Opportunist'ic Uplilnk Scheduling for 3G LTE Systems

Mohammed Al-Rawi t Riku Yintti t Johan Torsner Mats S'agfors I t,

t Helsinki University of Technology, Communications Lab, Finland.

{mohammed. alrawi, riku.jantti} @ tkk.fi
+
Wireless Access Networks, Nomadic Lab, Ericsson Research, Finland.
{johan torsner mats sagfors}@ericsson corn

Abstract sideration point of view due to the fact that the require
ments on the power amplifier are weaker for lower PAPR.
In this paper we propose a heuristic opportunistic sched- SC-FDMA is based on transmitting frequency chunks con-
uler that is consistent with the resource allocation con- sisting of multiple subcarriers at a time. These chunks are
straints of the aplink channel for 3G long term evolution usually referred to as resource blocks (RB). SC-FDMA has
(LTE) systems. Retransmission processes as well as chan- two modes: localized-FDMA (L-FDMA) where users are
nel conditions are taken into account in the formation of the assigned RBs of adjacent subcarriers and the other mode
scheduling decision The heuristic scheduler could be seen is distributed-FDMA (D-FDMA) where the subcarriers of
a reasonable choice for solving the constrained resource al- one RB are distributed over the entire frequency band but
location maximization problem since it is computationally with an equal distance of each other. D-FDMA has the ad-
feasible and can find a practical solution. We also formulate vantage of being robust against frequency selective fading
an integer programming problem that provides a theoretical because its information is spread across the entire signal
optimal solution. The heuristic scheduler was found to per- band. Therefore, it offers the advantage of frequency di-
form relatively well when compared to the optimal solation versity On the other hand L FDMA can potentially achieve
multi-user diversity in the presence of frequency selective
fading if it assigns each user to subcarriers in a portion of
the signal band where that user has favorable transmission
characteristic (high channel gain). Multi-user diversity re-
1. Introduction lies on independent fading among dispersed transmitters.
Opportunistic channel adaptive scheduling has drawn a
3G long term evolution is steadily establishing itself as lot of attention since its introduction [2] Adaptive coding
being the step paving the way toward 4G mobile commu- and modulation is utilized to match the users data rate to the
nications technology LTE will stretch the performance of current channel condition. Such a scheduling benefits from
3G technology, thereby meeting user expectations in a 10- the multiuser diversity.
year perspective and beyo nd. The fundamental targets of Lim et al [3] propose an opportunistic scheduler for
this evolution - to further reduce user and operator costs and SC-FDMA systems. The authors suggest assigning RBs to
to improve service provisioning - will be met through im- users who obtain the highest marginal utility regardless of
proved coverage and system capacity as well as increased the location of other RBs that are already allocated The
data rates and reduced latency [I]. The air interface access algorithm in [3] can be interpreted as a special case of the
technologies considered for LTE are orthogonal frequency gradient scheduling rule discussed e.g. in [6]. Lim et al. as-
division multiple access (OFDMA) for the downlink and sumed perfect channel knowledge. In [4], a channel depen-
single carrier frequency division multiplexing access (SC- dent time domain scheduler that assigns all RBs to the user
FDMA) for the uplink. SC-FDMA is technically similar to who has the largest average gain to interference ratio (GIR)
OFDMA except that it has restrictions in frequency alloca- in every transmission time interval (TTI) is suggested. The
tion. SC-FDMA has the advantage of having a lower en- work also suggests a time-frequency scheduler that assigns
velope peak to average power ratio (PAPR) than OFDMA. groups of multiple consecutive RBs to users with the high-
This feature made SC-FDMA a better choice than OFDMA est average GIR over the RBs of a group in every TTI. The
for the uplink from the user equipment (UE) battery con- number of the groups can be equal to the number of active

978-1-4244-1841-1/08/$25.00 )2008 IEEE 705

users as long as the number of active users does not exceed P k is the amount of power allocated to subcarrier k for
the total number of resource blocks. user i. Gi,k is the path gain for that subcarrier. S(Oo) is a
Our paper follows the approach taken in [3]. That is, we 0 degrees directional antenna gain. P is the noise power.
consider schedulers that seek to maximize throughput based Pm,.. is the maximum transmission power. We assume all
aggregate utility. Our work differs from the previous work users transmit at maximum power. Li is the number of RBs
in the way RBs are allocated to the users. In this paper, we assigned to user i. VII is the number of subcarriers in one
study localized allocation of RBs. That is, all the subcarri- RB and is the same for all RBs since we assume they all
ers in the RBs assigned to a single user must be adjacent to have the same size.
each other In addition we will consider the effect of imper
fect channel state information and hybrid automatic repeat 2.1. Sound'ing
request (HARQ) on the scheduling decisions.
This paper is organized as follows. Section 2 describes Since the channel-dependent scheduling can only be ap-
the system model for this study Section 3 proposes a the plied to low-speed UEs, usually the localized RB is as-
oretical scheduler that provides an optimal solution to the signed to transmit the channel-dependent scheduling traf-
block allocatioun problemi. Sectioun 4 iintroduces the heuris- fic. In the case of localized data transmission the reference
tic scheduling algorithm. Section 5 contains the numerical signal is also localized which means that the reference sig
results and finally Section 6 concludes the paper. nal occupies the same spectrum as data transmission in two
short blocks. Only one sounding pilot is required for each
2. System model UE in each RB. Therefore in each localized RB, multiple
uplink sounding channels can be supported for UEs which
We consider the uplink of a single-cell model that uti- are not transmitting in the current RB and the current sub-
lizes L-FDMA. The cell contains one base station commu- frame. On the other hand a sounding pilot should be trans-
nicating simultaneously with N mobile user terminals. The mitted in every RB in order for the node B to sound the
bandwidth W consists of M subcarriers that are grouped channel over the whole transmission bandwidth for each
int oL-- w RBss where r AfJC denotes coherence band
e the coeec UE Although this limits the number of sounding UEs in
Af
width of the channel. Each RB will contain MI L consecu- each sub-frame, more uplink sounding channels can be ob-
tive subcarriers. The channel is assumed to be slowly fading tained by TDM because each low-speed UE can perform
such that the channel state stays essentially constant during uplink channel sounding over multiple sub-frames [5].
one TTI. That is, the coherence time of the channel is as-
sumed to be longer than the duration of the TTI and thus 2.2. RNetransm'iss'ions
the channel exhibits block fading characteristics. The RBs
fade independently, but the fading seen by individual sub- Assume that the number of transport formats (modula-
carriers in a RB is approximately the same since the subcar- tion and coding schemes) is large so that the correspond-
rier spacing is small compared to the coherence bandwidth ing data rates can be approximated with a real number
of the channel. We assume that the channel is subjected c. In case of perfect channel state information (CSI), the
to Rayleigh fading. We assume that the amount of power rate assigned to block Ti is matched to the channel state
available for every user will be constant and is represented ,n(t) C,(t). This would require timely knowledge
by the maximum transmission power Pma ,. The transmis- of the channel state of all the RBs at the transmitter In
sion time interval for the uplink is I ms and consists of two practice, we would need to select the rate based on possi-
O. ms sUibframes. The capacity of RB ri for useri at TTI k bly outdated and imperfect CSI. If the selected rate exceeds
is given by the instantaneous block capacity jij r(t) > Ci,n t), then
the transmitted data cannot be decoded at the receiver. In
C i, = B 1log2 ( +-vi,f)
(eff)0
(n) that case, we assume that synchronous non-adaptive HARQ
protocol is utilized. The characteristics of this solution is
where Bn is the bandwidth of RB I, yI(eff) (t) denotes the that retransmissions occur at a predefined (normally fixed)
effective (SNR) of user i on RB n and is computed as fol- time after the previous (re)transmission using exactly the
lows same RBs. The same code rate is used for retransmissions
-1 even though the channel might have changed. The bene-
(eff) I fit with synchronous non-adaptive HARQ is that the uplink
i,n
-
I
(2) control signaling can be minimized since the HARQ pro-
,k+ ' ) cess ID does not need to be signaled explicitly and a re-
S(P
-i ,k.Gi,k ) ceived HARQ negative acknowledgment (NACK) can be
Yi k 1k
Li.M
(3) used as an implicit grant for a HARQ retransmission. In
Pn

706
chase combining HARQ [7], the receiver coherently com- We introduce yi,, which will denote a selection variable:
bines the original code block and the retransmitted block. Yi,n = 1 if RB n is assigned to user i, otherwise yi,n = 0
All the transmitted bit energy can be harnessed by the re- We assume that a mobile divides its available power evenly
ceiver by combining the erroneously received code block among the assigned RBs. Based on the channel sounding,
with the consecutive copies transmitted by the ARQ pro- the scheduler forms an estimate of the rateti,r that user
cess. If the receiver is able to decode the code block after i expects to obtain if RB n is assigned to it. Given the
combining the original packet and the retransmitted repli- estimated throughput xi, the scheduler needs to solve the
cas of that packet, the actual
, 1i(t , where T
p rate would become ~~~~~~~~~~m.
following assignment problem.
m = 1, 2, m. a is the number of retransmissions and
..

T is a fixed parameter denoting a predefined time after the

previous (re)transmission. Decoding fails if this rate still y(t) m
ar max
y- i=l uixiA (5)
n=l
exceeds the capacity of the channel.
subject to
3. Localized Grad'ient Algorithm LGA
Yi,ni {O 1}
We consider the gradient algorithm as the scheduling N
metric for our scheme. The gradient algorithm is a nat- \
Yr< 1 -2 129 C-/
ural generalization of the proportional fair (PF) algorithm I=1
in that it applies to any concave utility function U(X) and Yi,n -Yi,(n+ ) + Yi,m < 1rTn = T + 2, I + 3, ~C
to the systems where multiple users can be served at a Where N is the total number of users. C is the total
time. Stolyar [6] has proved the asymptotical optimality of number of RBs. We could see that the first inequality limits
this algorithm for multiuser throughput allocation. Users the RB to one user only. The second inequality enforces
n - 1,- , N are served by a switch in discrete time the requirement for consecutive blocks. If yi,n =I and
t 0 1 2? * * * Switch state m = (m(t), t = 0? 1, 2? ** )
Yi (n+l) = 0 then yi m < O for m > n + 1. If on the
is a random ergodic process. In each state m, the switch other hand both Yin 1 and Yi-(n 1 the inequality
can choose a scheduling decision k from a set K(Tn) Each
decision k has the associated service rate vector m (k)
requires that Yi,, Yi,n 0 then the inequality states
1 If =

<-1- (1 -Yi,(n+l) )) That is, the inequality

=
that Yi,
(,rm (k),, pm (k)). This vector represents the amount of becomes redundant.
data of each user which will be served (transmitted) in one The gradient scheduler discussed above is optimal for
"time slot" if decision k is chosen. The gradient algorithm perfect channel state information In case of measurement
is defined as follows. If at time t the switch is in state m, delays and estimation errors, the selection rule occasionally
the algorithm chooses a (possibly non-unique) decision picks rates that do not match the channel capacity. We as-
sume that synchronous non-adaptive HARQ is utilized to
k(t) = arm max VU(X(t))T.Pm(k) (4)
kEK(m) take care of the errors Now the scheduler has to reserve
those RBs to the UE that has scheduled retransmissions. To
Where U is a strictly concave smooth utility function. take this into account in our integer programming problem,
X(t) is a vector representing exponentially smoothed aver- we need to add a constraint
age service rates xi Typically the utility function has the
aggregate form U(X) = ui (xi) It is worth noting Y, =
, if user i has an ARQ process on RB n (6)
that if we select ui(i) = log(i), the gradient scheduler
It is worth noting that the integer programming ap-
becomes equivalent to the well-known Proportional Fair
proach presented here does not provide the optimal solution
scheduler. In [6], it has been shown that (4) converges to
in case of imperfect channel estimates. However, it is ex-
the optimal solution of maxx U(X) as t oo.
The set K(t) denotes all feasible RB assignments that pected still to provide a close to optimal solution that we
will use as a reference. To validate this claim, we note that
can be made at time instant t. The set is confined by the
the performance loss due to retransmissions is low as shown
channel capacity as well as the constraints that we have on
in Section 5.
the allocation of the RBs. In what follows, we formulate an
integer programming assignment problem for solving k (t)
under the constraint that all RBs assigned to a single UE 4. Heuristic Localized Gradient Algolrithm
must be consecutive in the frequency domain. The integer HLGA
programming solution is then used as a reference to validate
the performance of the suggested heuristics to be introduced The localized gradient algorithm described in the pre-
in the next section. vious section requires that the scheduler solves an integer

707
programming problem for every TTI. As the number of processes. T is a fixed predefined time. Iterating for F1 with
users and available resource blocks grow, the computational Fi(t +T)(0){RB1 RB2<. C} wehave
complexity and time required to solve the problem soon be-
comes infeasible. Hence, there is a need for simpler algo- F>Y+)(t+ T) F> Y)(t+ T) \ Jr(t +T), t
rithms that can find good enough solutions very fast. In
this section we suggest a scheduling algorithm that follows Numercal Example
a simple heuristics in allocating the resource blocks to the For a better understanding of the heuristics we present
users while maintaining the required allocation constraint this simple example. Assume a system with 3 users and
and taking retransmission requests into consideration. 6 RBs Assume perfect channel estimation The selection
Let J, denote the set of RBs assigned to user i. Fi denotes metric forms an i x j matrix that has the following values
the set of RBs that could be allocated to user i, (i e the RBs for time-slot t.
that do not violate the localization constraint if assigned to
user i). We Initialize by defining Ji and Fi for all i and t.
026 1.65 0.10 1.60 0.85 0.88
T(0) 0.82 0.50 0.30 0.90 0.63 0.87
{RBI ,RB2, , C} 041 039 047 062 089 059,!
Where C is the total number of RBs in the frequency band. The scheduler will start by allocating RB2 to UEI since
Step 1: Iterate by finding the user-RB pair that has the max- it forms the highest value in the matrix (1.65) and naturally
imumuu value any RB that is allocated to a user will be excluded for all
other users The next highest value is (1 60) with UEI on
(t ,j ) arg max u (x (t))pj,j(t RB4. UEI is allocated RBs 4 and 3 due to the fact that
jEEF(k ,i RB3 will fall between two RBs that belong to the same user
Step 2: Assign RB j to user i* and update Fi. (ULI) and to maintain localization it cannot be allocated to
anv other user Next is the value (0 89 with UE3 on RBS
j(k+1) Uj leading to the exclusion of RBI from the set of possible
ji8 - Ji'* RBs for UE3. Following that is (0.87) with UF2 on RB6
F(k+)
F. F(k) \pS(k), i + i*
excluding RB I for UE2. Finally RB I is allocated to UEI
since there is no possibility to grant it to any other user due
J(k) {atn: n > J(k } for users who have been as- to localization The RB allocation will have the final form
signed RBs located before J(* and J(k) tn: n <
j, } for users with RBs located after J(k I) RB index 1 2 3 4 5 6
Step 3: If user i* is assigned an RB that is not consecutive UE 1 1 1 1 3 2
to the previously assigned RB(s) then all RBs in between in
between will be allocated to that user since assigning any of
these RBs to any other user will breach the localization for 5. Numerical results
user
j(k+l) A computer simulator is used to create a single cell en
).~*
ti :.j(k)
ji*M (k) vironment. The simulator generates N users with locations
{jk f
Ji* J J }: J ii(J(k uniformly distributed over the cell area. We assume pedes-
Jd Jl: j i c trian profiles for the users hence channel conditions are
slowly changing. Different users experience different chan-
Update F, in the same way as in step 3. nel conditions that vary depending on their distance from
Step 4 Repeat the previous steps until all RBs are assigned. the base-station and speed. The speeds of the mobile users
Step 5 If a user has failed transmissions on certain RBs, were independent random variables uniformly distributed
then these RBs plus any blocks located in between two non- between 3 km/h and 10 km/h. All users are assumed to have
consecutive ARQs will be reserved for retransmission. full buffers with different packet sizes. System parameters
are shown in Table 1. We use the proportional fair rule as
Jr(t + T) {RB Q,.. ,RBJ Q}, rcR the metric for selecting the RBs in every TTI for LGA and
HLGA. Retransmissions are included in the scheduling pro-
Where RB1 ARQ represents the block with the lowest order cess and are prioritized. We compare our scheme against the
RB(a)
that has an ARQ process and RBARQ is the is the ARQARQblock solution provided by the LGA as well as the solution from a
with the highest order. R is the set of users that have ARQ blind scheduler that assigns all RBs to one user at a time in a

708
 Table 1. System parameters
Parameter Value
Carrier frequency 2 GHz
RB bandwidth 375 kHz
Total number of blocks 10 (25 subcarrier/RB)
TTI duration lms
Fading model One path Raleigh
No. of terminals 5 (Single cell)
Site to site distance 1000 m
Number of Tx antennas 1
Max. Tx power 21 dBm
Noise power -108.5 dBm

Figure 1. Normalized cell throughput with

round-robin (RR) fashion. Fig. I shows the cell throughput perfect and imperfect channel estimation
for cases of perfect and imperfect channel estimation. The
figure is normalized to the performance of a restriction-free,
retransmission-free case where there is no constraint in the
block allocation and channel estimation is assumed to be challenging task due to the fact that some resources have to
be allocated to satisfy the constraint rather than the channel
perfect. The scheduling used in this reference case is sim- condition. Finding the exact optimal solution is effort re-
ply the original non-localized version of the gradient algo- quiring. Therefore, a heuristic approach to the scheduling
rithm, non localized gradient algorithm (NLGA). We notice problem would be a suitable choice for finding a practical
that there is only a 4% gap between the LGA optimal solu-
solution to the allocation optimization problem. For this
tion and the NLGA optimal solution. The gap represents the
impact of the localization requirement. This implies that the purpose, we suggested the HLGA which provided a good
performance when compared to the optimal solution pro-
localization constraint has a low impairment on the perfor- vided by the LGA The LGA on the other hand was con-
mance of the LGA. It could also be seen in the figure that
sidered realizable only in theory due to its complexity. The
the HLGA provides a close to optimal performance when
scheme could therefore present a good benchmark to nea-
compared to the NLGA and LGA optimal solutions. For sure the perfortance of other channel dependent scheduling
the imperfect channel information case we see that the LGA
rules for localized SC-FDMA.
and HLGA still perform well with retransmissions now as-
sociated in the block scheduling decision. The gap between
the LGA solution which is now a sub-optimal solution and References
the NLGA optimal solution grows to 10%. Thus, the impact
of the retransmissions on the performance of the LGA was [Li HI Ekstrom A Furuskar J Karlsson, S Meyer, Mv Park all, J
Torsner and M. Wahlqvist, "Technical solutions for the 3G long-term
only 5c. The HLGA also maintains a good position with a evolution," IEEE Communications MaIgazine, Vol. 44, 2006.
drop in performance of only 7%. Table 2 shows the exact [2] P Bender, P Black M. Grob, R Padovani, N Sindhushayana, and A
throughput values for both cases. Viterbi, "CDMA/HDR A bandwidth efficient high speed wireless
data service for nomadic users," IEEE Communications Magazine,
Vol. 38 No. 7 pp. 70-77 2000.
Table 2. Cell throughput values in Mbps for [3] J. Liimi, H. Myung, K. Oh and D. Goodman "Proportional fair
scheduling of uplink single -carrier FDMA systems," in Proc IEEE
perfect and imperfect channel estimation PIMRC 06 pp 1-6 2006.
NLGA LGA HLGA RR [4] K. Jersenius "Uplink channel dependent scheduling for future cellu
lar systems," Master thesis, Link5pirig university, 2007.
Perfect 2678 2564 2305 169 [5] Nortel "Sounding channel for UL channel-dependent scheduling,"
Imperfect 24.08 21.29 9 83 3GPP TSG-Ran Working Group I Meeting RI-060907, May 2006.
[6] A. Stolyar "On the asymptotic optiinality of the gradient scheduling
algorithm for multiuser throughput allocation", Operations research,
Vol. 53, No. l, pp. 12-25, 2005.
6. Conclud'ing Remarks [7] S Sanayei, A. Nosratinia, N. Aldhahir, "Code Combining: A
maximum-likelihood decoding approach for combining an arbitrary
The constraint in resource allocation for the uplink chan- number of noisy packets", IEEE Troans. on Cormmnun. 1985, Vol. 33,
pp. 593-607, 1985.
nel in LTE systems makes channel dependent scheduling a

709