Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Uplink QoS Scheduling for LTE System

Sungoh Kwon Neung-Hyung Lee

School of Electrical Engineering Digital Media and Communications Division


University of Ulsan, Ulsan, Korea Samsung Electronics Co. LTD, Suwon, Korea
Email: sungoh@ulsan.ac.kr Email: nhlee.lee@samsung.com

Abstract— In this paper, we propose a QoS uplink scheduling 㩺㪇 㩺㪈 㩺㪉 㩺㪈㪏 㩺㪈㪐


algorithm for LTE collaborating with delay estimation. Un- 㩷㩷㫊㫃㫆㫋
like downlink scheduling, uplink scheduling cannot incorporate 㩷㩷㩷㪪㫌㪹㪄㪽㫉㪸㫄㪼
packet delay information due to specification constraints of LTE. 㩷㩷㪦㫅㪼㩷㫉㪸㪻㫀㫆㩷㪽㫉㪸㫄㪼㩷㪔㩷㪈㪇㫄㫊
The limited information results in difficulty ensuring a QoS
when conventional scheduling algorithms are employed. Using
a delay estimation tailored to LTE, we have shown that the Fig. 1. 3GPP LTE radio frame structure
uplink performance of the proposed QoS scheduling scheme can
be improved more than 11% even with a simple estimation via
simulations.
service flows which satisfy their QoS constraints. In LTE, all
I. I NTRODUCTION services have delay bounds to be met. For such delay-sensitive
services, assuming that all of the incoming packets have time-
Recent advanced technologies enable broadband wireless
stamps, schedulers monitor the head-of-line service delays
systems and the demand for various services increases rapidly
and accordingly allocate resources taking into account channel
in a mobile environment. At the same time, the service quality
conditions [2], [3], [4], [5]. However, the uplink scheduler for
of wireless systems is expected to be at the level of wired-
LTE is not able to be aware of packet time delays.
line networks, since applications over the Internet rapidly
An LTE uplink scheduler is located at an eNB and allocates
migrate to mobile environments. Many new applications and
resources based on reports from user entities (UEs). Since the
their mobile network demands result in continually evolving
reports do not provide time information for each packet, delay-
wireless networks.
aware scheduling algorithms proposed in the literature [2], [3],
In first and second generation wireless networks, voice
[4], [5] cannot be directly employed for LTE. This problem can
services were provided as circuit services. The third genera-
also occur in other systems, which use UE feedbacks lacking
tion (3G) network provides voice and data services on separate
of time information.
networks. In the 3G network, only the quality of service (QoS)
In this paper, we propose a QoS scheduling algorithm
of voice service is guaranteed by the radio access and core
tailored to LTE uplink scheduling. To that end, we estimate
networks, while the data services are provided in a best effort
packet delays based on feedback from UEs, and study the
method through the radio access network.
impact of delay uncertainty on QoS scheduling performance.
In contrast to the previous wireless networks, the next
generation wireless networks including the 3GPP (Generation The rest of this paper is organized as follows. In Section II,
Partnership Project) Long Term Evolution (LTE) [1] provide we describe the system model and state QoS scheduling
all services as packet services based on IP networks, including issues for LTE. In Section III, we develop a delay estimation
voice services. Hence, QoS of all services should be guaran- algorithm. In Section IV, we provide numerical results to study
teed in wireless networks including the radio access network. the efficacy of the scheme. We conclude in Section V.
To satisfy wide-range QoS constraints, schedulers are em-
II. S YSTEM M ODEL AND P ROBLEM
ployed in the network in order to guarantee QoS requirements
across service flows. Many scheduling algorithms have been In this section, we summarize the LTE system [6] and
studied to provide enhanced QoS performance for wireless describe uplink QoS problems during LTE QoS scheduling.
networks [2], [3], [4], [5].
The random access scheduler located at the MAC layer of A. LTE system
an eNB 1 schedules service packets according to the QoS
The 3GPP LTE FDD (Frequency Division Duplex) is frame-
profiles received from the upper layer. The QoS profile is
structued as shown in Fig. 1. One radio frame for downlink
composed of such information as priority, allowable delay, and
and uplink transmission is 10 ms long and consists of 20
minimum bit rate. The scheduler maximizes the number of
0.5-ms long slots. A subframe is defined as two consecutive
1A 3GPP LTE basestation is called an eNB slots (1 ms). Scheduling for uplink and downlink transmis-

Authorized licensed use limited to: KLE Technological University. Downloaded on March 22,2024 at 12:14:31 UTC from IEEE Xplore. Restrictions apply.
978-1-4244-8331-0/11/$26.00 ©2011 IEEE
First arrival Packet arrival BSR report
UE eNB

PDCP PDCP
UE i
t0 t1 tk-7 tk-6 tk-5 tk-4 tk-3 tk-2 tk-1 tk time
RLC RLC

BSR: Bis-3

BSR: Bis-2

BSR: Bis-1
(Padding)
(Periodic)

(Periodic)

(Periodic)
(Regular)
BSR: Bi0

BSR: Bis
MAC MAC

PHY PHY eNB


T0 Ts-3 Ts-2 Ts-1 Ts time
I3 I2 I2 I0

Fig. 2. 3GPP LTE protocol stack Delay estimation

Fig. 3. Example of packet arrivals and BSR


sions occurs at the eNB every subframe, referred to as the
Transmission Time Interval (TTI) 2 .
The protocol stack for the 3GPP LTE consists of the PHYsi- scheduler according to channel condition and QoS require-
cal (PHY), Medium Access Control (MAC), Radio Link Con- ments, while special service packets are persistently transmit-
trol (RLC), and Packet Data Convergence Protocol (PDCP) ted without QoS scheduling. Therefore, the QoS scheduler
layers, as depicted in Fig. 2. is necessary in order to adhere to the QoS requirements of
Scheduling for uplink and downlink transmissions is per- multiple services.
formed at the MAC layer according to such conditions as The QoS requirement of each service is defined by a QoS
priority, time delay, and channel condition. For downlink Class Identifier (QCI) and a minimum data rate, referred to as
scheduling, the scheduler timely gets information regarding the Guaranteed Bit Rate (GBR). When the service is initially
buffer status including time delay from a buffer manager connected, the network assigns one and only one QCI to each
located at the RLC or PDCP layer. service, while a minimum data rate is assigned only to GBR
For uplink scheduling, UEs buffer service packets in their type services. The QCI and minimum data rate can be also
storage and report buffer statuses to their serving eNBs changed by negotiation between the UE and the network in
via Buffer Status Report (BSR) messages periodically or service.
eventually [7]. Since multiple UE services can be provided, The assigned QCI specifies the service packet treatment in
each service has its own buffer, and each UE simultaneously terms of the following performance characteristics: resource
manages multiple buffers. type; priority; Packet Delay Budget (PDB); and Packet Error
The BSR provided by the 3GPP LTE consists of two fields: Loss Rate (PELR).
a two-bit logical channel ID 3 and its buffer size. The buffer LTE provides two resource types, GBR and non-GBR. GBR
size is the sum of the corresponding channel data stored in services request guaranteed minimum data rates, while non-
the corresponding buffer up to the BSR triggering time. Using GBR services have no bit rate requirement. A conversational
BSR, UEs support QoS-aware packet scheduling. voice service and an e-mail service are examples of the
The BSR is triggered in three cases. First, UEs transmit a GBR and the non-GBR types, respectively. Since applications
BSR to their serving eNB when uplink data in the RLC (or have different levels of allowable time delay and error, LTE
PDCP) entity become available for transmission or when standardizes multiple QoS classes as in Table I [8]. The QoS
uplink data of a higher priority arrive (Regular BSR). UEs class of an application service is determined by the chosen
also report their buffer status either when a predefined timer policy. For example, for live streaming video service, QCI 2
expires (Periodic BSR), or when the amount of padding is will be assigned for urgent or high priority UEs, otherwise
equal to or greater than the size of the BSR (Padding BSR). QCI 7 will be selected.
Figure 3 shows an example of packet arrivals at UE and The priority level in Table I defines relative importance
BSR. Service packets randomly arrive and are accumulated between different service aggregates. A PELR value specified
i
at a buffer. UE i reports BSR Bm to the serving eNB only for a standardized QCI wholely applies to the radio interface
when at least one of the BSR triggering conditions is satisfied. between a UE and an eNB so that the PELR is not a QoS
When BSR is triggered, UE i notifies the serving eNB of the scheduling parameter [8].
i
amount of packet data stored at the buffer at time Tm . The PDB in Table I is a soft upper bound for the allowable
B. Standardized QoS requirements in LTE packet delay time between the UE and the serving network.
Data packets in LTE are scheduled by a scheduler at the The PDB is interpreted as a maximum delay with a confidence
MAC layer; most are dynamically scheduled by the QoS level of 98 percent [8]. In other words, the QoS of the service
is considered to be satisfied when the ratio of service packets
2 LTE TDD (Time Division Duplex) scheduling is similar to LTE FDD that violates the PDB is less than 2 percent, i.e.,
scheduling except that either the uplink or the downlink but not both is
scheduled at every TTI. Pr{Packet delay ≥ PDB} ≤ 0.02. (1)
3 A UE can use up to four logical channels in LTE. If the number of services
is greater than four, then the services are categorized into four groups. Since the QoS constraint (1) is identical to that of the

Authorized licensed use limited to: KLE Technological University. Downloaded on March 22,2024 at 12:14:31 UTC from IEEE Xplore. Restrictions apply.
TABLE I
QCI TABLE THAT SPECIFIES THE SERVICE PACKET TREATMENT IN TERMS OF THE FOLLOWING PERFORMANCE CHARACTERISTICS : R ESOURCE TYPE ;
PRIORITY; PDB; AND PELR.

QCI Resource Type Priority PDB PELR Example Services


1 2 100ms 10−2 Conversational Voice
2 4 150ms 10−3 Conversational Video (Live Streaming)
GBR
3 3 50ms 10−3 Real-Time Gaming
4 5 300ms 10−6 Non-Conversational Video (Buffered Streaming)
5 1 100ms 10−6 IMS Signaling
6 6 300ms 10−6 Video (Buffered Streaming), TCP-based (e.g., www, e-mail, chat, ftp, p2p file sharing, etc)
Non-GBR
7 7 100ms 10−3 Voice, Video (Live Streaming), Interactive Gaming
8 8
300ms 10−6 Video (Buffered Streaming), TCP-based (e.g., www, e-mail, cat, ftp, p2p file sharing, etc)
9 9

M-LWDF, we briefly describe the M-LWDF algorithm and guaranteed. In particular, the delay-sensitive non-GBR services
explain the QoS scheduling problem that commonly arises in in Table I are susceptible to QoS failure.
the LTE system. One can use the BSR report time as packet arrival time.
Since the buffer size in BSR indicates a total length of the
C. M-LWDF and LTE QoS scheduling issues
packets buffered by the reporting time, the time delay of the
The M-LWDF algorithm is tailored to schedule multiple oldest packet is also renewed by every BSR so that the eNB
data users sharing a wireless channel. The constraint of M- will consider the packets as a new arrival. For example, in
LWDF is, for a given allowed violation probability δi , Fig. 3, let us assume that a scheduler schedules packets at
Pr{Wi ≥ τi } ≤ δi , (2) time Ts and that the packets arrived at time tk−7 are not
scheduled by time Ts and are buffered in the UE. Then, even
where Wi is the delay of service i and τi is a deadline to be if the packets that arrived at time tk−7 experience time delay
served. (Comparing between (1) and (2), we use τi and δi as Ts − tk−7 in the buffer, the QoS scheduler considers all of the
the PDB and the allowed violation rate of service i during packets reported at Ts as new arrival packets.
scheduling, respectively.) To overcome this problem, more sophisticated delay estima-
To that end, at every TTI the algorithm selects the best tion is necessary for uplink QoS scheduling. Hence, we study
service k such that an estimation method for LTE uplink QoS scheduling.
− log δi ri (t) III. Q O S ALGORITHM WITH DELAY ESTIMATION
k = arg max Wi (t), (3)
i∈Ω τi ri (t) We propose a delay estimation with BSR for uplink QoS
where Ω is the set of connected services, ri (t) is the available scheduling in this section.
data rate of service i at time t, ri (t) is the average transmission To estimate the time delay, we use a queue to store received
data rate, and Wi is the head-of-line delay of service i. The BSRs for each service. According to the depth of the queue,
average data rate ri (t) of service i can be internally computed we call the algorithm the M -step delay estimation with BSR.
by the scheduler based on the amount of the transmitted data. To describe our method, we assume that delay estimation
The available data rate ri (t) of service i can be estimated is performed at time Tsi when the sth BSR with Bsi arrives at
using Adaptive Modulation and Coding (AMC) depending on the eNB and define the parameters as follows:
the wireless condition. M the number of estimation steps
i
The main issue is how to procure time delay information Ts−m the time at which the m th previous BSR with respect
Wi (t) for QoS scheduling. For downlink scheduling, since a to Tsi is received for 0 ≤ m ≤ M
i
downlink scheduler and buffers for downlink service packets Bs−m the BSR value for service i at time Ts−m for 0 ≤
are located at the same eNB, the scheduler can obtain time m≤M
i i i
delay information for scheduling. In contrast to the downlink Im time interval between times Ts−m and Ts−m−1 for
case, buffers for uplink service packets are located at UEs, yet 0≤m≤M
the eNB schedules the stored uplink packets. Furthermore, as Ni the net amount of new arrival data of service i during
described in Section II-A, LTE does not provide time delay time interval I0i
i
information of individual packets for uplink. Rm the amount of residual data of service i in the mth
The BSR from a UE to the associated serving eNB provides queue for 0 ≤ m ≤ M
i i
only the amount of packet data stored in the UE at the BSR Dm the amount of scheduled data in Rm during time
i
triggering time. Since the BSR does not provide time infor- interval I0
i
mation, the serving eNB does not possess delay information Wˆm i the virtual delay of the remaining data Rm for 0 ≤
for uplink QoS scheduling, so QoS performance cannot be m ≤ M.

Authorized licensed use limited to: KLE Technological University. Downloaded on March 22,2024 at 12:14:31 UTC from IEEE Xplore. Restrictions apply.
The algorithm we propose is as follows. The scheduler Start

checks for the arrival of BSRs from UEs every TTI 4 . If there
are no BSR reports, the scheduler increases all of the estimated
delays Wˆm i s by one.
Received BSR?
No

When receiving a BSR of service i, the eNB determines


if the elapsed time from the previous BSR arrival time Yes

(BSR Arr T ) to the present time (Current T ) is greater


than a prefixed time interval, Delay Est Int. As the number Current_T - BSR_Arr_T > No
Delay_Est_Int?
of UEs increases, the number of BSRs also increases. Hence,
to control the computing load due to the excessive BSRs, Yes

we here define Delay Est Int as a control parameter. If the Parameter update
elapsed time (Current T − BSR Arr T ) from the previous (N i , I i m , Di m )

BSR arrival time is shorter than Delay Est Int, then the
eNB increases the estimated delays by one. Otherwise, the Packet classification
(Ri m )
eNB follows the delay estimation procedure outlined below.
First, the scheduler computes the parameters associated with
Delay estimation Delay update
service i; net arrival data N i , time interval Imi
and the amount (Wˆ i m ) (Wˆ i m Wˆ i m  1)
i
of scheduled data Dm . Since BSR implies only a total of the
buffered data, new arrival N i is the difference between the
End
amount of current data M Bsi and the amount of unscheduled
i i
previous data Bs−1 − m=0 Dm . The updated parameters are
expressed as Fig. 4. Flow chart of an N-step delay estimation with BSR
  M
 

N i = max Bsi − Bs−1 i
− Dm i
, 0 and
m=0
From (2) and (6), the LTE uplink QoS algorithm with delay
i
Im = i
Ts−m − i
Ts−m−1 for 0 ≤ m ≤ M. estimation selects the best service k at time t such that
  
Next, the scheduler classifies packets into M groups (Rmi
s W̃i (t) ri (t)
k = arg max f (pi ) ,
for m = 1 · · · M ) in the decreasing order of m based on the i∈Ω τi ri (t)
i i
stored data Rm s and the scheduling information Dm s, which
are defined as where f (pi ) is a priority function of service i, and τi is the
 i i PDB of service i. The priority function depends on the charge
i Rm−1 − Dm−1 if 1 ≤ m ≤ M − 1,
Rm = i i i i (4) and scheduling policy of the carrier. A simple example is a
RM + RM −1 − DM − DM −1 if m = M. reciprocal of pi so that the higher priority services have higher
The scheduler then updates R0i with N i . values than do the lower priority services.
After classifying the packets of Bsi into M + 1 groups, Even although we have proposed a QoS scheduling scheme
virtual delays are assigned to the corresponding groups and for uplink based on the LTE system, the scheme can be readily
are defined as extended to wireless systems that do not support individual
⎧ i
⎪ I packet information for uplink scheduling. In the next section,
⎨ 20 if m = 0,
I0i m−1 i we evaluate the algorithm using simulations.
Wˆm
i = + I if 1 ≤ m ≤ M − 1, (5)

⎩  2
m−1 i
k=0 k
k=0 Ik if m = M. IV. S IMULATIONS
In the delay estimation of (5), we set the time delay Ŵ0i of the For simulations, we set up two-tier cells composed of 19
initial group R0i , which recently received packets, to half of hexagonal cells. We assume that each cell has three sectors
the elapsed time I0i , since the average of n packet arrival times and that three UEs per sector are uniformly distributed in the
Ii coverage. Each UE has one non-GBR service with a PDB of
is 20 [9] assuming that the arrival process is Poisson [10]. The
150 ms. Since packets experience the average delay of 20 ms
longest time delay WˆM i is set to the sum of the elapsed time
i in the core network [8], the scheduling time budget at the eNB
delays. If Rm is zero, then we reset the corresponding delay
is calculated by subtracting the average delay from the given
Wˆmi to zero.
PDB. We set the traffic load to 1:2:3 for three UEs in a sector,
From the estimated delays, Wˆm i s, we define the head-of-line
use ITU Ped-A 30 Km/h as a channel model [10], and set the
delay W̃i (t) of service i at scheduling time t, defined as
bandwidth to 10MHz.
W̃i (t) = max Wˆm
i . (6) For QoS scheduling, each eNB devises the proposed
0≤m≤M
scheduling algorithm and schedules packets every TTI. All
4 In implementation, the scheduler can check BSR arrival with a certain simulations run for 10000 TTI (Ten seconds). Delay Est Int
frequency. For simplicity, we assume that the frequency is one TTI. is set to 40 ms and the periodic BSR timing is set to 20 ms.

Authorized licensed use limited to: KLE Technological University. Downloaded on March 22,2024 at 12:14:31 UTC from IEEE Xplore. Restrictions apply.
We consider two metrics for comparison, packet drop ratio 7
and cell throughput. For packet drop ratio, we count the
number of packets that violate the PDB and compute the 6.5

Cell throughput (Mbps)


ratio of the number of violated packets to the total number
of incoming packets. For cell throughput, we calculate the 6

total data rate at the PHY layer. We define the ideal case as
5.5
the case in which the buffer size and delay per packet are
perfectly known and use the performance of the ideal case as
5
an upper-bound to achieve in the simulations. Ideal
Figure 5 shows the packet drop ratios when the average Step2
4.5
traffic load per UE increases from 1.0 Mbps to 2.0 Mbps. Step1
Step0
Compared to scheduling without estimation (Step 0), the 4
1 1.2 1.4 1.6 1.8 2
scheduling algorithms with estimation (Step 1 and Step 2) Traffic load per UE (Mbps)
show better performances. When the average traffic load per
UE is 1.25 Mpbs, the packet drop rate improves from 2.0%
Fig. 6. Cell throughput versus traffic load
to 1.8% (11% improvement) in the case of Step 2.
Figure 6 shows the cell throughput when the traffic load
varies. Compared to the ideal case the other algorithms are V. C ONCLUSION AND F UTURE W ORK
slightly inferior, although all of the algorithms have similar
In this paper, we have proposed a QoS uplink scheduling
performance in terms of cell throughput. The degradation of
algorithm for LTE collaborating with delay estimation. Un-
the Step 2 case is less than 0.5% with respect to the ideal case.
like downlink scheduling, uplink scheduling does not possess
The data in Figs. 5 and 6 indicate that even simple estimation
packet delay information due to specification constraints.
can significantly improve UL QoS scheduling performance.
The limited information results in difficulty supporting QoS
without delay estimation. With delay estimation tailored to
12 LTE, we have shown that UL scheduling supports QoS. The
proposed scheme can also apply to wireless systems that use
10 imperfect feedback for uplink scheduling. Via simulations, we
Packet drop ratio [%]

have shown that our proposed algorithm with only a small


8
number of steps can achieve adequate performance.
Since system parameters such as BSR period, delay estima-
6
tion interval, and PDB can affect QoS scheduling performance,
we are analyzing the parameter impacts and studying optimal
4
Step0 parameters for UL QoS scheduling for further work.
Step1
2
Step2 R EFERENCES
Ideal [1] Motorola, “Long Term Evolution (LTE): A Technical Overview.”
0
1 1.2 1.4 1.6 1.8 2 [Online]. Available: www.motorola.com/lte
Traffic load per UE (Mbps) [2] H. Fattah and C. Leung, “An overview of scheduling algorithms in
wireless multimedia networks,” IEEE Wireless Communications, vol. 9,
(a) Packet drop ratio no. 5, pp. 76–83, 2002.
[3] W. Ajib and D. Haccoun, “An overview of scheduling algorithms in
mimo-based fourth-generation wireless systems,” IEEE Network, vol. 19,
Packet drop ratio normalized by Ideal

1.8
Step0 no. 5, pp. 43–48, 2005.
1.7 [4] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, and P. Whiting,
Step1
Step2 “Providing quality of service over a shrared wireless link,” IEEE
1.6
Ideal Communications Magazine, vol. 39, pp. 150–154, 2001.
1.5
[5] S. Shakkottai and A. Stoylar, “Scheduling algorithms for a mixture of
real-time and non-real-time data in hdr,” in 17th Int. Teletraffic Congress,
1.4 2001.
[6] 3GPP, TS 36.300 Evolved Universal Terrestrial Radio Access (E-
1.3 UTRA)and Evolved Universal Terrestrial Radio Access Network(E-
UTRAN); Overall description;, 2009.
1.2 [7] ——, TS 36.321 Evolved Universal Terrestrial Radio Access (E-
UTRA);Medium Access Control (MAC) protocol specification, 2009.
1.1 [8] ——, TS 23.203 Policy and charging control architecture, 2008.
[9] S. M. Ross, Stochastic process, 2nd ed. John Wiley and Sons, 1996.
1
1 1.2 1.4 1.6 1.8 2 [10] Recommendation ITU-R M.1225, Guidelines for the evaluation of radio
Traffic load per UE (Mbps) transmission technologies for IMT-2000, 1997.

(b) Packet drop ratios normalized by the ideal case

Fig. 5. Packet drop ratio versus traffic load

Authorized licensed use limited to: KLE Technological University. Downloaded on March 22,2024 at 12:14:31 UTC from IEEE Xplore. Restrictions apply.

You might also like