You are on page 1of 5

2015 24th Wireless and Optical Communication Conference (WOCC)

An On-Demand QoE Resource Allocation Algorithm


for Multi-flow LTE eMBMS

Hsiang-Yun Meng1 , Ching-Chun Chou2 , Rafael Kaliski3 , Hung-Yu Wei4


Department of Electrical Engineering
National Taiwan University
r02921057@ntu.edu.tw1 , f95921098@ntu.edu.tw2 , rkaliski@ieee.org3 , hywei@ntu.edu.tw4

1 Resource Block 1 Symbol


Abstract—Recent multimedia studies show that the Quality- Frequency
of-Experience(QoE) more accurately represents the user’s level of
satisfaction than the Quality-of-Service (QoS) does. Long Term Totally
Evolution (LTE) Multimedia Broadcast Multicast Operation 12 subcarriers
(MBMS), via MBMS Operation On-Demand (MooD), enables 12 subcarriers
dynamic resource configuration of multicast flows. In order to = 1 carrier
maximize the QoE of all users in a LTE MooD system we (180kHz)

propose a resource allocation method which efficiently allocates


resources based on both the demand of each live video stream 2 slots = 1 subframe(1 millisecond)
Time
and the channel conditions of the users. We compare our method
against other methods; our method achieves the highest QoE and Fig. 1. LTE time-frequency radio resource
demonstrates efficient resource allocation regardless of whenever
resources are sufficient or not.

allocation. In our proposed algorithm, the RBs are dynamically


I. I NTRODUCTION AND R ELATED W ORKS allocated to different video flows based on the system objective
Recent multimedia distribution schemes emphasize end- of reducing the buffering ratio and increasing the average bit-
users’ Quality of Experience (QoE) metrics over traditional rate. Our algorithm maximizes the QoE of the LTE MooD
performance metrics, such as throughput or error rate. Due to multimedia services.
the time varying wireless channel conditions experienced by
end-users, several works analyze the varying channel condi-
tions problem in terms of both multicast and broadcast services II. P ROBLEM F ORMULATION
[1]. In terms of QoE related research, Dobrian et al. [2] utilized In this section we first briefly introduce our objective func-
data-mining to analyze the relationship between video quality tion, the QoE utility, then we present the OTT live streaming
and user engagement. Their research found that a video’s resource allocation problem.
bitrate and its buffering ratio dominates a system’s QoE, as
captured by their user engagement metric. Based on Dobrian’s
results [2], [3] and [4] derived QoE functions for both wired A. Features of QoE Utilities
and wireless multicast systems.
The most important parameters in the user engagement-
3GPP (3rd Generation Partnership Project) recently decided oriented QoE utility function [2] are the buffering ratio (buffer-
that LTE MBMS (Multimedia Broadcast / Multicast Service) ing ratio is the percentage of time spent re-buffering a video)
Operation On-Demand (MooD) [5] should support Over- and the average bitrate. The higher the buffering ratio is, the
the-Top (OTT) multimedia service. MooD enables on-the-fly less likely the viewers are to watch the video for an extended
MBMS service configuration and seamless service migration. period of time. The higher the average video bitrate is, the
For example, when it becomes more efficient to run a unicast more likely the viewers are to watch the video for a longer
service as a MBMS service, the system may activate a previ- period of time.
ously inactive MBMS session for the service. As such, future
LTE MBMS services may be dynamic and configured based The utility in [3], shown in equation (1), is a linear function
on each user’s requirements and/or the system’s preferences. based on buffering ratio and average bitrate. The utility pre-
sented in [4], shown in equation (2), is an exponential utility
To enable active dynamic MBMS configuration, LTE re- function based off the same data from [2]. The linear utility
source allocation is further studied in this paper. As shown function, while simpler, becomes distorted when the buffering
in figure 1, Resource Blocks (RBs) have a duration of 0.5 ratio exceeds approximately 10%, this is due to the limitations
milliseconds and a width of 180kHz; RBs are the basic of linear-regression. In contrast to the linear utility function,
resource allocation unit used for both unicast and multicast the exponential utility function more accurately represents the
services. In this paper we investigate multiple methods of RB overall user engagement curve.
allocation based on each users’ demand and their respective
channel conditions.
To enhance the QoE of LTE MooD multimedia services, Bitrate
Ulinear = −3.7 × Buf f Ratio + (1)
we propose a QoE-based algorithm for LTE MBMS resource 20
4 corresponding author Uexp = V ideoLength × RBitrate × RBuf f Ratio (2)

978-1-4799-8854-9/15/$31.00 ©2015 IEEE 93


2015 24th Wireless and Optical Communication Conference (WOCC)

allocated to which group, i. Our method maximizes the system


utility, which results in an optimal user QoE. All notations used
eNB in this paper are contained in table I.
Symbol Description

Variables
UE1 UE2 UE3 UE4 UE5 UE6
Xji the j-th channel is allocated to video group [i]
X the decision variable at each round, an array of {Xji }
S [i] trunk size for video group [i] at one moment
[i]
mj the group MCS of group [i] in j-th channel , define in Eq.(4)
Group 1 Group 2 Group 3 [i]
mj,k the MCS of user-k of video group [i] in j-th channel
[i]
Fig. 2. Example of 3 Multicast Video Group bF the packet size of the current frame of video group[i]
now
V [i]

Input
s set of members of video group [i]
V number of video group
N [i] the population of video group [i], N [i] = |V [i] |
where RBitrate and RBuf f Ratio are: NRB number of carriers(channels) / resource blocks (RB) in this system
[i]
RBitrate = 1 − e−0.0001024×AvgBitrate Nbuf f number of buffering events of group [i]
Bits[i] number of total bits which has been transmitted for group [i]
RBuf f Ratio = e−0.04606×Buf f Ratio×100 T a given period that an allocation will be used for
U [i] utility of the video group [i]

Functions
U utility of the system (i.e. utility of this eNB)
In Eq.(1), a buffering ratio increase of 1% results in a
M (m) a mapping function which maps MCS m to Bitrate per RB
decrease of user engagement by 3.7 minutes; while an average g(n) a weighted function to weight a group with n users
bitrate increase of 20kbps results in the user engagement wre coefficient of Re-Buffering term in utility, wre = −3.7

Const
increasing by 1 minute. Similarly, equation (2) shows that an wkb coefficient of Avg.Bitrate term in utility, wkb = 0.05
increase in the buffering ratio or a decrease in the average wb the same as wkb while the unit is in bit
bitrate results in an exponential decrease of the utility. Notice TABLE I. TABLE OF NOTATIONS
that the buffering term dominates the utility function as [2]
showed that for a live video the buffering ratio is the most
significant factor in determining user engagement, while video We need to determine the bandwidth based on the trunk
bitrate is the second most significant factor in determining user size, S [i] , allocated to each group, [i], in order to determine
engagement . each group’s corresponding average bitrate and buffering ratio.
Given a channel allocation, Xji , the trunk size allocated to a
video is defined per equation (3).
B. Problem Formulation
In this work we consider a single cell LTE eMBMS sce- N RB
[i]
X
nario. In this scenario, which is based on [6], the surrounding S [i] = Xji × M (mj ) × T (3)
cells act as static interference sources with frequency flat j=1
distributions.
In our problem formulation there are multiple user equip- Where the selected modulation coding scheme (MCS) used
[i]
ments (UEs) in the cell. Each UE subscribes to a single video for channel-j, assigned to video [i], is mj .
stream. UEs which subscribe to the same video are classified as
belonging to the same video / multicast group. Figure 2 shows [i]
mj = min mj,k
[i]
(4)
an example of 3 separate multicast groups. Each multicast k∈Vi
group may have multiple channels on which it can receive
data. Due to the nature of multicast, all the members of a The MCS assigned to the group is dependent on the
[i]
group share the same coding scheme on any given channel / group member(s) with the worst channel conditions. M (mj )
carrier. represents a mapping function which maps each MCS to its
There are two critical factors used for resource allocation: corresponding per RB bitrate.
the each video packet’s size / video packet’s deadline and each In the next section we discuss our resource allocation
UE’s channel conditions. As all videos are streamed live, the method.
Evolved Node B (eNB) distributes the multimedia content on
the fly, i.e. the eNB is unable to obtain information such as III. P ROPOSED S CHEME
packet size until the live streaming data is generated. This is
a general feature of live streaming. We assume that the eNB In order to evaluate QoE metrics, we must know whether
can obtain the downlink quality of all carriers used by each a video frame is decodable or not. The typical unit used
UE. to evaluate the QoE metric, a video frame, exists at the
[i]
application-layer. Due to the limited capacity of the assigned
The packet size of a video i is denoted as bFnow . The RBs in a LTE subframe, a video frame may require multiple
channel conditions of the k th user of video i in channel j are subframes in order to be completely transmitted to a UE. As
[i] the capacity of a RB is dependent on the MCS and channel
denoted as mj,k . For convenience, all symbols with superscript
[i] indicate that they correspond to the video group [i]. conditions (channel conditions change over time), the eNB
must manage resources at the MAC-layer (RBs are assigned
We formulate the resource allocation problem as a multi- at the MAC layer). To perform QoE-based resource allocation,
round integer optimization problem. In each round the eNB the QoE must be evaluated at the MAC layer. Thus, we propose
runs our algorithm and determines a resource allocation for a single round QoE utility function, which operates at the MAC
each multicast group, which maximizes the system utility. layer, and an Integer Linear Programming Resource Allocation
The resulting allocation, Xji , indicates which channel, j, is algorithm to perform LTE subframe resource allocation.

94
2015 24th Wireless and Optical Communication Conference (WOCC)

A. Single Round QoE Utility Function In order to formulate our problem as an Integer Linear Pro-
gramming (ILP) problem, we apply the following 3 constraints
Based on the current video information and channel con- (The 3 constraints reflect the features of an LTE eMBMS
ditions, we predict the trend of the QoE and assign resources system.)
such that the predicted QoE is maximized.
V
In order to predict the QoE in each round we reformulate X
the linear QoE utility (1) as it requires less time to calculate. Xji = 1 (12)
For performance evaluation purposes we use the more accurate i=1
exponential QoE utility (2), which was presented in section IV. S[i]
Bits[i]
[i]
≤1− [i]
(13)
As the original version of the utility function does not bFnow bFnow
directly reflect the utility gain in terms of MAC-layer resource
allocation, we reformulate the linear QoE utility function (1) Xji ∈ {0, 1} , ∀i, ∀j (14)
below.
The constraints (12) and (14) limit each channel to a single
video at a time, these constraints are boolean variables. The
[i] 1
Nbuf f + (1 − u[i]
) constraint (13) places an upper-limit on the trunk size allo-
U [i] = wre × ×L cated to each video; the upper-limit prevents over-allocating
tnow + T resources for packets from the content provider as the over-
Bits[i] + S [i] allocated resources would go unused.
+wkb × (5)
(tnow + T ) Using the aforementioned constraints and the linear objec-
where tive equation (7), we use Linear Programming plus Branch
[i] and Bound to obtain the allocation pattern Xji in P-time. As
bFnow
u[i] = (6) a result our solution is able to operate fast enough to be used
S [i] for live video. We call our resource allocation method ILP
Resource Allocation.
The wre term determines the amount of a utility’s contri-
bution towards each slice of a video frame. The term 1/u[i] In each round, after obtaining the channel conditions and
captures the utility gain obtained by reducing the buffering the video packet information, we use the ILP Resource Allo-
ratio. The duration of a video frame is translated via the L cation algorithm to allocate resources. After multiple rounds,
term into the corresponding number of LTE slots; when the the QoE utility is used to determine how the eNB allocates
Frames Per Second (FPS) is 30, L is 2000. The term tnow resources such that the aggregate set of users’ QoE is max-
represents the time tick, which is referenced with respect to imized. The proposed algorithm is summarized in algorithm
the beginning of the live video transmission. 1.

B. Integer Linear Programming Resource Allocation IV. P ERFORMANCE E VALUATION


After reformulating the equation, we derive a linear system In this section we describe our simulator, simulation set-
utility U for video group [i] with respect to the resource tings, and discuss the results.
allocation Xji by multiplying by the group weight function
g (The group weight function g is with respect to each video
group.) A. Enviromment
The physical-level settings of our simulation are based on
U = max
X
g(N )U [i] [i]
(7) the LTE specifications [8] and [9]. The system-level settings
Xji are based on LTE simulator [10].
i
where The LTE simulator [10] doesn’t contain an eMBMS simu-
[i] lation. As such, we build our eMBMS simulator in MATLAB.
Nbuf f + 1 Bits[i] Using our simulator we compare the performance of our
U [i] = wre × × L + wb × proposed resource allocation method against 3 other resource
tnow + T (tnow + T )
PNRB [i] PNRB [i] allocation methods (The 3 other resource allocation methods
wre × L T j=1 Xji mj T j=1 Xji mj are Baseline, Throughput-Oriented, and Water-Filling). We
+ [i]
+ wb × (8) summarize the simulation settings in table II.
tnow + T b tnow + T
Fnow
We set our system bandwidth to 10MHz (per the LTE spec-
[i] ification [9] there are 50 effective channels when the system
where bFnow is the current video frame size in bits, which
will be transmitted next, and function g is a group weighted bandwidth is 10MHz.) We assume 40 of the 50 channels are
function which can be tuned by the network operator as per used for MBMS purposes, while the remaining channels are
[7]. In section IV our simulation applies the linear version of reserved for unicast or other applications.
g, per equation (11). Other multicast configuration settings, such as the system
population, are similar to those in [7]. The video used for our
simulation, Foreman [11], is in CIF format, has a duration of
Constant : g(N [i] ) = 1 (9) 300 frames, and a frame rate of 30 FPS. We encode the test
Logarithmic : g(N [i] ) = log(N [i] ) (10) video into a H.264/AVC video with a bitrate of 1.493 Mbps.
Our system capacity and the video rate is similar to AT&T’s
Linear : g(N [i] ) = N [i] (11) test scenario [12].

95
2015 24th Wireless and Optical Communication Conference (WOCC)

Algorithm 1 Single Round ILP Resource Allocation simulation.


Input: T : Allocation-Period • Throughput-Oriented (TO): Resources are allocated
1: t: Time with the objective of maximizing throughput. In terms
[i]
2: mj : Estimated Channel Condition of QoE, only the bitrate term is considered.
3: Ni : User in each video
4: RB = NRB : Number of Carriers • Water-Filling (WF): This method refers to the re-
[i] source allocation used in [13]. The demand for the
5: bFnow : Packet-Size of current video frame (remaining size)
video is divided by the weight (population) as the
6: Bits[i] : Current amount of Received Bits
priority. In order to make a fair comparison with our
7: Current Re-Buffering Times
method, we set the demand equal to the packet size
8: V : Number of Videos
divided by channel condition. In other words, a video
9: F P S: Frame Rate
with small packet sizes, good channel conditions,
Output: Channel Allocation Pattern, Xji or with a large group will have a higher priority.
10: Aeq =[ ];
Resources are assigned to the videos in the order of
11: L = 2000; // Slots per second
their respective priority.
12: for i = 1 to V do
13: Aeq = [Aeq INRB ×NRB ] //I is an identical matrix; • ILP Resource Allocation (ILP): Resources are allo-
14: end for cated according to the users’ channel conditions and
15: beq = 1NRB ×1 ; the video packet sizes. The resources are assigned such
16: for i = 1 to V do //bi is an 4 × 1 vector that the QoE utility, which is with respect to both the
[i] bitrate and the buffering ratio, is maximized.
17: bi = bFnow − Bits[i] ;
18: end for
19: for i = 1 to V do C. Simulation Result
[i]
20: Calculate Mapping M [i] (mj )
The performance metrics in our simulation are:
// i.e. from MCS to BitsPerSlot;
[i]
21: Aij = M [i] (mj ) × T ; • Utility: The expected user engagement per user is
 evaluated per equation (2) for more accuracy and to
wre L wb
22: Ci = T [i] × FPS + t+T ; help ensure that the distortion due to the linear QoE
(t+T )bF
now
[i] does not adversely effect the results.
23: fji = Ni × Ci × M [i] (mj );
24: end for • Buffering ratio: The number of Re-Buffering events
//Run divided over the entire video length.
Linear-Programming + Branch-and-Bound to solve X. • Bitrate: The average bitrate (kbps) per user.
(E.g. ILP)
//ILP maximize f s.t. constraints: In figures 3a-3c, the aforementioned 3 metrics all improve
AX ≤ b; Aeq = beq ; Xji ∈ {0, 1}; when the number of available channels increase. In all cases,
25: Xji = ILP (f, A, b, Aeq , beq , 0, 1); baseline has the worst metrics, while TO and WF have the
26: Return Xji ; second best set of metrics. WF metrics improve over TO
metrics as the number of available channels increase, this is
Parameter Setting
due to the fact that WF considers the demand of the video and
Symbols per slot 6 symbols (Extended Cyclic Prefix) the channel while TO does not. WF only assigns resources to
Number of Cells 1 Cell and see neighbor cell as interference the video with the highest priority, yet ILP always efficiently
Cellular layout Hexagonal grid, 3 sectors/cell assigns resources to all videos be determining the allocation
ISD 1732m [3gpp TR36814 Table A.2.1.1-1 3GPP case 3] which maximizes the QoE utility.
Central-frequency 2.0 GHz
System bandwidth 10 MHz We show the other scenarios, fixing the number of videos
Locations of UE Uniform distributed in each sector and number of channels yet varying the group size, in figures
BS Power 46 dBm 3d-3f. The 3 metrics all degrade when the population of each
Number of video streams User Defined (defaults = 4) group increases, as it becomes more likely that at least one user
LTE BLER upper-bound 10%
has poor channel conditions. To serve a user with poor channel
N0 (AWGN Noise) Thermal noise density = -174dBm/Hz
Path loss and SF UMa scenario in [3gpp TR36814 Table B.1.2.1-1]
conditions a slow MCS must be used, as determined per
CQI definition same as [3gpp TS36.213 Table 7.2.3-1] equation (4), i.e. the channel efficiency decreases. Regardless
Allocation Period T 1 slot (time unit for resource allocation in table I) of the group size, ILP always obtains the maximum QoE as it
TABLE II. TABLE OF S IMULATION PARAMETERS AND M ODEL allocates resources more efficiently.

V. C ONCLUSION
B. Algorithm for Comparison An on-demand resource allocation method is a necessity
due to the high volume of video traffic in LTE networks.
We compare our ILP resource allocation method against the A QoE-oriented resource allocation algorithm achieves higher
Baseline, Throughput-Oriented, and Water-Filling approaches. user satisfaction than a traditional throughput-oriented resource
All approaches are summarized below: allocation method.
• Baseline (BL): Allocate channels in the order of the In this paper we proposed a QoE-Based resource allocation
video group. The number of channels allocated to each method which efficiently allocates resources based on both the
group are proportional to the population of group. demand of video and the channel conditions. Our algorithm
The allocation is fixed for the entire duration of the maximizes the QoE utility over the aggregate set of all users.

96
2015 24th Wireless and Optical Communication Conference (WOCC)

Utility (VideoNum=4,GroupSize=4) ReBufferRatio (VideoNum=4,GroupSize=4) Avg.BitRate (VideoNum=4,GroupSize=4)

BL
TO 1200
7
0.6 WF
ILP
6 1000
0.5

AvgBitRate(kbps)
800
0.4

BufferRatio
Utility(sec)

4
600
0.3
3

0.2 400
2
BL BL
TO 0.1 200 TO
1
WF WF
ILP ILP
0 0 0
30 35 40 45 50 30 35 40 45 50 30 35 40 45 50
Channel num Channel num Channel num

(a) Utility v.s. ChannelNum (b) BufferRatio v.s. ChannelNum (c) Avg. Bitrate v.s. ChannelNum

Utility (VideoNum=4,CHNum=40) ReBufferRatio (VideoNum=4,CHNum=40) Avg.BitRate (VideoNum=4,CHNum=40)


9 1400
BL
TO 0.8
8 WF
1200
ILP 0.7
7
0.6 1000
6

AvgBitRate(kbps)
BufferRatio

0.5
Utility(sec)

5 800

0.4
4 600

3 0.3
400
2 0.2
BL BL
TO 200 TO
1 0.1 WF WF
ILP ILP
0 0 0
2 4 6 8 10 2 4 6 8 10 2 4 6 8 10
Population Per Group Population Per Group Population Per Group

(d) Utility v.s. Population per Group (e) BufferRatio v.s. Population per Group (f) Avg. Bitrate v.s. Population per Group
Fig. 3. Performance Metrics for Different Scenario

We built an LTE eMBMS simulator whose environment is [4] W. Kuo et al., “A QoE-Based Link Adaptation Scheme for H.264/SVC
based on the LTE specifications and whose system capacity Video Multicast Over IEEE 802.11,” IEEE Trans. Circuits Syst. Video
is set per AT&T’s settings. We evaluated the performance of Technol., vol. 25, no. 5, pp. pages 812–826, May 2015.
our ILP resource allocation method against 3 other resource [5] 3GPP, “3GPP TR 26.849 V12.0.0. MBMS operation on Demand;
MBMS Improvements, Technical Report,” 3rd Generation Partnership
allocation methods. Our ILP resource allocation method al- Project (3GPP), Tech. Rep., December 2014.
ways achieves the highest QoE utility regardless of whether [6] S. Lu et al., “Channel-Aware Frequency Domain Packet Scheduling
the resources are sufficient or not. for MBMS in LTE,” in Vehicular Technology Conf., 2009. VTC Spring
2009. IEEE 69th, April 2009, pp. 1–5.
[7] J. Chen et al., “Fair and Optimal Resource Allocation for LTE Multicast
ACKNOWLEDGMENT (eMBMS): Group Partitioning and Dynamics,” in 2015 Proc. IEEE
INFOCOM, April 2015.
Hung-Yu Wei was supported in part by the Ministry of [8] 3GPP, “3GPP TS 36.213 V12.5.0. Physical layer procedures; Evolved
Science and Technology of the Republic of China under Grants Universal Terrestrial Radio Access, Technical Specification,” 3rd Gen-
102-2221-E-002-077-MY and 103-2221-E-002-086-MY3. eration Partnership Project (3GPP), Tech. Rep., March 2015.
[9] ——, “3GPP TR 36.814 V9.0.0. Further advancements for E-UTRA
physical layer aspects; Evolved Universal Terrestrial Radio Access,
R EFERENCES Technical Report,” 3rd Generation Partnership Project (3GPP), Tech.
Rep., March 2010.
[1] R. Afolabi et al., “Multicast Scheduling and Resource Allocation [10] J. C. Ikuno et al., “System level simulation of LTE network,” in Proc.
Algorithms for OFDMA-Based Systems: A Survey,” IEEE Commun. 2010 IEEE 71st Vehicular Technology Conf., Taipei, Taiwan, May 2010.
Surveys Tuts., vol. 15, no. 1, pp. pages 240–254, First 2013.
[11] (2005) SVC test sequences. [Online]. Available: ftp://ftp.tnt.uni-
[2] F. Dobrian et al., “Understanding the Impact of Video Quality on hannover.de/pub/svc/testsequences/
User Engagement,” in Proc. of the ACM SIGCOMM 2011 Conf., ser.
SIGCOMM ’11. New York, NY, USA: ACM, 2011, pp. 362–373. [12] S. J. Crowley. (2011, April) The Challenge of HD Video Streaming
[Online]. Available: http://doi.acm.org/10.1145/2018436.2018478 on LTE. [Online]. Available: http://stevencrowley.com/2011/04/22/
streaming-hd-video-on-mobile-broadband/#more-1977
[3] X. Liu et al., “A Case for a Coordinated Internet Video
Control Plane,” in Proc. of the ACM SIGCOMM 2012 Conf. on [13] C. Ko et al., “Strategy-Proof Resource Allocation Mechanism for Multi-
Applications, Technologies, Architectures, and Protocols for Computer Flow Wireless Multicast,” IEEE Trans. Wireless Commun., pp. 1–14,
Communication, ser. SIGCOMM ’12. New York, NY, USA: ACM, 2015.
2012, pp. 359–370. [Online]. Available: http://doi.acm.org/10.1145/
2342356.2342431

97

You might also like