The Impact of QoS Support on the End User Satisfaction in LTE Networks with Mixed Traffic

miststizzaMobile - Wireless

Dec 10, 2013 (3 years and 10 months ago)

78 views

The Impact of QoS Support on the End User
Satisfaction in LTE Networks with Mixed Traffic
Iana Siomina and Stefan W
¨
anstedt
Ericsson Research,Sweden,E-mail:{iana.siomina,stefan.wanstedt}@ericsson.com
Abstract—How good is a best effort network for different
services in multi-service LTE networks?In this paper,we try to
answer the question in the context of capacity of an LTE network
with users concurrently running multiple services.We present a
system simulation study for two-service scenarios with VoIP and
a second service represented by real-time video,mobile TV,or
web surfing.We present results for a best effort network and
compare to those for a network with QoS provisioning.The study
demonstrates that traffic differentiation and service prioritization
are particularly crucial when a delay-critical service,e.g.,VoIP,
is in combination with a delay-insensitive intensive traffic.By
prioritizing VoIP,we achieve VoIP capacity comparable to that in
pure VoIP simulations at a cost of a few-percent capacity loss of the
second service.A simple model for network capacity assessment
for the two network types is presented to support our observations
theoretically.
Indexterms – LTE,QoS,capacity,downlink,traffic mix,VoIP,
video,TCP.
I.I
NTRODUCTION
Long-Term Evolution (LTE) [5] is an emerging radio access
network technology standardized in 3GPP [1] and evolving as
an evolution of Universal Mobile Telecommunications System
(UMTS).LTE is a converged all-IP network,and Quality of
Service (QoS) provisioning is crucial for providing a range
of IP-based services in the new generation networks.Hence
an evolved 3GPP QoS concept [10] has been developed.In
radio networks,QoS implies traffic differentiation and using
multiple bearers (a bearer is a point-to-point communication
service between two network elements) with configuration and
priorities optimized to ensure sufficient service quality for each
user.
Network-initiated bearer establishment and network-
controlled simplified QoS profiles based on QoS Class
Identifiers (QCIs) are among the key elements of the evolved
QoS concept.Both aim at ensuring consistent QoS and policies
across different User Equipment (UE) vendors and models as
well as in roaming scenarios.A QCI is a pointer to a pre-
configured set of node specific parameters [2],e.g.,scheduling
weights,admission thresholds,packet discard timer,etc.A
set of pre-defined QCIs is used to ensure class-based QoS.
Bearer-level QoS is further detailed by the Policy Charging
Enforcement Function assisted by a Subscriber Policy Register.
A wide range of QoS-related functions and parameters will
allow operators to adopt the level of QoS support in the
network aligned with their own strategy.Furthermore,initially
an operator may be primarily interested in achieving high peak
rates.There even exist some scenarios in which QoS may be
not needed,e.g.,low traffic load (a likely case during the initial
deployment phase) and mostly non-delay-critical data traffic
(an older network can,for example,be used to provide voice
service).Note,however,that it is not true that QoS is only
needed at high network loads,irrespective of the traffic pattern.
This is because services with some special QoS requirements
(e.g.,jitter,end-to-end delay,packet loss,buffering time,etc.)
have to compete with delay-insensitive intensive best effort
(BE) traffic,which may be difficult even at low loads when
the services are being run at the same UE.With BE service,
the network neither provides any delivery or service quality
guarantee nor prioritizes any service.This results in unspecified
variable user bitrate and delivery time,depending on the current
queue length,channel quality,and network load.
Our objective is to investigate the performance when deliv-
ering different services over a BE bearer in terms of capacity
of delay-critical services,such as Voice over Internet Protocol
(VoIP),in various mixed traffic scenarios.We present an LTE
performance simulation study with the focus on downlink
(DL) two-service scenarios with VoIP and a second service
represented by real-time video,mobile TV,or web surfing.
We compare per-service and combined performance in each
traffic mix in a BE network to that in a network with QoS
provisioning.We also present a simple model for network
capacity assessment for the two network types,to support our
observations theoretically.
Numerous theoretical and simulation-based studies have been
presented in the literature addressing various QoS aspects in
cellular networks,but far fewer specifically in LTE.Among
them,many disregard important protocol aspects and inter-
layer communication,assume unrealistic traffic models or
single-service scenarios.Recently,the authors of [7] presented
a bottom-up approach for modeling cumulative performance
degradation along protocol layers and predicting the perfor-
mance of different services in single-service scenarios.The
effects of QoS scheduling strategies on service performance in
a traffic mix consisting of VoIP,streaming video,and Session
Initiation Protocol (SIP) were studied in [14].In [4],the authors
presented a QoS-aware scheduling approach with an adaptive
VoIP priority mode which aims at decreasing the negative
impact of VoIP packets prioritization on the overall system
throughput.Mixed traffic scheduling has also been studied for
earlier UMTS networks,e.g.,for HSDPA [6],[11].
II.LTE R
ADIO
A
CCESS
In the following we describe the data flow through Layer 2
and Layer 1 delivered on one or multiple bearers in the form
of IP packets from the network (i.e.,IP) layer and transmitted
in DL.Layer 2 has three sub-layers,PDCP,RLC,and MAC.
Robust Header Compression (ROHC) is optionally performed
at the PDCP sub-layer.The data unit is then ciphered.The
PDCP header thus carries the required information for ROHC,
decompression,and deciphering.From PDCP,the protocol data
unit (PDU) is delivered to RLC which performs concatenation
of several PDUs from the same bearer or segmentation if
needed and adds an RLC header.In MAC,PDUs from several
bearers can be multiplexed,i.e.,form one MAC PDU.RLC
concatenation and segmentation as well as MAC multiplexing
are affected by scheduling decisions,i.e.,transport blocks are
formed according to the amount of available resources.
Finally,transport blocks are delivered fromMAC to the phys-
ical layer,where a Cyclic Redundancy Check (CRC) is added.
The transmission time interval (TTI),or the transmission time
978-1-4244-2644-7/08/$25.00 © 2008 IEEE
of a single transport block,in LTE is one sub-frame of 1 ms.
The amount of data transmitted in one transport block depends
on protocol and signalling overhead,selected modulation and
coding scheme,and the amount of allocated resources,i.e.,
power and resource blocks
1
(RBs).If not properly decoded from
the first attempt,multiple transmissions of the same transport
block can be combined at the receiver.Orthogonal Frequency
Division Multiplexing (OFDM) and Single Carrier Frequency
Division Multiple Access (SC-FDMA) have been adopted as
the transmission schemes for DL and uplink (UL),respectively.
III.A
N
A
PPROACH FOR
T
HEORETICAL
C
APACITY
E
STIMATION IN
M
IXED
T
RAFFIC
S
CENARIOS
Next we present an approach for capacity assessment in a BE
network with mixed traffic and a single queue where packets
are placed in their arrival order.The approach is extended in
Section III-B for a case with a separate queue for each service,
where different services have different priorities.
A.Without Service Prioritization
Assume a user i running two services concurrently during
time interval T.The two services have average packet transmis-
sion rates of f
1
and f
2
,respectively,and average packet sizes of
s
1
and s
2
,respectively.Without loss of generality,we assume
that f
1
≥ f
2
.All packets become available for the eNodeB DL
scheduler in the same order and at the same rate as they were
generated.The total amount of data available for transmission
during time T is thus (s
1
f
1
+s
2
f
2
)T.
The amount of data that can be transmitted to user i over
a radio link in one TTI depends on the scheduling decision
and the selected transport format,both of which depend on
available resources and channel quality.Let s
max
i
be the average
amount of data that can be transmitted to user i in a single
transport block.Given for user i a set of assigned RBs,the
average signal-to interference-plus-noise (SINR) ratio (may vary
by RB),and the target block error rate (BLER),s
max
i
can be
found by mapping the SINR into the effective bitrate obtained
from link-level simulations and subtracting the physical layer
overhead after that.Note that the effective bitrate should also
take into account retransmissions.
We say that a VoIP user is satisfied with the service quality
if there is no delay accumulated due to insufficient scheduling
capacity of the cell.We therefore require the ratio between the
data available for transmission and the scheduled data to not
exceed
1
1−ε
,where ε is the maximum ratio of delayed and lost
packets with which the VoIP service quality perceived by a user
is still satisfactory.
The number of scheduled users per TTI is limited by the
number of available RBs in the frequency band (e.g.,25 in
the 5 MHz band) and the total available power.There may be
some other limitations,like the number of available control
channels,which impose additional constraints on the network
capacity.Furthermore,all users in a cell share the available
resources in time and the amount of assigned resources to each
user is determined by a scheduler.With a classical round robin
(RR) scheduler,active users get equal time shares and in the
order defined by their previous transmission time,i.e.,the most
recently scheduled user has the least chance to be scheduled
in the next TTI.Let N be the number of users in the cell and
n denote the number of users scheduled at every TTI (or the
1
A resource block is a two-dimensional unit spanning over a predefined
number of consecutive sub-carriers and a time slot of 0.5s,i.e.,half sub-frame.
number of control channels).Then,the maximum amount of
data that can be transmitted to user i during time T is given by
T +d
max
￿
N
n
￿
· Δ
· s
max
i
,(1)
where d
max
is the maximum scheduling delay,and Δ is the
TTI length.The user satisfaction requirement thus reads
(s
1
f
1
+s
2
f
2
) · T ·
￿
N
n
￿
· Δ
(T +d
max
) · s
max
i

1
1 −ε
,(2)
or
(f
1
+αf
2
) · T ·
￿
N
n
￿
· Δ
(T +d
max
) · β

1
1 −ε
,(3)
where α =
s
2
s
1
and β =
s
max
i
s
1
.Recall that (3) is defined for
a single user (user i),i.e.,N is the maximum cell load that
the VoIP satisfaction criteria admits for user i.VoIP capacity of
the cell can be defined as the maximum cell load (the number
of users in the cell) under which at least 95 % of VoIP users
are satisfied.In a scenario where all users run the same traffic
mix the capacity is defined as the number of users per cell that
satisfies (3) for the fifth percentile of the s
max
-values of all
users in the cell.When the number of allocated RBs is the same
for all users in a TTI,the fifth percentile of the s
max
-values
corresponds to the fifth percentile of the SINR values.
Figure 1 shows the relation between α,β,and N when (3)
is a strict equality and the other parameters are set as follows:
f
1
= 50fps,f
2
= 15fps,ε = 0.02,d
max
= 0.23s,n = 4,
Δ = 0.001s,and T = 30s.For a known traffic mix,we can
also compute α,e.g.,α ≈ 30 in a scenario with 12.2 kbps
VoIP and 110 kbps video.With this setting,the capacity of
15 users per cell can only be achieved if the fifth percentile of
the β-values is at least 1.95,i.e.,the “worst” user out of 95 %
best should be able to transmit at every occasion the amount
of data equal to 1.95 VoIP packets in average.As previously
mentioned,β depends on SINR and to some extent,i.e.,within
a TTI,can also be controlled by the scheduler.Observe from
Figure 1 that the absolute capacity gain of such control is much
larger for small α,i.e.,when the traffic loads of both services
do not differ too much.Another observation is that a large
difference in the amount of generated traffic among the services
has a strongly negative impact on the capacity of delay-critical
services,e.g.,even though the model does not take into account
TCP retransmissions,we can expect that TCP traffic,especially
coming from a fixed network,can easily block VoIP packets
and thus result in low VoIP capacity.
0
10
20
30
0
0.5
1
1.5
2
2.5
0
50
100
150
200
β
α
Number of users per cell
Fig.1.Maximum cell load for a satisfied VoIP user.
B.With Service Prioritization
The model presented in Section III-A assumes a single-queue
scenario without traffic differentiation.If both traffic types have
the same delay budget,then ideally they have also the same
capacity (in reality,there may be some difference due to,for
example,adaptive transmission rate).With multiple bearers it is,
however,possible to differentiate among different traffic types
and apply different scheduling priorities.Adopting model (3),
we assume that if the first service has a relative priority of ρ > 1
over the second then we can disregard
1
ρ
of the data traffic from
the second service when computing capacity of the first one,i.e.,
the amount of traffic to be scheduled is f
1
+
α
ρ
f
2
.This means we
increase capacity of the first service at a cost of some capacity
loss of the second service.Note that ρ →+∞models a special
case corresponding to absolute priority of the first service.When
computing the capacity of the second service,the amount of
traffic to be scheduled is ρf
1
+αf
2
,where ρ ≤ T.
The relative capacity gain of the first service and the relative
capacity loss of the second service compared to the BE case
can be computed respectively as
f
1
+αf
2
f
1
+
α
ρ
f
2
and
f
1
+αf
2
ρf
1
+αf
2
.
Observe that even with small ρ the gain of the first service
capacity can be much larger than the capacity loss of the second
service.In our VoIP+video example with β = 1.95 for the fifth
percentile,VoIP priority of ρ = 2.25 allows us to double VoIP
capacity at a cost of 11 % video capacity loss.
IV.A S
IMULATION
S
TUDY
In this section,we present a DL network study conducted
by means of simulations for three traffic mixes and compare
the observations with our theoretical findings.An advanced
dynamic system-level simulator with detailed protocol imple-
mentation has been used to conduct the study for a network with
a regular hexagonal layout.The simulator adopts a wrap-around
technique to resolve the interference problemin border cells and
ensure uniform cell load for homogeneously distributed traffic.
Some important parameters are detailed in Table I.
A.Scenarios and Scheduling Description
Two scenarios have been considered.In the first scenario,
all user traffic goes into the same queue and is further not
distinguished by the scheduler.In the second scenario,two
queues with different priorities are used for user traffic.The
queue with a higher priority is intended for VoIP traffic,whilst
user traffic generated by the second service goes into the second
queue.The scheduling algorithm in both scenarios is RR,i.e.,
all users get the same time share for their transmissions and the
user with the highest priority is scheduled first.The service-
specific priorities in the second scenario affect the final user
priority in the way that the maximum of the service priorities
among active contexts is added to the user priority.As a result,
a user with VoIP data in the queue has a higher priority than the
one having only data from the second service,and a user with
VoIP data only has the same priority as a user under the same
conditions with VoIP data and the other data.In both scenarios,
a user with a pending retransmission has the highest priority,
which is a fixed value not depending on the transmitted data.
The maximum number of scheduled users per TTI is limited by
the number of control channels.
Furthermore,we experiment with two bearer configurations,
tcp-optimized and voip-optimized.The latter is only used for
VoIP and video bearers in the second scenario.Voip-optimized
bearers imply RLC unacknowledged mode (RLC UM),i.e.,
disable ARQ retransmissions,and allow for maximum 10 and
8 HARQtransmission attempts for DL and UL,respectively.For
tcp-optimized bearers,the acknowledged RLC mode (RLC AM)
TABLE I
N
ETWORK CONFIGURATION AND PARAMETER SETTING
Bandwidth 5MHz
Inter-site distance 500m
Number of sites/cells per site 7/3
Antenna type Directional,SIMO
Number of RBs in the band 25
Maximum DL power 20W
Simulation time interval 30s
DL assignment channels (PDCCH) 4
Channel quality report interval 0.1s
Channel quality estimation error Log.normal,3dB std.dev.
ROHC Yes
User mobility 3km/h in average
VoIP traffic
Voice codec AMR 12.2 kbps
Encoding/decoding delay 0.015s/0.005s
Voice activity factor 0.5
SID frames during inactivity period Yes
Maximum end-to-end delay 0.23s
Maximum delayed and lost frames 2%
Related RTP and RTCP Yes,on the same bearer
Real-time video traffic
Video codec H.264,CBR
Resolution QCIF (176 × 144)
Average bitrate/frame rate 110kbps/15fps
Encoding/decoding delay 0.04s/0.04s
Playout frequency 60Hz
Related RTP and RTCP Yes,on the same bearer
Mobile TV traffic
Video codec H.264,Baseline profile
Resolution CIF (352 × 288)
Average bitrate/frame rate 242kbps/25fps
Request message size 400Byte
Buffer length 1s
Maximum pre-buffering time 10s
Maximum total re-buffering time 10s
Playout frequency 60Hz
Related RTP and RTCP Yes,on the same bearer
Web surfing traffic
HTTP request size 400Byte
Web page size 100kByte
Number of web page downloads 1
is used in combination with at most 9 and 7 HARQ transmission
attempts for DL and UL,respectively.Also,we use a slightly
shorter T1-timer for triggering out-of-order delivery for voip-
optimized bearers (0.04 s) compared to tcp-optimized (0.06 s).
B.Traffic Models
The following three traffic mixes have been considered,

mix1:VoIP & real-time video,

mix2:VoIP & mobile TV,

mix3:VoIP & web surfing.
One traffic mix per user is assumed in every simulation and
it is the same for all users.All users remain in the system
throughout the simulation,i.e.30 s.The initial user distribution
is random uniform.A simple handover model is used to support
user mobility.Each session comprises two participants,one is
always located on the network side (behind the Internet).
Packets generated by VoIP,video,and mobile TV services are
encapsulated into Real-time Transport Protocol (RTP) packets
with 12 Byte headers.Each RTP stream is accompanied by
an RTP Control Protocol (RTCP) stream of control packets
(RFC3550).When delivered from the network layer,Robust
Header Compression (ROHC) is applied in PDCP in order to
reduce header overhead.For each header type,a predefined
ROHC profile is selected:ROHC-RTP (RFC3095),ROHC-UDP
(RFC3095),ROHC-TCP (RFC4996),or ROHC-IP (RFC3843).
1) VoIP:A VoIP session starts in the beginning of the simu-
lation with a random offset uniformly sampled from [0 s,0.2 s].
The conversation between two users is continuous and has an
activity factor of 0.5.The talk spurt lengths are drawn from
an exponential distribution.During silence periods (discontin-
uous transmission,DTX),Silence Descriptor (SID) frames are
transmitted periodically (every 160 ms).
2) Real-Time Video:Real-time (RT) video traffic is repre-
sented by full duplex conversational video with frame sizes read
from a realistic video trace.Each video session starts with a
random offset uniformly distributed in [0 s,0.3 s].
3) Mobile TV:Unlike with RT video when a client just
passively receives frames from a server,mobile TV is a
streaming application where the traffic is initiated by a client
request transmitted to the server which then starts transmitting
a mobile TV stream.The clients have a playout buffer.The
initial buffering time in mobile TV should be small to avoid
long delays when switching channels.
An MTV music clip trace has been used to generate the
mobile TV stream.Except for a larger bitrate,the traffic also
has a larger variation in frame sizes compared to RT video
because of I-frames inserted every 25th frame.A client initiates
a session in the beginning of the simulation with a randomoffset
in [0s,0.3 s] and buffers frames until the buffer is filled.Re-
buffering occurs when the buffer is empty.The session lasts
for 30 s,but it can be stopped when either the maximum pre-
buffering or the total re-buffering times is exceeded.
4) Web Surfing:Every user sends a single Hypertext Transfer
Protocol (HTTP) request initiated randomly within the first 25 s
of the simulation and downloads a single web page.Note that
unlike other discussed traffic types,this is TCP traffic.
C.Service Quality and Capacity Metrics
There is no standard definitions of user satisfaction for
multimedia applications[12].It is,however,common to measure
user satisfaction of,for example,a VoIP user by the number of
delayed and lost frames [3].For conversational video users,
using the same metric can still be justified by the requirement
of synchronization with the voice service.The approach we
adopted uses a model that sets a relationship between bitrate,
delay/loss rate,and subjective video quality [8],[9].Given the
minimum required Mean Opinion Score (MOS) and a video
bitrate,the model gives us the maximum allowed loss/delay
rate.For web users,throughput is the most critical metric.
For mobile TV,we adopt the Video Streaming Quality Index
(VSQI) model [13],from which the quality of a video stream
can be computed by
Q
final
= (1−Q
BufRed
)(1−Q
PlRed
)(1−Q
BrRed
)(Q
max
−1)+1,
where Q
BufRed
is the quality reduction factor due to buffering
effects,Q
PlRed
is the reduction factor due to packet loss,
Q
BrRed
is the quality reduction due to the total bitrate,and
Q
max
is the maximum quality for QCIF videos.The first
parameter is a polynomial function of the initial buffering
percentage relative to the clip length,re-buffering frequency,and
total re-buffering time.Q
PlRed
is a power function of packet
loss,and Q
BrRed
is an exponential function of the average
bitrate.All the fitted functions have been parameterized by
means of the root mean square error regression for different
codecs,including H.264 which is used on our simulations.
Below we detail the user satisfaction criteria for the four
service types.

A user is satisfied with the VoIP service if the experienced
ratio of lost and end-to-end delayed frames is at most 2 %.

A user is satisfied with the video service if the MOS-score
for the video session is at least 3.0.

A user is satisfied with mobile TV streaming if the MOS-
score given by the VSQI model is at least 3.0.

A user is satisfied with a TCP download if the experienced
throughput is at least 300 kbps.
The cell capacity is defined for each service separately as
the maximum number of users in the cell among which at least
95 % are satisfied with service in question.We also compute
combined capacity requiring each satisfied user to be satisfied
with all the services at the same time.Note that all presented
results are relevant for relative comparisons.
D.Simulation Results
Table II presents results for the single-queue (BE network)
and two-queue (QoS network) scenarios.For each scenario,
we show per-service capacity and the combined capacity for
each traffic mix.As expected,both services in each of the
first two mixes in the BE network have similar capacity —
the result of a single queue and comparable delay budgets.
The small differences are mainly due to variable packet sizes
and fast fading.Large I-frames in mix2 further degrade VoIP
performance.In mix3,VoIP heavily suffers from web surfing,
emphasizing the importance of traffic differentiation and QoS-
aware scheduling.Throughput-based satisfaction criterion for
web traffic explains the difference in per-service capacity in
mix3.
The results for the two-queue scenario show that by priori-
tizing VoIP traffic we can gain a lot in VoIP capacity at a cost
of very small capacity loss of the second service.The VoIP
capacity gain is 105 %,258 %,and 710 % for mix1,mix2,and
mix3,respectively,i.e.,it increases with traffic intensity of the
second service.The combined capacity increases in mix2 and
mix3,i.e.,only if it has been limited by VoIP in the BE network.
In Figure 2 we show user satisfaction as a function of the
average cell load in each traffic mix.Per-service user satisfaction
is depicted by solid and dashed lines,and the dotted lines denote
combined user satisfaction and the 95 % capacity threshold.
Per-service capacity is further compared to the maximum
achievable,i.e.,obtained in a single-service BE network (REF).
The reference results are shown in Table III.We note that video
and web traffic capacities in Table II are close to the reference
results.The gap is larger for mobile TV.VoIP capacity in the
QoS network is also much below the reference capacity.The
latter is mainly explained by the scheduler design.Recall that if
TABLE II
R
ELATIVE CELL CAPACITY IN
BE
AND
Q
O
S
NETWORKS
Traffic
mix
Service
BE network QoS network
Per-service Combined Per-service Combined
mix1 VoIP 14.56
14.25
152.85
14.07
Video 14.26 14.07
mix2 VoIP 6.36
6.34
164.29
6.53
Mobile TV 6.61 6.53
mix3 VoIP 2.05
2.05
145.66
6.6
Web surfing 6.5 6.6
TABLE III
R
EFERENCE CAPACITY IN SINGLE
-
SERVICE SCENARIOS IN A
BE
NETWORK
Service Relative capacity
Service Relative capacity
VoIP 231.0
Mobile TV 10.19
Video 16.8
Web download 7.43
11.5
12
12.5
13
13.5
14
14.5
15
15.5
0.88
0.9
0.92
0.94
0.96
0.98
1
Relative cell load
Satisfied users share


BE: VoIP
BE: Video
BE: VoIP & Video
QoS: VoIP
QoS: Video
QoS: VoIP & Video
(a) Mix1
4.5
5
5.5
6
6.5
7
7.5
0.8
0.82
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
Relative cell load
Satisfied users share


BE: VoIP
BE: Mobile TV
BE: VoIP & Mobile TV
QoS: VoIP
QoS: Mobile TV
QoS: VoIP & Mobile TV
(b) Mix2
1
2
3
4
5
6
7
8
0.5
0.6
0.7
0.8
0.9
1
Relative cell load
Satisfied users share


BE: VoIP
BE: Web
BE: VoIP & Web
QoS: VoIP
QoS: Web
QoS: VoIP & Web
(c) Mix3
Fig.2.Per-service and combined user satisfaction as a function of the average cell load.
1
3
5
7
9
11
13
0
250
500
750
1000
1250
1500
1750
2000
2250
Relative cell load
User throughput, [kbps]


BE
QoS
REF
mean
5th percentile
95th percentile
Fig.3.TCP throughput in BE and QoS networks vs.REF network.
there are no pending retransmissions,users with pending VoIP
packets have absolute priority,which does not mean though that
only VoIP traffic is scheduled.Also,resegmentation of transport
blocks that need to be retransmitted may result in that some
transport blocks do not carry VoIP packets and thus “steal”
resources from VoIP traffic.This is more likely to happen in
mix3 which results in the lowest VoIP capacity in the QoS
network.Furthermore,for the same network,VoIP capacity is
larger in mix2 than in mix1,which is explained by that most of
the mobile TV frames (except I-frames) are smaller than those
from video and cause less disturbance to prioritized VoIP.
For web traffic,we also present the average throughput and
the 5th and the 95th percentile throughput as functions of the
average cell load in the BE and QoS networks as well as in
the REF network.Not surprising,the BE network is the best
for users with good channel conditions.These users experience
the largest throughput degradation when VoIP is prioritized over
other traffic types.The loss ranges from zero to 15 % for some
loads.QoS reduces the average throughput by at most 10 %
at high loads.Users with bad channel conditions are almost
not affected by QoS,except for the lowest load when the QoS
network is noticeably better.In general,our observations on the
throughput degradation with prioritized VoIP are in line with our
conclusions in Section III-B.
We have also computed the fifth percentile of the user average
β-values in mix1 and found it comparable to that computed
in our example in Section III-A (1.76 vs.1.95 for 15 users
per cell).The ratio between the VoIP capacity gain and the
second service’s capacity loss is significantly larger than that
computed in Section III-B,even though the results are not
directly comparable because of scheduling.One reason for
the large capacity ratio is the conservative channel quality
measurements filtering and link adaptation that limit high-bitrate
traffic capacity in the interference-limited environment and also
make it less sensitive to prioritization of low-bitrate traffic.
Another reason can be seen from the low average number of
retransmissions,high power and RB utilization in some cells
and at the same time small average number of scheduled users
(below 1.25 in all scenarios and traffic mixes at the capacity
limit of the second service).
V.C
ONCLUSIONS
Service differentiation and prioritization of delay-critical traf-
fic are important not only at high loads but at any load when
multiple services are concurrently run at a user terminal which is
a likely scenario in LTE networks.This is especially important
when a delay-critical service,e.g.,VoIP,is in combination
with a delay-insensitive intensive traffic,web surfing or TCP
download.We have shown theoretically and by simulations that
prioritization of such a service as VoIP typically does not cause
large quality degradation of other services due to small VoIP
packet sizes but allows more efficient radio resource utilization.
Consequently,in the QoS scenario we could allow more (VoIP)
users in the system together with the mixed traffic users.
R
EFERENCES
[1] 3GPP TS 36-Series,http://www.3gpp.org/.
[2] 3GPP TS 23.107,Quality of Service (QoS) concept and architecture.
[3] 3GPP TR 26.975,Performance characterization of the Adaptive Multi-
Rate (AMR) speech codec.
[4] S.Choi,K.Jun,Y.Shin,S.Kang,B.Choi.MAC scheduling scheme
for VoIP traffic service in 3G LTE,in Proc.of the 66th IEEE Vehicular
Technology Conference (VTC2007-Fall),Sep.2007.
[5] E.Dahlman,S.Parkvall,J.Sk
¨
old,and P.Beming.3G Evolution:HSPA
and LTE for Mobile Broadband.Academic Press,2007.
[6] M.Ericson and S.W
¨
anstedt.Mixed traffic HSDPA scheduling —Impact
on VoIP capacity,in Proc.of the 65th IEEE Vehicular Technology
Conference (VTC2007-Spring),pp.1282–1286,Apr.2007.
[7] G.G
´
omez,J.Poncela Gonz
´
alez,M.C.Aguayo-Torres,J.F.Paris,J.T.
Entrambasaguas.QoS modeling for performance evaluation over evolved
3G networks,in Proc.of the 3rd ACM Workshop on QoS and Security
for Wireless and Mobile Networks (Q2SWinet),Oct.2007.
[8] J.Gustafsson.Study of multimedia streaming packet losses and inter-
ruptions,in preparations for a parametric opinion model for multimedia
streaming services,Contr.29 to ITU-T Study Gr.12 meeting,Jan.2007.
[9] F.Hultin.Congestion notification and rate-adaptation for real-time services
in all-IP radio networks,Master’s Thesis,Lule
˚
a Un.of Technology,2007.
[10] R.Ludwig,H.Ekstr
¨
om,P.Willars,and N.Lundin.An evolved 3GPP
QoS concept,in Proc.of the 63rd IEEE Vehicular Technology Conference
(VTC2006-Spring),pp.388–392,May 2006.
[11] K.I.Pedersen,T.E.Kolding,and P.Mogensen.QoS Considerations for
HSDPA and Performance Results for Different Services,in Proc.of 64th
IEEE Vehicular Technology Conference (VTC2006-Fall),Sep.2006.
[12] J.Saliba,A.Beresford,M.Ivanovich,and P.Fitzpatrick.User-perceived
quality of service in wireless data networks,Personal and Ubiquitous
Computing,9(6),Nov.2005.
[13] X.Tan,J.Gustafsson,G.Heikkil
¨
a.Perceived video streaming quality
under initial and rebuffering degradations,in Proc.of the 5th Intl.
Conference on Measurement of Audio and Video Quality in Networks
(MESAQIN 2006),May 2006.
[14] M.Wernersson,S.W
¨
anstedt,and P.Synnergren.Effects of QoS schedul-
ing strategies on performance of mixed services over LTE,in Proc.
of the 18th Intl.Symposium on Personal,Indoor and Mobile Radio
Communications (PIMRC ’07),Sep.2007.