1. System configuration: a BS with Q antennas aiming to
group g: 1 5 g 5 G, there is no interference among users and the transmit simultaneously to groups of Q users (mobile terminals).
same signal level is achieved for all of them. Let us first consider
the signal model for this MISO link'. The received signal vector 2.1. Equivalent problem
for the group at time instant n is y ( n ) :
We show here that the solution to Eq. (6) is equivalent to the min.
Q imization of the maximum BER of all users and groups. First,
y(n) = x H b i s x ( n ) E C'"'; (1) recall that an additional parameter a b can be inserted in Eq. (2).
k=l which denotes the desired signal power at the receiver4. With this
where bb is the transmit beamvector for the kth user and H is the approach, the optimum solution is that minimizing the global BER
QxQ complex flat-fading channel matrix, the ith row (1 5 i 5 Q) of all the users in the cell, e.g. if the symbols are BPSK and there
of which contains the 1 x Q vector of the channel gains for the ith is no interference among users in the same group',
user, i.e. hi. At position k, 1 5 k 5 Q, of vector y(n) we find
the received signal for the user kth. The transmitted symbols have
unitary mean energy E { Isc(n)/'} = 1.
(7)
As stated before, for each user we want to minimize the power
of the beamvector according to a ZF criterion:
min bfbk s.t. Hbk =lk, (2) where S N R k , , = w, being the noise power, is the SNR
of user kth in group g. This is equivalent to the minimization of
where the vector 1 k is 1 at position kth and 0 elsewhere, It is
the I-norm of the BERs for all users in all groups. As this problem
straightforward to obtain the solution to this problem:
is difficult to solve, we can choose the infinity-norm instead of the
bk = HH (HH")-' 1 k E CQx'.
I-norm. Therefore, Eq. (7) is now transformed into:
(3)
Applying an eigenvector decomposition to the hennitian ma-
trix HHH = UXUH , il is easy lo verify that the power from
the kth transmit beamvector reduces to: This problem is in fact the minimization of the maximum BER
among all the users and groups. Intuitively, let us have an scenario
(4) where a user has a high SNR (thus very low BER) and the other
users equally low SNR. Then, we can always increase the SNRs
where A, is the mth eigenvalue of matrix HHH and um is the of bad users so that their BER is reduced. In the end, all the users
normalized eigenvector associated with that eigenvalue (the mth will have the same SNR (and BER), therefore (ckk,g/' = a,Vk,g,
column of matrix U). If we sum up the powers of the beamvectors thus the ZF solution if we set LI = 1. Finally. note that the ZF
for the Q users in group, the total power can be expressed as: approach does not guarantee the minimization of the mean BER.

3. PROPOSED METHODS

Several options have been evaluated so as to solve the combinato-

Inserting the index g denoting the group, the total transmit power rial problem of Eq. (6). First, exhaustive search in the possibility
P,added up over all the groups is finally expressed as: space or the graph-based solution in [4J,are prohibitively expen-
sive in terms of complexity when the number of users increases.
Instead, we have applied the stochastic method SA, which may
yield to the optimum solution with a reasonable computing time.
Moreover, we have developed some low complex heuristic algo-
In order to minimize the global transmit power P, in Eq. (61, rithms that achieve a low degradation with respect to SA.
we have to find the best distribution of the users in groups (among
all alternatives). Indeed, this is a combinatorial problem. As it
will be stated. the global transmit power may vary substantially 3.1. Simulated annealing
depending on how the users are grouped together.
SA is an iterative algorithm which is capable of finding the global
'The main current swdards are Hiperlad2 and IEEE 802.1 la. optimum of a function even if the problem is not convex. SA has
'Suictly speaking, we should also add the group index g. But far sim- analogies with the physical process of annealing, i.e. cooling of a
plicity, this is omitted until Eq. (6).
'E is b e diagonal matrix of the positive eigenvalues and the columns 'The lk vector will have now air at position kth
of the unitary matrix U contain the eigenvectors associated to them. 5Thiscondition is satisfi ed by the ZF solution.

IV- I86
 Table 1. Simulated Annealing Table 3. Heuristic algorithms
I . Choose an initial solution G. This G x Q matrix contains 2 1. Build matrix D as in Eq. (IO).
user index at each position g,li, i.e. each row contains the 4 2. For the selected criterion, either MAX-MIN or MAX-RATIO.
users belonging to the same group. store the cost values in the vector c and keep the information

3. Initially. P,,,*" - P (G) and G,i, -

2. Compute the global transmitted power P (G) as in Eq. (6).
G.
4. Generate a new solotion G' exchanging two users' positions ir
about the user and dso about the group that has the minimum
cost for that user, i.e. the minimum of each row of D.
3. Sort c is descending order, c' is the soned vector. Set position
index 1 to 0 . Note that g ( l ) is the group having minimum cost
differentgroups. for the user at position 1.
5. As in Eq. (6). compute the power P (G') and the differenct
4. Increase 1 and SeleCt that position in the soned vector c a . If
with the previous solution A(P) = P (G')- P ( G ) .
the group g(1) has not been 6 lled in C u m n t iteration, assign to
6. Accept 6' with probability min
( I,e
"1.
T If G' is ac
group g(1) the user corresponding to index 1.
5. Repeat step 4 until 1 points to the laSt element of c 9 or if the G
cepted. G t G'.
1. IfP (G') 5 Pm<n. then P,,. t P ( G ' )and G,,, G'
8. If stopping condition not satisfi ed, decrease T and repeat 4.
- groups have k e n fi lled by a differentuser at this iteration.
6. Repeat step I until every user is assigned to a group.

9. Retum the solution: G,,, and P,,,. 3.2. Heuristic techniques

In this section, we propose much simpler methods than exhaustive
system [6]. The key parameter is the temperature T , which helps searcb or SA, which yield to suboptimal performances but reduce
avoiding local minima, as it is explained below. The application of drastically the computational load. Basically, we propose two ap-
this algorithm to our problem is summarized in Table 1. Now, let proaches: MAX-MIN and MAX-RATIO. Essentially, they separate
us briefly describe the technique. the users that are close in terms of angle of arrival, the spatial sig-
natures of which are "cry similar or highly co-linear, see Table
Given a current solution, a new one is proposed by exchang- 4. If we grouped these users together, the required power would
ing two users belonging to different groups. If it is better than increase substantially.
the previous one, it is accepted as the current solution, else, it Their initialization procedure is described in Table 2. After
is accepted with a certain probability. This mechanism is called that, we have the G users that are most difficult to separate asso-
kill-climbing, and avoids finding a local minimum. The parameter ciated to G different groups. We can then proceed in assigning
that controls the acceptance probability is the temperature T . The remaining USCIS to groups. Therefore, at each iteration of the d-
higher it is, the higher the acceptance probability. Therefore, T gorithm we need to build the K' x G' matrix D, where K' is
shall be lowered gradually, so that asymptotically only better s o b the number of remaining users and G' the number of uncompleted
tions are accepted and a minimum is achieved. Better in our case groups. At the k'th row and g'th column we put the cost of assign-
means having less global transmitted power, see Eq. (6). ing user index k' to group g', i.e.
Initially, the temperature shall be high enough in order to ac-
cept most of the proposed solutions, in our case we increase T until
the acceptance ratio is 95%. In this work, we run N,, = 50 iter-
ations per value of temperature. After that, we run the algorithm
until the acceptance ratio is lower than 0.05 for 5 times (stopping where H(g') is the Q' x Q channel matrix of group 9'. the last row

ponential profile T -
condition). After each run, we lower the temperature with an ex-
PT, fi = 0.9. We have chosen a low 0in
order to speed up the algorithm
of which corresponds to the channel of user k'. Note that Q' - 1
is the number of users belonging to group g' at that iteration.
The basic procedure for both algorithms is depicted in Table
3. They differ in the way the cost values are computed (step 2).
Whereas the MAX-MIN approach selects the minimum cost of as-
Table 2. Initialization procedure for heuristic algorithms signing a user to a group, i.e. the minimum of each row of D,
I, Compute the cost of all possible cumbinations of two users. The MAX-RATIO calculates the ratio between the maximum and mini-
cost of clustering together usen i and j is: mum cost for each user, i.e. the maximum and minimum values at
every row of matrix D. We shall remark that at each iteration we
assign at most G users, i.e. we do not assign two users to the same
group nor fill groups that are not optimal for the users.
where the matrix H,, is defined as in Eq. ( I ) , but in this cast Now, let us briefly explain the ideas behind the selected crite-
only for users i and j. ria. The MAX-MIN approach selects the best-suited group for each
2. Son the cos1 values of the matrix C in descendine " order in the user and it assigns then first those who are most difficult to cluster,
vector cs, keeping the information of the index 2, j in the matrix i.e. those having a maximum value of the minimum cost. In terms
c.Set 1 to 0. of performance. it does not take into account that a user may not
3. Increase 1. Select the lth position in vector cB. If the uscrs be assigned to its best-suited group, e.g. because it is already full.
i , j corresponding to the index I have not yet been assigned. Then, a big degradation may occur.
separate them in different groups g ( i ) # g ( j ) . Note that g ( i ) Instead, the selection criterion shall be related with the disper-
is the group where we put user i and that only one user shall be
&signed per group. sion of the cost values. We shall first allocate in the groups the
users that may provoke a high penalty in terms of power if they
4. lithe number of assigned users is not G. repeat step 3. are not assigned to the best group. This is the behavior of the

IV - 187
Fig. 2. Example of two users that have to be assigned to groups.
MAX-MIN first allocates user j whereas MAX-RATIO selects user
i. If user i is not assigned to group 1 a high degradation occurs,
while if user j is not in group 1,the loss in performance is negligi-
ble. MAX-RATIO allocates first the users incurring in a big loss if
they are not in the best group, thus achieving a better performance. '4L w_DI-s

Fig. 3. Outage Powers (90% and 95%) for Q=2

i

Table 4. Spatial distribution of users. Single run, G = 4, Q = 3.

MAX-RATIO proposal: instead of selecting the users according to

the minimum costs, we son them by means of the ratio between
the maximum and the minimum costs. Note that other approaches
might also be possible as long as they consider the variation of
the cost values. For an illustrative example of the behavior of the
heuristic algorithms, see Fig. 2.
l
'a1
.
IS
.
1
,
,I 8
,
*---r
.I
, I
. I,
, .
, .I
, .l
Fig. 4. Outage Powers (90% and 95%) for Q=4.
4. SIMULATIONS
of anival) are separated in different groups. If we grouped them
N , = 10000 rnns of the simulations have been conducted for a together, the increase in power would he remarkable.
number of transmit antennas ranging from Q = 2 to Q = 4, and
a number of groups' from G = 2 to G = 7. The channel is flat 5. CONCLUSIONS
fading, and static users are distributed uniformly between -60'
and 60" (sectored antennas at the BS ). No noise is included in the There are two main contributions of this paper. First, we deal with
simulations. The aim is to see what is the performance in terms of the spatial scheduling a set of users in groups for simultaneous ser-
power for the techniques that have been previously described. vice and propose a new joint MAC-PHY spatial channel allocation.
We have also simulated the random scheduler, i.e. the users are Second, we focus on the techniques for the resolution of this NP-
arbitrarily allocated to the groups. We see in Figs. 3 and 4 that this complete problem. Therefore. we apply the stochastic technique
solution is not convenient at all as the output power' is at least one SA and propose suboptimal solutions that are shown to have a low
magnitude order worse than the MAX-RATIO solution. We also degradation with respect to the SAperformance with much lower
see that the MAX-MIN approach yields to had performances due computational load. As this is a first insight in the joint Physical
to the fact explained in the example (see Fig. 2): it does not take and Medium Access Control (MAC-PHY) design, the next step is
into account all the possible options far a user, i.e. it may not be to take fairness issues [7] into consideration.
assigned to the best-suited group, so this worsens the performance
because of a high penalty in terms of power. 6. REFERENCES
As expected, SA has the hest performance. Note that the total
power when G is low is greater than in the high range. This is [ I ] E Rashid-Fmkhi et al., "Transmit Beamforming and Power Control
far Cellular Wireless Systems:' lEEE JSAC, Oct. 1998.
due to the fact that, if we have very few users and they are close
together, we cannot cope with their separation. On the other hand,
121 R. SVidh and M. Benglssan and B. Ottersten, "System Evaluation
of Optimal Downlink Beamforming in Wireless Communication," in
when we increase G, we also increase the degrees o f freedom of VTC Fall. 2001.
the system. This allows a better separation of the groups. In turn,
[3] R.W. Heath and M. Airy and A.J. Paulraj. "Multiuser Diversity far
the MAX-RATIO technique approaches the minimum power of SA. MIMO Wireless Systems with Linear Receivers:' inilrilomar, 2001.
The degradation is mainly between 3 dB and 6 dB. The key advan-
[4] R. Zhang, "Scheduling for Maximum Capacity in SDMA,TDMA Net-
tage is that the the computational load is redwmd drastically paying works: in ICASSP. 2002.
the price of a slight increase in the global transmit power. (51 L.J. Cimini, "Analysis and Simulation of a Digitll Mobile Radio
Finally. we show in Table 4 the spatial distribution of users Channel using Orthogonal Frequency Division Multiplexing," IEEE
given by a single Tun of the simulation with the MAX-RATIO SO- Transrcrionr on Communications, July 1985.
lution. It is seen that the users that are closer (in terms of angle [61 P.J.M. van Laarhaven and E.H.L. AaRr, Simuiared Annealing: Theory
and Applicatk" Kluwer Academic Publishers, 1987.
6For simplicity, we assume K to be a multiple of Q. 171 D.N.C. Tse, "Multiuser Diversity in Wireless Networks," Wireless
'In those figures, we show the ourage power for a cenain threshold Z . Communicorions Seminnr Standford Univerriry. April ZOOl.
That means that the x% of the obtained powers are below the platted value.