LTE Uplink Scheduling Algorithms: Performance and

Challenges
Haidar Safa and Kamal Tohme
Department of Computer Science
American University of Beirut
Beirut, Lebanon
{hs33, kgt02}@aub.edu.lb

Abstract LTE uses the SC-FDMA radio access technology for
its uplink transmission. As a result, resources assigned to the
same user must be contiguous in the frequency domain. Several
uplink scheduling algorithms were proposed in the literature to
fit that constraint. These algorithms take as input a matrix which
is used by the Packet Scheduler for an efficient resource
allocation. The performance of these algorithms is affected by the
paradigm that is used to construct that matrix. Two main
paradigms exist in the literature, channel dependent and
proportional fairness. In this paper, we evaluate the performance
of some LTE uplink scheduling algorithms for both channel
dependent and proportional fairness paradigms. As a result, we
identify the weaknesses of the existing paradigms and define
some challenges for future enhancement.
Keywords- LTE; uplink scheduling; channel condition;
I. INTRODUCTION
Long term evolution (LTE) [1] was developed by 3GPP to
cope with the increasing demand for better QoS and the
emergence of bandwidth consuming multimedia applications
[2]. LTE does not fully comply with the IMT's requirements
for 4G standards, but offers revolutionary performance when
compared to its predecessors [3]. To make this possible, a new
architecture, the system architecture evolution (SAE) [4] was
created, and several technologies such as orthogonal frequency
division multiplex (OFDM) [5] and multiple inputs multiple
outputs (MIMO) [6] were introduced to the cellular network
turf.
The SAE is an all-IP network, consisting of a user plane,
the evolved UTRAN (eUTRAN), and a control plane, the
evolved packet core (EPC) [4], as shown in Figure 1. Some of
the user plane's functionalities were moved from the core of the
network to its edge, leaving the eUTRAN with only one
simplified entity, the evolved NodeB (eNodeB), that is
responsible only for radio resource management (RRM),
making the core of the network flatter and less complex, thus
allowing for reduced latencies. Moreover, a new interface
between the eNodeBs was incorporated, allowing for faster
data routing.
OFDM is the radio access technology used by the eUTRAN
[5]. It offers high spectral efficiency and reduces bit errors
greatly. OFDM breaks down the information to be transmitted
and sends each piece in parallel on different subcarriers, which
can be placed tightly and even overlap in frequency. OFDM's
main disadvantage is that it has a high peak to average power
ratio (PAPR), which makes it unsuitable for the uplink. For
this reason, the single carrier-frequency division multiple
access (SC-FDMA), a variation of OFDM, that incorporates
the advantages of OFDM with the low PAPR trait of single
carrier systems was adopted by 3GPP [7]. In SC-FDMA, to
have the low PAPR trait needed, resources assigned to the
same user equipment (UE) must be contiguous in the frequency
domain, making packet scheduling for the uplink an
unprecedented problem. The resources to be assigned to users
are called resource blocks (RBs), are 180 KHz each, and
constitute the entire bandwidth.

Figure 1. System Architecture Evolution.
3GPP did not define scheduling algorithms for LTE, but
left them for the vendors to implement. In this paper, we study
the LTE uplink scheduling. We first present the LTE packet
scheduler and its main entities and describe the channel
dependent and proportional fairness paradigms which are used
to construct a users-to-resources allocating matrix that is used
later by the scheduling algorithms. We then survey the LTE
uplink scheduling algorithms found in the literature. These
algorithms are then evaluated using the network simulator NS-
3. Their performance results are analyzed and discussed. Based
on these results, we identify their weaknesses and define some
challenges for future enhancement.
II. UPLINK PACKET SCHEDULING IN LTE
A. LTE Packet Scheduler
The packet scheduler (PS) [9] is the controlling entity of
the MAC layer, and deals with allocating RBs to UEs every
transmission time interval (TTI) of 1ms. Scheduling decisions
are carried out on a per-user basis, even though each user may
have several data flows. Figure 2 shows the PS interacts with
several radio resource management (RRM) functionalities
which provide, for instance, the channel quality of UEs on
every RB frequency, and the QoS requirements of a UE. The
PS interacts closely with the hybrid ARQ (HARQ) manager,
19th International Conference on Telecommunications (ICT 2012)
978-1-4673-0747-5/12/$31.00 2012 IEEE

which is responsible for scheduling retransmissions. The link
adaptation (LA) unit provides information to the PS about the
supported modulation and coding schemes for a user.

Figure 2. Interaction between the PS and other RRM functionalities.
To provide the channel quality of UEs on every RB
frequency and the amount of data pending for transmission and
their priorities to the serving eNodeB, LTE uses mainly two
signaling messages, the sounding reference signal (SRS) and
buffer status report (BSR). LTE uses a channel sounding
technique that allows the eNodeB to monitor the channel
condition of every UE over the entire bandwidth, i.e. over all
RBs [8]. Every 1 ms, each UE send a SRS to its serving
eNodeB. The latter then extracts channel state information
(CSI) and passes it to the CSI manager. The CSI manager, in
its turn, generates a metric value for each RB for each UE,
creating a matrix called Channel Conditions Matrix. This
matrix, shown in Figure 3, is used by the PS for an efficient
resource allocation. To request resources from the eNodeB,
LTE defined an enhanced scheduling request, the BSR, which
also allows the UE to inform the eNodeB about the amount of
buffered data and their priority. In LTE, data are classified into
four groups that are called radio bearer groups (RBGs), each
of which has a different priority level. Two formats are
defined for the BSR: a short one if only one RBG is reported,
and a long one that allows four RBGs to be reported. Each
RBG holds data from applications of the same priority. The BS
manager stores these BSRs, allowing the eNodeB to know
exactly how much data each user has in each of its four
transmission queues as shown in Figure 3.

Figure 3. (left) Channel Conditions Matrix, (right) Buffer Status.
LTE Packet Scheduler is divided into two algorithms. The
first algorithm is that of the time domain (TD), in which N
users are selected for potential scheduling based on their
sessions required QoS, which are handled by the BS manager,
and such that these users have pending data for transmission
and are configured for scheduling in the next TTI. These
selected users are passed to the frequency domain (FD)
scheduling algorithm, which allocates RBs to them ensuring
that each RB is allocated to at most one UE and all RBs
allocated to a single UE are contiguous in frequency. The FD
algorithm uses the channel conditions matrix available from the
CSI manager to schedule users on RBs with high channel
quality avoiding RBs where they experiences deep fades.
B. Channel Dependent Vs. Proportional Fairness Paradigm
In channel dependent (CD) paradigm [12], the channel
quality of UEs on every RB frequency is considered when
creating the allocation matrix, while respecting the contiguity
constraint. As a result, users with high channel quality are
being assigned most resources. Therefore, the scheduling
algorithm would take as input the matrix consisting of metric
values r
c
i
(t), which represent the instantaneous channel rate for
user i on RB c at time t. The CD approach achieves the best
throughput; however it suffers from the starvation problem as
users with low quality channels are being assigned the least
resources, or even may not be assigned any resources at all.
This disadvantage is partially solved with proportional fairness
paradigm (PF) [13, 14]. Indeed, PF allocates resources to users
based on the ratio of the channel condition over the lifelong
service rate of a certain user, instead of just its channel
conditions. As a result, users with low channel conditions will
get some resources but fewer than those with better channel
conditions. Hence, fairness is proportional to the channel
conditions. By doing so, a high throughput is still achieved.
Therefore, PF constructs the matrix
c
i
(t) = r
c
i
(t)/R
i
(t), where
R
i
(t) is the long term service rate of user i till time t, and that
matrix is used as input for the scheduling algorithm.
The goal of the TD PF algorithm is to maximize the
logarithmic utility function
i
log R
i
, where R
i
is the long term
service rate of user i. To maximize this function,
i
d
i
(t)/R
i
(t)
should be maximized, where d
i
(t) is the total data transmitted to
user i at time t, and R
i
(t) is R
i
up till time t, as was proven in
[14]. To adapt this algorithm to the FD, the following notations
are defined. Let x
i
c
(t) denote if RB c is assigned to user i at
time slot t; let r
i
c
(t) be the instantaneous channel rate for user i
on RB c at time t; and let
c
i
(t) = r
c
i
(t)/R
i
(t) be the PF metric
value that user i has on RB c at time slot t. Then the goal of the
FD PF algorithm would be to maximize
i

c
x
i
c
(t)
c
i
(t), while
respecting the RB contiguity constraint.
However, incorporating the RB contiguity constraint into
scheduling algorithms was proven to be NP-hard [14], thus
exhaustive search is impractical. For this reason, all proposed
algorithms are greedy heuristic ones, i.e. they find a good
enough solution, gaining computational performance and
conceptual simplicity at the expense of accuracy. At their worst
case, they might not perform well, yet in practice, their overall
performance is very good.
C. Uplink Scheduling Algorithms
Most LTE uplink scheduling algorithms can be applied to
both Channel Dependent matrix and Proportional Fairness
matrix even though they were originally proposed as Channel
Dependent algorithms [10, 11, 12] or proportional fairness
ones [13, 14]. This is true because channel dependent and
proportional fairness contribute only to the way the allocation
matrix is formed.
First Maximum Expansion (FME) algorithm was proposed
in [12] and was used with the channel dependent matrix. It first
assigns RB j to User i such that M[i, j] is the highest metric
value, where M is the CSI matrix. Then it expands the
allocation on either column j+1 or j-1of M, depending on
which has the higher metric value as shown in Figure 4. If that
metric value belongs to UE i, the RB is allocated to it. If not,
that RB is allocated to the new UE and the allocation continues
with it. RBs keep getting allocated to the same UE as long as
that UE has the highest metric value for the current RB, or

higher metric values belong to already served UEs, which
would break the contiguity constraint if the RB gets assigned to
them. Each UE is considered served whenever another UE
having better metric is found. When the expansion on one side
reaches the end of the bandwidth, the allocation is spread out
on the other side.

Figure 4. First Maximum Expansion Algorithm.
Recursive Maximum Expansion (RME) was also proposed
in [12] and was used with channel dependent matrix. It differs
from FME by performing a recursive search of the maximum.
RME's flowchart is shown in Figure 5. Instead of expanding on
one side then the other, RME searches for the maximum,
assigns it to the corresponding UE, and expands the allocation
for only that particular UE on both the left-hand and right-hand
side, until a better RB allocation is found for another UE. Then,
the UE is put in idle mode, and the algorithm recurs. The
algorithm terminates when all RBs get allocated. If all UEs
become idle and not all RBs are allocated, RME keeps
searching for highest metric values, idling UE/RB metrics this
time, until one is found that belongs to a UE having a RB
allocated to it that is "adjacent" to the RB just found; i.e., there
are no RBs between them or RBs between them are not
allocated. When this RB is found, it is allocated to the
corresponding UE, and the allocation is expanded on both its
sides, till the contiguousness with the previous allocation is
achieved on one side, and till a better allocation is found on the
other side. The algorithm then recurs till all RBs get allocated.
The Carrier-by-Carrier in Turn was proposed in [13] and
used with a proportional fairness matrix. As shown in Figure 6,
it starts assigning resources from left to right to active users
having the highest
c
i
on the RB that the iteration in progress is
working on, deactivating users that cannot get any more RBs
due to the contiguity constraint along the way. The
proportional fairness matrix element
c
i
(t) = r
c
i
(t)/R
i
(t) is the PF
metric value that user i has on RB c at time slot t where r
i
c
(t) is
the instantaneous channel rate for user i on RB c at time t; and
R
i
(t) is the long term service rate of user i till time t

Figure 5. Recursive Maximum Expansion Algorithm.

Figure 6. Carrier-by-Carrier in Turn Algorithm.
Carrier-by-Carrier in Turn algorithm starts at one end and
iteratively goes to the other, thus it is not likely to produce a
good enough schedule. Thus the greedy strategy take the
largest first is applied in the Largest-Metric-Value-RB First
1. Let U be the set of schedulable users
2. Let A[m] be RB-to-user assignment status
3. for RB c =1 to m do
4. Pick the best user i in U with largest metric value
5. Assign RB c to user i (A[c] i)
6. Let I be RBs already assigned to user i (Prior to his
iteration)
7. if I is empty then
8. U = U {A[c-1]}
9. end if
10. end for

algorithm [13]. This algorithm, which is shown in Figure 7,
allocates RB i to User j in one iteration if there are no RBs
assigned to another user between RB i and the RBs already
allocated to User j, such that
i
j
is the highest remaining metric
value. Otherwise, it picks the second highest metric and so on.
When a RB is allocated to a user, all RBs between that RB and
RBs already allocated to that user are assigned to it as well.
Several variation of this algorithm can be implemented. The
length of the in-between RBs can be taken into
consideration, or the sum of the metrics of the in-between
RBs and the user can be compared to the sum of the metrics if
the RBs would get assigned to other users.

Figure 7. Largest-Metric-Value-RB First Algorithm.
The Riding Peaks algorithm [13] that is shown in Figure 8
tries to use each users highest valued RBs as much as possible.
It relies on the Doppler Effect, i.e. in multi-carrier systems, the
channel states of a user are correlated in both time and
frequency. So if User i has a good channel quality on RB c,
then it is highly probable that he will also have a good channel
quality on RB c-1 and RB c+1. The main idea of this algorithm
is to ride users peaks in the FD. This algorithm assigns a RB
to a user that already has allocated RBs if they are neighbors.

Figure 8. Riding Peaks.
III. PERFORMANCE EVALUATION
To study the performance of these algorithms we have
implemented and integrated into NS-3, the Recursive
Maximum Expansion (RME) algorithm, First Maximum
Expansion (FME) algorithm, and the Riding Peaks algorithm.
The simulation parameters used are summarized in Table 1.
Each of these algorithms was implemented twice, once with a
proportional fairness matrix and once with a channel dependent
matrix.
TABLE I. SIMULATION PARAMETERS
Parameter Value
System Type Single Cell
Channel Model Urban
# of Active Users in Cell 8, 24, 48, 96, 144, 192, 240
Users Distribution Random
Traffic Model Infinitely Backlogged
User Transmit Power 125 mW
System Bandwidth 5 MHz
# of RBs 25
# of Subcarriers per RB 12
RB Bandwidth 180 KHz
Transmission Time Interval (TTI) 1 ms
Maximum Delays 10 ms, 40 ms, 90 ms, 150 ms
Mobility Random < 30 Km/h
Simulation Time 10000 TTIs
A. System Throughput
We first study the total throughput achieved by the three
approaches. Results are shown in Figure 9. We can observe
that Channel Dependant (CD) Riding Peaks and CD RME
provide the best system throughput, followed by CD FME,
then PF Riding Peaks and PF RME, and finally, PF FME.
These can be justified since CD algorithms allocate RBs to
UEs with the highest c7hannel conditions, and the transmission
rate and the channel condition are proportional to one another.
Therefore, CD algorithms generate the best system throughput.
However, PF algorithms lower the system throughput in order
to provide fairness among UEs. Since FME finds the first
maximum and then works iteratively on the RBs, its lower
performance is expected. Both RME and Riding Peaks
recursively find the maximums, providing better performance
than FME. Riding Peaks achieves a slightly better system
throughput than RME, which is mainly because it finds the
maximum at each step, as opposed to RME, which finds the
maximum, then iteratively allocates RBs close to that
maximum till it encounters a RB whose metric value for the
UE in question is not the best, and then recurs.

Figure 9. System Throughput
B. Fairness Index
We then study the fairness of the algorithms using Jain's
Fairness Index. The formula is given as
1. Let V be the sorted list of all metric values in decreasing order
2. Let S be the set of not-yet-assigned RBs
3. k 1
4. while S is not empty do
5. Pick RB c with k
th
largest metric value from S
6. Let I be RBs assigned to user i
7. if (c is adjacent to I) or (I is empty) then
8. Assign RB c to user i
9. S = S c;
10. remove cs metric value from V;
11. k 1
12. else
13. k k++
14. end if
15. end while
1. Let V be the sorted list of all metric values in decreasing order
2. Let S be the set of not-yet-assigned RBs
3. k 1
4. while S is not empty do
5. Pick RB c with k
th
largest metric value from S
6. Let I be RBs assigned to user i
7. if none is yet assigned to RBs between I and c then
8. Let C be all RBs located between I and c
9. C = C U {c}
10. Assign all RBs in C to user i
11. S = S C;
12. remove all metric values of RBs in C from V;
13. k 1
14. else
15. k k++
16. end if
17. end while
(x
1
, x
2
, , x
n
) =
( x
i
n
i=1
)
2
n x
i
2 n
i=1
, w
UEs in the system and x
i
is the number of r
UE i. Jains Index returns a number betwee
being perfectly fair. Our results are presented
Figure 10. Fairness.
Riding Peaks is the best among the
followed by RME, and finally, FME. We
that fairness is very low with CD algorithm
with PF ones, which are expected results,
imminent in the CD paradigm, a problem s
algorithms resulted in a very unfair resource
is understandable, since the only UEs
resources are those with the best channel c
those not having the highest CQI w
Proportionally Fair algorithms all produces
resource allocation scheme, which slowly
number of users increases.
C. Users Served
With this very low fairness index when
used, one has to wonder about the number
being served. The left side of Figure 11 show
algorithms are because out of 240 users in th
20 where being granted resources in all 3 CD
study. PF algorithms give resources to all u
We then study the number of users being
i.e., those whose QoS requirements are m
Figure 11 shows that CD algorithms prov
number of users with their required QoS. Thi
only a small number of users are being grante
first place. PF algorithms, on the other han
results, but as the number of users increases
users get served efficiently. We can conc
Peaks, which provides the best results in all
far, is definitely the best among the algorit
followed by RME, and finally, FME.
But seeing the decreasing number of
requirements are being met as the number of
the system, we can conclude that one of the
the CD and the PF paradigms is that QoS is
is true since the TD packet scheduling alg
responsible for providing QoS to users, b
scheduler is the one that allocates resources
the highest priority users are being selec
scheduling, there are no guaranties that thos
assigned resources.

where there are n
resources given to
en 0 and 1, with 1
d in Figure 10.

three algorithms,
can clearly notice
ms, and very good
since starvation is
solved by PF. CD
e allocation, which
that will receive
conditions, and all
will be starved.
s a relatively just
decreases as the
a CD algorithm is
r of users that are
ws how unfair CD
e system, less than
D-algorithms under
users, as expected.
served efficiently;
met. Right side of
vide a very small
is is justified since
ed resources in the
nd, give far better
in the system, less
clude that Riding
l metric studied so
thms under study,
users whose QoS
f users increases in
limitations of both
s jeopardized. That
gorithm is the one
ut the FD packet
s. So even though
cted for potential
se users would get
D. Packets Arriving after their
We next study the rate of p
their average transmission de
account, in order to validate
Figure 12 shows that the pack
most belong to RBG1, which ar
priority. Less packets are bein
RBG4, which has a 0% rate of
explain also why the number o
decreases fast as the number o
By not taking RBGs into accou
equally. But lower priority da
higher priority data. So if R
consideration and higher pri
advantages, the total number o
would increase, by delaying lo
requirements would nonethele
priority data faster.
Figure 11. (left)User Served, (
Figure 12. Rate of RBG1, RBG2
r Maximum Allowed Delay
packets not arriving in time and
elays while taking RBGs into
the limitation we just defined.
kets that are being delayed the
re the packets having the highest
ng delayed working our way to
f delayed packets. These results
f users getting served efficiently
of users increases in the system.
unt, all packets are being treated
ta are more delay tolerant than
RBGs were to be taken into
iority data were being given
of users being served efficiently
ower priority users, whose QoS
ess be met, and serving higher

(right) Users Efficiently Served

2, RBG3, RBG4 Delayed Packets.


E. Mobility
This experiment shows that if a UE with good channel
conditions moves fast away of the eNodeB, or its channel
condition degrades suddenly for any reason, it will definitely
be denied resources for the services it is running if channel
dependant algorithm was used, and will most probably have its
services interrupted with proportional fairness, leading us to
another limitation in both paradigms. The simulations involved
47 stationary users randomly distributed around the eNodeB,
and 1 user close to the eNodeB, who goes underground after
transmitting for 1 minute, i.e. his CQI decreases drastically
instantaneously. We monitor that users throughput over the
simulation time. Figure 13 shows that with CD algorithms, the
user was denied service for good. With PF algorithms, on the
other hand, that UE was denied resources for around 400 ms,
and it then continued to be served with a lower throughput.
These 400 ms were the time needed for the system to stabilize,
i.e., for the long term service rate of the users who have his
new channel quality to match his own.

Figure 13. Mobiliy problem
IV. CONCLUSION AND CHALLENGES
In this paper, we have surveyed the uplink frequency packet
scheduling for LTE networks. To evaluate the performance of
some proposed algorithms, we have implemented three of them
and integrated them into NS-3. Likewise, we implemented
channel dependant and proportionally fair versions of each
algorithm in order to study their performance with both
paradigms. Results show that the riding peaks algorithm has
the best performance with both channel dependent matrix and
proportional fairness matrix.
By analyzing deeply the obtained results, we could identify
several challenges present in both the CD and the PF
paradigms. Neither CD nor PF paradigms take the UE QoS
requirements into consideration. Therefore, QoS may not be
provided. This can be illustrated by considering the users with
the lowest channel quality having the highest priority
applications. In a CD algorithm, these users would be starved.
In PF, on the other hand, these users would get assigned
resources, but since fairness is proportional to the channel
conditions, they would most probably not get assigned enough
RBs to fulfill their QoS requirements. Furthermore, there are
cases where UE mobility might lead to an unfair treatment and
disruption of service. For example, consider a user with high
channel condition who suddenly goes underground. As a
result, the channel quality of this UE will decrease swiftly; but
its long term service rate would be high, since it used to have a
favorable channel quality. This would make the ratio of its
new low channel condition and its long term service rate very
low, which denies it from receiving resources until the system
stabilizes.
Last but not least, all the literatures assume that there are n
UEs, and solve the problem of allocating RBs to them. But
one of the fundamental properties of cellular networks is their
dynamic nature, which wasn't studied in any literature. For
example, UEs close to the eNodeB may keep getting admitted
at a high frequency, run a service for a short period of time
and disconnect. These UEs will have favorable conditions, and
a long term service rate starting at 0 and finishing at a very
low value. The other UEs in the system can be divided into 2
groups. Those with low channel quality and those with
favorable channel quality but a higher long term service rate.
Both these groups will get starved, since their instantaneous
channel rate over their long term service rate ratio will be
lower than the UEs getting repeatedly admitted. This let us
conclude that even the PF paradigm has a starvation problem,
when the dynamic nature of cellular networks is taken into
account.
REFERENCES
[1] "LTE," 3GPP. Available from http://www.3gpp.org/LTE; accessed 14
October 2010.
[2] ITU World Telecommunication/ICT Indicators Database. Available
from http://www.itu.int/ITU-D/ict/statistics/; accessed 27 February
2011.
[3] M. Rinne, and O. Tirkkonen, LTE, the Radio Technology Path toward
4G, in Computer Communication, Volume 33, Issue 16, pp. 1894
1906, October 2010.
[4] TR 23.882, System Architecture Evolution (SAE), 3GPP.
[5] Long Term Evolution (LTE), Overview of LTE Air Interface,
Motorola, Technical White Papers, 2007.
[6] A.J. Paulraj, D.A. Gore, R.U. Nabar, and H. Bolcskei, An Overview of
MIMO Communications Key to Gigabit Wireless, in Proceedings of
the IEEE, Volume 92, Issue 2, pp. 198128, February 2004.
[7] M. Rumney, "3GPP LTE: Introducing Single-Carrier FDMA," in
Agilent Measurement Journal, Issue 4, Second Quarter 2008, pp. 18--27,
2008.
[8] TS 36.321, E-UTRAN, MAC protocol specification, 3GPP.
[9] TS 23.203, RAN; E-UTRA; E-UTRAN; Overall Description, Stage 2,
3GPP.
[10] J. Lim, H.G. Myung, K. Oh, and D.J. Goodman, Channel-Dependent
Scheduling of Uplink Single Carrier FDMA Systems, in Vehicular
Technology Conference, 2006, VTC Fall 2006, IEEE, Montreal, QC,
Canada, September 2006.
[11] J. Lim, H.G. Myung, and D.J. Goodman, Single Carrier FDMA for
Uplink Wireless Transmission, in IEEE Vehicular Technology
Magazine, Volume 1, Issue 3, pp 3039, 2007.
[12] L. R. de Temino, G. Berardinelli, S. Frattasi, and S. Mogensen,
"Channel-aware scheduling algorithms for SC-FDMA in LTE uplink,"
in Personal, Indoor and Mobile Radio Communications, 2008. IEEE
19th International Symposium, Cannes, France, December 2008.
[13] F. Calabrese, P. Michaelsen, C. Rosa, M. Anas, C. Castellanos, D. Villa,
K. Pedersen, and P. Mogensen, "Search-tree based uplink channel aware
packet scheduling for UTRAN LTE," in Vehicular Technology
Conference, 2008, VTC Spring 2008, IEEE, Singapore, pp. 1949--1953,
May 2008.
[14] L. Suk-Bok, I. Pefkianakis, A. Meyerson, X. Shugong, and L. Songwu,
Proportional Fair Frequency-Domain Packet Scheduling for 3GPP LTE
Uplink, in IEEE Proceedings of the 28th Conference on Information
Communications (INFOCOM) 2009, Rio De Janeiro, Brazil, pp. 2611
2616, June 2009.