Optimized Network-coded Scalable Video Multicasting over eMBMS Networks

1.7 MB·English

by  Andrea Tassi

Optimized Network-coded Scalable Video
Multicasting over eMBMS Networks
Andrea Tassi∗, Ioannis Chatzigeorgiou∗, Dejan Vukobratovic´‡ and Andrew L. Jones∗
∗School of Computing and Communications, Lancaster University, United Kingdom
‡Department of Power, Electronics and Communication Engineering, University of Novi Sad, Serbia
Abstract—Delivery of multicast video services over fourth to deliver the same PtM service in a synchronous fashion.
generation (4G) networks such as 3GPP Long Term Evolution- Hence, base stations of the same SFN do not interfere with
Advanced (LTE-A) is gaining momentum. In this paper, we
each other.
addresstheissueofefficientlymulticastinglayeredvideoservices
3GPP addresses the reliability issues of PtM communi-
by defining a novel resource allocation framework that aims to
maximize the service coverage whilst keeping the radio resource cations by proposing the adoption of Application Level-
footprint low. A key point in the proposed system mode is that ForwardErrorCorrection(AL-FEC)schemesbasedonRaptor
the reliability of multicast video services is ensured by means codes [4]. However, Magli et al. [5] noted that this kind of
of an Unequal Error Protection implementation of the Network
5 sparse codes, as well as LT codes, are usually designed to be
Coding (UEP-NC) scheme. In addition, both the communication
1 appliedoverlargesourcemessages.Asaconsequence,theuse
parameters and the UEP-NC scheme are jointly optimized by
0
the proposed resource allocation framework. Numerical results of Raptor or traditional LT codes can lead to a non-negligible
2
show that the proposed allocation framework can significantly communication delay. As noted in [5], this issue can be

n increase the service coverage when compared to a conventional mitigated by using Random Linear Network Coding (RLNC)
a Multi-rate Transmission (MrT) strategy.
strategies, applied over short source messages.
J
Index Terms—Unequal error protection, random linear net-
Thispaperdrawsinspirationfrom[6]and[7],whichutilize
0 work coding, multimedia communication, resource allocation.
2 the RLNC principle to protect the reliability of Point-to-
Point (PtP) and PtM communications, respectively. Unlike [6]
I. INTRODUCTION
] and [7], this paper takes into account the fact that the layers
T
Over the last few years, the technological evolution of of a video stream have different importance levels and hence
I
. communication devices has fuelled a surge in demand for we adopt the Unequal Error Protection (UEP) implementation
s
c new multimedia services over fourth generation (4G) and of RLNC [8], which we shall refer to as UEP-NC. UEP-NC
[ next-generation cellular networks. To give an example, up to allows the transmitter to adjust the error protection capability

67%oftheglobalmobileInternettrafficwillberepresentedby ofthecodeaccordingtotheimportancelevelofthetransmitted
2
v videocontentby2018[1].Variousstandardsforvideocoding video layer.
1 and compression have been proposed to enable layered video In addition to tackling reliability issues, the definition of
1 streaming. Among them, the H.264 Scalable Video Coding efficient resource allocation models suitable for delivering
3
(H.264/SVC) standard is gaining popularity [2]. PtM services is also a challenging task. Among candidate
3
The key idea behind H.264/SVC compression is the gener- resource allocation models, the family of Multi-rate Trans-
0
. ation of a layered video stream, which consists of one base mission (MrT) strategies is the most fitting for exploiting the
1
layerandmultipleenhancementlayers.Thebaselayerensures layered nature of a service [9]. MrT strategies attempt to split
0
5 a basic reconstruction quality, which progressively improves users into subgroups. The transmission parameters of each
1 with the number of recovered enhancement layers. Hence, a videolayerarethenoptimizedbyconsideringthepropagation
: scalable video service can be transmitted to multiple network conditions experienced in each user subgroup. In this way, all
v
i users and potentially be decoded by them at different quality subgroups receive the same video stream but each subgroup
X
levels, depending on the radio propagation conditions. is able to recover a different set of video layers and, thus,
r In this paper, we consider a 3GPP Long Term Evolution- experience a different level of video quality [10], [11].
a
Advanced (LTE-A) communication network multicasting AMrTresourceallocationmodelcanbedesignedtoachieve
H.264/SVC video services. Since its first release, the LTE-A a variety of objectives. For instance, [10], [11] maximize the
standard manages Point-to-Multipoint (PtM) services by total service level experienced by all the users. On the other
means of the evolved Multimedia Broadcast and Multicast hand, [7] concentrates on minimizing the amount of radio
Service(eMBMS)framework[3].APtMservicecanbedeliv- resources needed to broadcast a video service. In contrast
ered by using the Single-Cell (SC-) or the Single Frequency to[7],[10],[11],thispaperinterpretsthenumberofrecovered
Network (SFN-) eMBMS modes. In the SC-eMBMS mode, video layers as the system profit, while the amount of radio
the service is delivered by each base station independently resourcesrequiredtodeliverthePtMserviceisconsideredthe
of the others. On the other hand, the SFN-eMBMS mode system cost. According to a fundamental economics principle,
allows multiple neighbouring base stations (forming the SFN) we propose a novel MrT-based allocation model, which we
K
callUEPResourceAllocationModel(UEP-RAM).Ourmodel 3
K
maximisesthesystemprofit-costratioandensuresthatatleast 2
K
a desired fraction of users achieve a predetermined service 1
......
level. A key aspect in the proposed UEP-RAM is that system x x x
1 2 K
and UEP-NC parameters are jointly optimized.
The rest of the paper is organized as follows. Section II k1 k2 k3
presents the theoretical framework for the assessment of the
Fig.1. ExampleoflayeredsourcemessagexcomprisedofL=3layers.
service level achieved by each user. The integration of the
UEP-NC principle into the LTE-A protocol stack is also
In this paper, we map coded elements associated with an
presented. Section III describes the proposed UEP-RAM and
expanding window onto Packet Data Units (PDUs). The PDU
provides a novel heuristic strategy to efficiently derive a good
description and its role in the LTE-A protocol stack will be
qualitysolution.ResultsarediscussedinSectionIVwhilethe
clarifiedinSectionII-B.Forthesakeofsimplicity,weassume
main findings of the paper are summarized in Section V.
that coded elements associated with different expanding win-
II. SYSTEMMODEL dowsarenotmixedwithinthesamePDU.Theselectionofthe
coding coefficients for the generation of each coded element
We consider an eMBMS network that delivers the same
is governed by a Random Number Generator (RNG). As
target scalable video stream to a multicast group of U users.
explainedin[6],theRNGseedassociatedwiththefirstcoded
If the eMBMS network consists of a single base station,
element in a PDU is transmitted along with standard LTE-A
the video service is transmitted according to the SC-eMBMS
signalling information. The RNG seeds associated with the
mode. On the other hand, if the eMBMS network is formed
remaining coded elements in the PDU are then incrementally
by two or more spatially contiguous base stations, they will
computed from that associated with the first coded element.
form a SFN [3].
Let us focus on one of the N PDUs associated with the
(cid:96)
A. UnequalErrorProtectionRandomLinearNetworkCoding (cid:96)-th expanding window x . This PDU can hold n coded
1:(cid:96) (cid:96)
elements1, which are transmitted over a broadcast erasure
Wedenotebyx={x ,...,x }ascalablesourcemessage
1 K
channel using a Modulation and Coding Scheme (MCS) with
composed of K elements. As shown in Fig. 1, the elements
index m . The probability that user u will fail to recover the
of x can be grouped into L layers. Each layer signifies a (cid:96)
PDU is denoted by p . Note that the larger the value of m
different importance level and consists of a fixed number of u,(cid:96) (cid:96)
is, the higher the modulation order or the lower the error-
elements. For example, the (cid:96)-th layer consists of k elements.
(cid:96)
correcting capability of the corresponding MCS is. We can
We assume that layers are arranged in descending order of
infer that the value of m has an impact on both p and n ;
importance, that is, the first layer – which comprises the first (cid:96) u,(cid:96) (cid:96)
if m increases, then n and p are also likely to increase.
elements of the message – is the most important layer, while (cid:96) (cid:96) u,(cid:96)
Based on our adopted model, if N PDUs associated with
the L-th layer is the least important layer. We define the user (cid:96)
expanding window x are transmitted, r ≤N of them will
Quality-of-Service (QoS) level as the number of consecutive 1:(cid:96) (cid:96) (cid:96)
be successfully received by user u. We initially refer to the
layers that a user can recover, starting from the first layer.
case where source elements of expanding window x can be
In an effort to improve reliability, each source message is 1:(cid:96)
retrievedfromPDUsrelatedtothefirst(cid:96)expandingwindows.
transmitted to users according to the RLNC principle [12].
For convenience, we introduce sets N ={N ,...,N } and
A typical RLNC PtM scheme generates a stream of N ≥K 1:(cid:96) 1 (cid:96)
r ={r ,...,r } to enumerate all transmitted and received
coded elements y = {y ,...,y } obtained by linearly com- 1:(cid:96) 1 (cid:96)
1 N
PDUs,respectively,whichareassociatedwitheachofthefirst
bining elements of x. Hence, coded element y is defined
j
as y = (cid:80)K g x , where each coding coefficient g is (cid:96)expandingwindows.Theprobabilitythatuseruachievesthe
j i=1 j,i i j,i QoS level (cid:96) can be quantified by the probability P (N ) of
selected uniformly at random over a finite field GF(q) of size u 1:(cid:96)
recovering the source elements of the (cid:96)-th expanding window,
q. A user can recover the source message x as soon as it
expressed as
collects K linearly independent coded elements [12].
A key issue of conventional RLNC-based schemes is that a Ru(r1:(cid:96))
user is not able to recover all the K source elements if it has (cid:122) (cid:125)(cid:124) (cid:123)
not collected at least K linearly independent coded elements. P (N )=(cid:88)N1··(cid:88)N(cid:96) (cid:89)(cid:96) (cid:18)Ni(cid:19)(1−p )ripNi−rig(r ) (1)
Toaddressthisissue,weadopttheUEPimplementationofthe u 1:(cid:96) ri u,i u,i 1:(cid:96)
r1=0r(cid:96)=0i=1
RLNC [8], which we shall refer to as UEP-NC. In this case,
where R (r ) represents the probability that user u success-
instead of generating coded packets by linearly combining u 1:(cid:96)
fully receives r out of N coded elements, for i = 1,...,(cid:96).
source packets over the whole message x, the RLNC process i i
Function g(r ) is the probability of recovering the (cid:96)-th
takes place over a nested structure of L expanding windows. 1:(cid:96)
expandingwindow,giventhatuseruhassuccessfullycollected
AsdepictedinFig.1,the(cid:96)-thexpandingwindowx consists
1:(cid:96) the set of PDUs described by r .
of all the source elements belonging to the first (cid:96) layers. 1:(cid:96)
In other words, x1:(cid:96) is formed by the first K(cid:96) = (cid:80)(cid:96)i=1ki 1Asa consequence,thetotalnumber oftransmittedcodedelements asso-
elements of the source message x. ciatedwiththe(cid:96)-thexpandingwindowisN(cid:96)n(cid:96).
1 1 Data$stream Data$stream ...
associated$with$v$ 1 $associated$with$v$ 2 $
0.8 0.8 Source$Message
AApppp..,,ℓℓ==12,,ppuu,,ℓℓ==00..11 MAC x1 x2 ... xK2 ... xK
NP()uℓ1:000...246 NP()uℓ1:000...246 ASSSAAASiiiippppmmmmpppp........,,,,,,,,ℓℓℓℓℓℓℓℓ========12313123,,,,,,,,ppppppppuuuuuuuu,,,,,,,,ℓℓℓℓℓℓℓℓ========00000000........11141444 gj,1 ⊗y1⊗...gj⊕y,2jyj .⊗.. gj,KMaws2iAsthoCc$vPia2DteUd frequency TBAwsitshocxia1te:e1MdBMS4capAwasibtshloecx&siau1tb:e2fdrames Awrsiatshdoicxoia&1ft:r3eadme
Sim.,ℓ=2,pu,ℓ=0.4
00 10 20 30 40 50 00 10 20 Sim3.,0ℓ=3,4p0u,ℓ=0.450 PHY TB *me
t t
(a)ConsideredLTE-Astack (b)Radioframemodel
(a)n(cid:96)=2for(cid:96)=1,...,L (b)n(cid:96)=5for(cid:96)=1,...,L
Fig. 2. Performance comparison between the approximated (“App.”) and Fig.3. GenerationofaMACPDUassociatedwithx1:2 (leftside)andTB
simulated(“Sim.”)versionofPu(N1:(cid:96)). allocationonanLTE-AradioframeforL=3(rightside).
To the best of our knowledge, a closed form expression for that the user QoS improves with the number of recoverable
g(r1:(cid:96))isnotknown.However,weknowthatifthefieldsizeq consecutive video layers, starting from v1.
tends to infinity, the probability of receiving a coded element, In our system model, each data stream associated with a
which is linearly independent of previously received coded video layer traverses the Packet Data Conversion Protocol
elements, approaches one. Following the rationale of [13] and (PDCP), the Radio Link Control (RLC) and the MAC layers
consideringlargevaluesofq,wesetg(r1:(cid:96))=1ifthenumber of the LTE-A protocol stack of the base station. In this paper,
r(cid:96)n(cid:96) of successfully received coded elements associated with similarlytotheprotocolstackpresentedin[6],wemakeuseof
the (cid:96)-th expanding window is equal to or greater than a a modified implementation of the standard MAC layer, which
threshold value rmin,(cid:96); otherwise, we set g(r1:(cid:96))=0. We can allows the transmission of the video stream by means of the
thus write UEP-NC principle as explained in Section II-A.
g(r1:(cid:96))(cid:39)I(r(cid:96)n(cid:96) ≥rmin,(cid:96)) (2) The output of an H.264 video encoder can be modelled
as a stream of Group of Pictures (GoPs), each GoP has a
where I(s) is the indication function, i.e., I(s) = 1 if
duration of d seconds. Since the video decoding process
statement s is true, otherwise I(s) = 0. Using the same GoP
takes place on a GoP-by-GoP basis [2], a GoP represents
reasoning as in [13], we can express r as the following
min,(cid:96) a layered source message. According to the layered model
recursion
presented in Section II-A, the (cid:96)-th layer of a source message
r =K −K +max(r − r n , 0) (3) consists of k elements. To fit our model, the data stream
min,(cid:96) (cid:96) (cid:96)−1 min,(cid:96)−1 (cid:96) (cid:96)−1 (cid:96)
associated with the (cid:96)-th layer of a GoP is segmented by the
for an initial value of r =K .
min,1 1 MAClayerintok sourceelements.IfH isthenumberofbits
In this paper, we invoke approximation (2) for the cal- (cid:96)
comprising each source element and b is the bit-rate of the
culation of (1). To validate the accuracy of the proposed (cid:96)
(cid:96)-th layer of the GoP, the number of source elements forming
approximation, we compare the probability values derived
the (cid:96)-th layer is k =(cid:100)(b d )/H(cid:101).
from (1) with those obtained by simulation. Fig. 2 shows (cid:96) (cid:96) GoP
As shown in Fig. 3a, the modified MAC layer generates
both the approximated and simulated probability values of
coded elements associated with the (cid:96)-th expanding window.
P (N ) in the case of a three-layer source message with
u 1:(cid:96) All coded elements are then mapped onto one or more MAC
K = 10, K = 50 and K = 100, when the finite field
1 2 3 PDUs,eachofwhichcanholdn codedpackets.Thesequence
has q = 28 elements. We assume that each PDU can hold (cid:96)
of N MAC PDUs is then forwarded to the physical layer.
the same number of coded elements (n = 2 or 5) and (cid:96)
(cid:96) The LTE-A standard imposes that a MAC PDU is mapped
is affected by the same erasure probability (p = 0.1 or
u,(cid:96) onto exactly one physical layer Transport Block (TB). For
0.4). Probabilities have been plotted as a function of t, where
this reason, we use the terms PDU and TB interchangeably.
N =t, for (cid:96)=1,2,3. We observe that the performance gap
(cid:96) Fig. 3b depicts the frequency-time structure of an LTE-A
between simulation and approximated calculations is smaller
radio frame. Every radio frame comprises 10 subframes, each
than7·10−3,whichconfirmsthattheproposedapproximation
of which has a time duration of one Transmission Time
is accurate for q ≥ 28. In particular, in the rest of the paper,
Interval (TTI), which is equal to d = 1ms. Since at
we will refer to q =28. TTI
most 6 subframes can convey eMBMS data traffic [3], we
B. Layered Video Multicasting over eMBMS Networks use feMBMS = 0.6 to refer to the maximum fraction of
eMBMS-capable subframes per radio frame. Notice that a TB
Let us first consider a network operating in SC-eMBMS
is a fraction of the radio frame. It has a fixed time duration
modewithinwhichasinglebasestationdeliversaH.264/SVC
of one TTI, and consists of N Resource Block Pairs2
video stream. The stream consists of L layers {v ,...,v }, RBP
1 L
where v1 represents the base video layer, and {v2,...,vL} 2A resource block pair can be modelled as a frequency-time structure
represent the L − 1 enhancement video layers. We remark spanningabandwidthof180kHzandatransmissiontimeintervalof1TTI.
(RBPs) [3], not shown in Fig. 3b. The number n of coded thisproblem,ourproposedoptimizationstrategysetsp =pˆ
(cid:96) u,(cid:96)
elements that a TB can hold depends on the adopted MCS, if m ≤m(u), otherwise p =1.
(cid:96) u,(cid:96)
characterizedbyindexm ,thevalueofN andtheelement Letδ beanindicationvariableassociatedwithuseruand
(cid:96) RBP u,(cid:96)
size H. Table I reports the values of the ratio n /N for QoS level (cid:96), so that δ =1 if the first (cid:96) service layers are
(cid:96) RBP u,(cid:96)
H = 2 KB and different MCS indexes. The reported values recoveredbyuseruwithaprobabilityofatleastQˆ;otherwise,
have been selected in order to fit the actual bit capacity of a δ = 0. We understand that user u will recover the first (cid:96)
u,(cid:96)
TB [14]. service layers if it successfully recovers the (cid:96)-th expanding
ConsiderFig.3b,weassumethatatmostoneTBassociated windoworanyoftheexpandingwindowshavingindexgreater
withthesameexpandingwindowcanbetransmittedduringthe than (cid:96). Thus, the indication variable δ can be expressed as
u,(cid:96)
samesubframe.Theremainingpartoftheradioframethathas
(cid:32) (cid:33)
L
notbeenreservedforthedeliveryofthelayeredvideoservice δ =I (cid:95)P (N )≥Qˆ (4)
under consideration (marked in white) can be used by other u,(cid:96) u 1:i
i=(cid:96)
LTE-A eMBMS services.
where ∨ denotes the logical OR operator.
We now shift our focus from the SC-eMBMS to the
We define the profit made by the system as the number of
SFN-eMBMS mode. In this case, base stations belonging to
videolayersthatanyusercanrecoverwithagivenprobability.
the same SFN are connected to the Multicell Coordination
Using (4), we can express the profit as (cid:80)U (cid:80)L δ . On
Entity (MCE), which is in charge of synchronizing the trans- u=1 (cid:96)=1 u,(cid:96)
the other hand, we represent the cost incurred by the system
missionsofallbasestations[3].TheLTE-Astandardimposes
as the number of required TBs for the delivery of a GoP,
that all SFN base stations multicast the same video stream
whichiscapturedby(cid:80)L N .Inordertooptimizetheprofit
using the same physical signals. To this end, we assume (cid:96)=1 (cid:96)
while keeping the cost low, the proposed UEP-RAM aims at
that all RNGs, used by the SFN base stations during the
maximizing the profit-cost ratio as follows
networkencodingprocess,arekeptsynchronizedbytheMCE.
Consequently, all SFN base stations generate the same set of (cid:88)U (cid:88)L (cid:46)(cid:88)L
coded elements. (UEP-RAM) max δ N (5)
u,(cid:96) (cid:96)
InbothSC-andSFN-eMBMSmodes,weassumethateach mN11,,......,,mNLL u=1(cid:96)=1 (cid:96)=1
user can provide Channel Quality Indicator (CQI) feedback
U
(cid:88)
on the downlink channel conditions. CQI feedback is sent to subject to δ ≥Utˆ (cid:96)=1,...,L (6)
u,(cid:96) (cid:96)
the only base station in SC-eMBMS or the base stations that
u=1
is closest to the user in SFN-eMBMS. In the latter case, the 0≤N ≤Nˆ (cid:96)=1,...,L. (7)
(cid:96) (cid:96)
reported CQI is forwarded to the MCE. In the CQI feedback,
each user specifies the index of the MCS that ensures the Objective (5) represents the system profit-cost ratio. Con-
maximum transmission rate and a TB error probability or, straint(6)ensuresthatthefractionofusersrecoveringthefirst
equivalently,aMACPDUerasureprobabilitynotgreaterthan (cid:96) video layers with a probability of at least Qˆ shall not be
pˆ=0.1 [3]. smaller than tˆ(cid:96). Constraint (7) imposes that the transmission
of data associated with the (cid:96)-th expanding window shall
III. SCALABLEVIDEOMULTICASTINGOPTIMIZATION not require more than Nˆ TBs. Of course, the transmission
(cid:96)
In this section, we propose UEP-RAM to maximize the of a GoP should not exceed the GoP duration d . For
GoP
QoS level experienced by each user during the transmission this reason, the values of Nˆ ,...,Nˆ have to be such that
1 L
of scalable video streams, while keeping the number of TB max(Nˆ ,...,Nˆ )≤(cid:98)f ·d /d (cid:99).
1 L eMBMS GoP TTI
transmissions low. The constraint set of the proposed model Unfortunately,theUEP-RAMisacomputationallycomplex
ensures that a predetermined fraction of users shall achieve optimization problem because of the coupling constraints
the desired QoS level with at least a given probability. In the among optimization variables introduced by (6). For this
case of SC-eMBMS, we assume that the proposed resource reason, we propose a novel heuristic strategy, summarized in
allocation strategy is performed at the MAC layer along with Procedure1,whichcanbeusedtofindagoodqualitysolution
the scheduling operations defined by the standard. In the case of the UEP-RAM in a finite number of steps.
ofSFN-eMBMS,theMCEisinchargeofschedulingservices Consider constraint (7) which allows N to be equal to
(cid:96)
within the SFN. For this reason, we assume that UEP-RAM 0. In that case, coded elements associated with x will be
1:(cid:96)
is also handled by the MCE. neither generated nor transmitted. Let us assume that the first
We denote by m(u) ∈ [1,15] the MCS index reported in sexpandingwindowsarenotdelivered,where0≤s≤L−1.
the CQI feedback of a user u ∈ [1,U]. Upon receiving the The desired service coverage can only be offered if (i) the
m(u) valuesforanyvalueofu,basestationsemploythesame first (s+1) service layers are recovered with a probability
MCSwithindexm tobroadcastTBsassociatedwiththe(cid:96)-th of at least Qˆ by the user fraction t(cid:48) = tˆ, and (ii) the
(cid:96) s+1 1
expanding window. If m ≤m(u) for a user u, the TB error remaining layers are recovered by the user fractions t(cid:48) = tˆ,
(cid:96) i i
probability p experienced by that user will not be greater where i=(s+2),...,L.
u,(cid:96)
thanpˆ.However,basestationscannotcomputeP (N )using The rationale for the proposed heuristic can be summarized
u 1:(cid:96)
(1)astheyarenotawareoftheexactvalueofp .Toalleviate as follows. The value of s is initially set to L−1 (line 1 of
u,(cid:96)
TABLEI
Procedure 1 Heuristic UEP-RAM SIMULATIONPARAMETERSANDTHEVIDEOSTREAMSCONSIDERED.
1: s←L−1
2: whiles≥0do Paramter Value
3: N(cid:96)←0andm(cid:96)←0,forany(cid:96)∈[1,...,L] Inter-Site-Distance(ISD) 500m
564::: tmfo(cid:48)sr+(cid:96)(cid:96)1←←←sostˆl1u+t,ioa1nn,do.ft.(cid:48)i.S,←1L-(cid:96)tˆdpior,ofbolreman,yfoir∈an[ys(cid:96)+∈2[,s.+..,1L,.]..,L] SCDyasurtrpeilmeerxBiFnragenqdMuweoniddcetyh 2F2D.00DMGHHzz
7: ifS2-(cid:96)canbesolvedthenN(cid:96)←solutionofS2-(cid:96) BasestatTiorannasmndisusisoenrAPonwteenrnaGains 4se6edTBamblepeAr.-2se.1c.t1o-r2[15]
8: ifconstraint(6)holdsthen
PathlossandPenetrationLoss seeTableA.2.1.1.5-1[15]
9: mINT,(cid:96)←m(cid:96)andNINT,(cid:96)←N(cid:96),forany(cid:96)∈[1,...,L] ChannelModel ITU-TPedA[16]
10: for(cid:96)←L,...,s+2do
H 2KB
1112:: ifNm(cid:96)−(cid:96)1←>m0a(cid:96)n−d1,Na(cid:96)nd>N0(cid:96)t←hen0 formn(cid:96)(cid:96)=/N4R,B.P.,.,15 22,03,,656,,67,28,10,12,14,17,
13: ifS2-(cid:96)canbesolvedthen Qˆ 0.99
11115467:::: if(cid:80)L(cid:96)=1eNlseINNmiTf,(cid:96)(cid:96)(cid:96)N←<←(cid:96)>(cid:80)smoNlL(cid:96)IuN=tITNi1o,T(cid:96)n,N,(cid:96)oa(cid:96)tfnhtdSeh2nNe-n(cid:96)(cid:96),a←ndNmIN(cid:96)T−,(cid:96)1←0 StreawmheArNeˆ[(cid:96)1n,7m]fionr=(cid:96)=m{(cid:96)1ρ∈,1m[.,4.,.i.n...,..,Lρ153,]}n[(cid:96)dB] {(cid:100)2nk7m(cid:96)./i9nn,=m3i5n2.(cid:101)N9+,R4(cid:100)B5pPˆ.(cid:100)8k}(cid:96)/nmin(cid:101)(cid:101),
18: returnmINT,1,...,mINT,LandNINT,1,...,NINT,L (L=3) {b1,...,b3}[Kbps] {47.3,326.1,1396.7}
19: else {tˆ1,...,tˆ3} {0.99,0.8,0.6}
20: returnm1,...,mLandN1,...,NL StreamB[18] {ρ1,...ρ4}[dB] {28.1,33.4,39.9,46.4}
21: else s←s−1 (L=4) {b1,...,b4}[Kbps] {36.8,79.4,303.4,835.9}
22: returnnosolutionfound. {tˆ1,...,tˆ4} {0.99,0.9,0.75,0.6}
StreamA,B dGoP 0.533s
Procedure 1). The while-loop (lines 2-21) iteratively tries
returns the refined solution. Otherwise, the procedure returns
to find a feasible solution to the UEP-RAM. However, if
the intermediate solution.
constraints (6) and (7) are not met, then the value of s is
decreased (line 21). In particular, for any (cid:96) ∈ [s+1,L], the IV. NUMERICALRESULTS
following problem is solved (line 5)
Theperformanceoftheproposedresourceallocationframe-
(cid:110)(cid:88)U (cid:16) (cid:17) (cid:111) work is investigated in this section. We consider a network
(S1-(cid:96)) argmax I m ≤m(u) ≥Ut(cid:48) . (8)
(cid:96) (cid:96) composed of 19 base stations, each of which controls three
m(cid:96)∈[1,...,15] u=1
hexagonal cell sectors. In the case of a SC-eMBMS network,
The solution to (8) is the largest possible value of m for
(cid:96) 18 interfering base stations are organized in two concentric
which at least U t(cid:48) users experience a TB error probability
(cid:96) rings centred on the base station delivering the target scalable
equal to or smaller than pˆ.
video stream. In the case of a SFN-eMBMS configuration,
LetP(N )betheuser-agnosticversionof(1)representing
1:(cid:96) we consider a SFN formed by 4 base stations surrounded by
theprobabilityofrecoveringthe(cid:96)-thexpandingwindowwhen
the remaining 15 interfering base stations. Table I describes
pˆis the erasure probability of all TB transmissions. The value
the main simulation parameters and target values for two
of N is obtained by iteratively solving, for (cid:96)=s+1,...,L,
(cid:96) H.264/SVC video streams, namely stream A and B [2].
the problem below3 (lines 6-7)
To assess the quality of the UEP-RAM solutions generated
(cid:110) (cid:111)
(S2-(cid:96)) argmin P(N )≥Qˆ ∧ 0≤N ≤Nˆ . (9) by our proposed heuristic strategy, we shall compare them to
1:(cid:96) (cid:96) (cid:96)
N(cid:96)∈[0,...,Nˆ(cid:96)] resource allocation solutions obtained by directly solving the
UEP-RAM using a genetic strategy [19], hereafter referred to
where ∧ denotes the logical AND operator.
as direct solution. Genetic approaches are known to provide
Both the MCS indexes and the number of TB transmis-
a tight approximation of the optimum solution to the con-
sions associated to each expanding window obtained so far
sidered problem class [20]. Both the direct and the heuristic
define the intermediate solution of the heuristic UEP-RAM
solutions shall be compared in terms of the profit-cost ratio
{mINT,1,...,mINT,L,NINT,1,...,NINT,L}. If the solution of τ = (cid:80)U (cid:80)L δ (cid:14)(cid:80)L N that they can achieve. For
S2-(cid:96) does not meet constraint (6), the value of s is decreased u=1 (cid:96)=1 u,(cid:96) (cid:96)=1 (cid:96)
clarity, we remark that genetic strategies are characterized by
and the while loop continues until all valid values of s are
a considerable computational complexity, which makes them
exhausted.Ifconstraint(6)isfulfilled,arefinementprocedure
unsuitable for use in practical network scenarios [19].
(lines9-20)aimingatimprovingthequalityoftheintermediate
To assess the effectiveness of the proposed UEP-RAM, we
solution is initiated. In particular, the MCS index m for
(cid:96) compare it against the MrT resource allocation idea proposed
(cid:96) ∈ [s+2,L] is potentially decreased by setting it to m ,
(cid:96)−1 in[10],[11],whichreliesonastandardLTE-AMAClayerand
N is set to 0 and N is recomputed. We remark that if
(cid:96)−1 (cid:96) does not employ NC or AL-FEC error protection strategies.
N is set to zero, coded packets associated to x will not
(cid:96) 1:(cid:96) If ρ denotes the highest Peak Signal-to-Noise Ratio (PSNR)
beengeneratedandtransmitted.Iftherefinedsolutionrequires (cid:96)
achieved by user u after the recovery of the first (cid:96) layers [2],
a total number of TB transmissions which is smaller than
while P˙ is the probability that user u will recover the first
that associated with the intermediate solution, the procedure u,(cid:96)
(cid:96) layers, the MrT strategy defines the user performance level
3Since S1-(cid:96) and S2-(cid:96) are characterized by small discrete feasible spaces, as
(cid:110) (cid:111)
they are simple enough to permit the use of exhaustive search to find their ρ˙ = max ρ ·P˙ . (10)
solutions. u (cid:96)=1,...,L (cid:96) u,(cid:96)
110 1
Dir.UEP-RAM,StreamA
Heu.UEP-RAM,StreamA
τRatio90 DHeiru..UUEEPP--RRAAMM,,SSttrreeaammBB bability00..68
fiProt-Cost3570003 4 5 6 7 8 9 10 11 12 Recoverypro00..0249 0 MHvvv111e,,ru1vvT.221,U0v3EP-R1A3M0 150 170 ˆt=0.63190 2ˆ1t=0.820 230 ˆt=0.991250 270 290
NumberofRBPsperTBNRBP Distance(m)
Fig.4. Profit-costratioassociatedwiththedirect(“Dir.”)andtheheuristic (a)StreamA
(“Heu.”)solutionoftheUEP-RAMmodelvs.thenumberofRBPsper-TB.
1
TsuhmecoofnsaildleuresedrMpreTrfoalrlmocaantcioenlemveetlhsobdyaiompstimatimsinagximthieziMngCtShes bility0.8
ba0.6
used for the delivery of each layer of the video stream. The pro
MrT optim(MizraTti)onmpr1mo,.b.a.l,xemm(cid:96)(cid:88)cUanρ˙bue summarized as (11) Recovery00..0249 0 MHvvvv1111e,,,ru1vvvT.2221,,U0vv33E,Pv4-R1A3M0 150 170 ˆt=0.64190 ˆt=0.753210 ˆt=0.92230 ˆt=0.991250 270 290
u=1 Distance(m)
subject to m <m (cid:96)=2,...,L (12) (b)StreamB
(cid:96)−1 (cid:96)
Fig.5. Probabilityofrecoveringthefirst(cid:96)videolayers(with(cid:96)=1,...,L)
where constraint (12) allows the model to adopt a different
vs.thedistancefromcentreofthecell.
MCSforeachvideolayerand,thus,exploituserheterogeneity
in terms of radio propagation conditions. of 90 + 47 · 2 = 184 m from the centre of the cell, shall
For the sake of clarity, note that we refer to the approxima- recover the first four video layers with probability of at least
tion of p , described in Section III only during the resource Qˆ =0.99.
u,l
resource allocation operations. When assessing the user QoS, Figs. 5a and 5b show the probability of recovering the first
as presented in the following sections, we refer to the actual (cid:96) video layers as a function of the distance from the base
PDU error probability experienced by each user. station in the centre of the cell. Vertical dashed lines denote
the target distance for each QoS level. We observe that the
A. SC-eMBMS Delivery Mode
proposed UEP-RAM comfortably meets the service coverage
In order to better study the behaviour of the proposed constraints.Inaddition,UEP-RAMensuresaservicecoverage
UEP-RAM strategy, we adopt a user disposition pattern which is considerably greater than that of MrT. For instance,
that ensures significant user heterogeneity. In the case of if UEP-RAM is applied on stream A (Fig. 5a), the coverage
SC-eMBMS, users are placed on the radial line represent- area will be 64% greater than that of MrT for base layer
ing the symmetry axis of one of the cell sectors, which is provisioning (i.e., layer v only) and 41% greater for all-layer
1
controlled by the base station transmitting the target video provisioning (i.e., layers v , v and v ). The same trend is
1 2 3
service. We consider U = 80 users, which are spaced 2 m observed for stream B (Fig. 5b).
apart with the first user being placed 90 m from the base
station transmitting the target video service.
B. SFN-eMBMS Delivery Mode
Fig. 4 compares the profit-cost ratios associated with re-
source allocation solutions obtained by directly solving the Similarly to the SC-eMBMS scenario, the user distribution
UEP-RAM and by using the proposed heuristic strategy, as a in the SFN-eMBMS case is characterized by significant het-
functionofN forbothstreamAandstreamB.Weobserve erogeneity. We consider a network composed of U = 1700
RBP
that the performance gap between the direct and the heuris- users located at the vertices of a regular square grid with step
tic solutions of UEP-RAM is negligible. Heuristic solutions 20 m. The user grid is placed on the playground in such a
deviate at most 4.3% from direct solutions in Fig. 4. Having way that the area spanned by the SFN is uniformly covered,
establishedthattheproposedheuristicstrategygeneratesgood starting from inside the SFN.
quality solutions, we utilize it in the performance comparison In this section, we use the maximum PSNR experienced by
of the MrT and UEP-RAM methods. a user as the performance metric. As in the previous section,
We first draw attention to the fact that tˆ, which is the theadopteduserdistributionallowsustoconvertthemaximum
(cid:96)
fraction of users that shall receive the first (cid:96) video layers PSNR at each user into contours of the playground. Conse-
with probability of at least Qˆ =0.99, can also be interpreted quently, a visual comparison of the coverage areas offered by
as a distance from the centre of the cell. For example, if MrT and UEP-RAM can be performed. We explained that the
we consider the value tˆ = 0.6 and we recall the user maximum PSNR for the MrT strategy, denoted by ρ˙ , can be
4 u
disposition pattern that we have adopted, we understand that obtained from (10). The formulation of the maximum PSNR
0.6·80 = 48 of the users, who are located within a distance for the proposed UEP-RAM can be easily adapted from (10)
TABLEII
yposition(m)yposition(m)−−−−12345671234567212100000000000000000000000000000000000000−−M559.7200r00T35.8.549−−233700.90042597.42.935.53785.−−8.54.8.399511.2900700.49355.2.9738.11594005..009845.458.8227.7.3349549.30085005..8935.9359.722.97U554.H009245E0075e.P8.u.9$8Rr3i5A.s395.9/77M9.7200c00 yposition(m)yposition(m)−−−−12345671234567212100000000000000000000000000000000000000−−M5500r002T8.31496..−−49330020042868.4331.9..94691.−−.4.649431133.39300.34.90034.444336969..294..811.494006.14300294..6849.13436433..4.6444006200.46.484333.3919.4..9391.382U55.H3004E4003e4P6.6u.'424.Rr84i399..A99.s3124.33/778M.001c00 FRRveAc1vCo,1vTvve,I12rvvOe,,21dNvv,23LvO,a3vFye4UrsSER333SM964R...-r911ET%%%CSOtrVeEaUmREI1HAN87P014e-G-0Ru%%%rAE.AMCHSE3333TM9643....Or9131T%%%%FSVtrIeDaUEmOE811HB6P700L5e-.00RA4u%%%r%AY.MERS.
xxppoossiittiioonn((mm)) xxppoossiittiioonn((mm)) REFERENCES
(a)StreamA (b)StreamB
[1] Cisco Visual Networking Index, “Global mobile data traffic forecast
Fig.6. MaximumPSNRvalues(dB)associatedwiththeMrTandheuristic
update,2013-2018,”2014.
UEP-RAMstrategies(blackcirclesrepresentsthebasestations).
[2] P. Seeling and M. Reisslein, “Video Transport Evaluation With H.264
VideoTraces,”IEEECommun.SurveysTuts.,vol.14,no.4,pp.1142–
as follows 1165,Fourth2012.
[3] E. Dahlman, S. Parkvall, and J. Skold, 4G: LTE/LTE-Advanced for
ρ = max {ρ ·P (N )}. (13)
u (cid:96) u 1:(cid:96) MobileBroadband. ElsevierScience,2011.
(cid:96)=1,...,L
[4] D. Lecompte and F. Gabin, “Evolved multimedia broadcast/multicast
Fig. 6 shows the aforementioned performance metric asso- service (eMBMS) in LTE-advanced: overview and Rel-11 enhance-
ments,”IEEECommun.Mag.,vol.50,no.11,pp.68–74,2012.
ciated with each point of the playground for the two resource
[5] E.Magli,M.Wang,P.Frossard,andA.Markopoulou,“NetworkCoding
allocationstrategiesandthetwovideostreamswhenN =5 MeetsMultimedia:AReview,”IEEETrans.Multimedia,vol.15,no.5,
RBP
RBPs. Coloured regions depict areas where users can achieve pp.1195–1212,2013.
[6] C.Khirallah,D.Vukobratovic´,andJ.Thompson,“PerformanceAnalysis
a maximum PSNR value that is equal to or greater than that
andEnergyEfficiencyofRandomNetworkCodinginLTE-Advanced,”
of ρ shown in Table I. Table II reports the fraction of users IEEETrans.WirelessCommun.,vol.11,no.12,pp.4275–4285,2012.
(cid:96)
in Figs. 6a and 6b that can recover the first (cid:96) video layers, [7] A. Tassi, C. Khirallah, D. Vukobratovic´, F. Chiti, J. Thompson, and
for (cid:96) = 1,...,L, with at least a target probability Qˆ = 0.99. R.Fantacci,“ResourceAllocationStrategiesforNetwork-CodedVideo
BroadcastingServicesoverLTE-Advanced,”IEEETrans.Veh.Technol.,
We observe that the heuristic UEP-RAM ensures a service Jul.2014.
coverage which is considerably greater than that of MrT. For [8] D.Vukobratovic´ andV.Stankovic´,“UnequalErrorProtectionRandom
LinearCodingStrategiesforErasureChannels,”IEEETrans.onCom-
instance,theheuristicUEP-RAMresourceallocationsolutions
mun.,vol.60,no.5,pp.1243–1252,May2012.
allow65%ofuserstorecoverallfourvideolayersofstreamB [9] R. Afolabi, A. Dadlani, and K. Kim, “Multicast Scheduling and Re-
andalluserstorecoverthebaselayer.Incomparison,theMrT sourceAllocationAlgorithmsforOFDMA-BasedSystems:ASurvey,”
IEEECommun.SurveysTuts.,vol.15,no.1,pp.240–254,2013.
strategy successfully delivers all four video layers to about
[10] D. Munaretto, D. Jurca, and J. Widmer, “Scalable video broadcast in
33% of users, while the base layer is received by less than cellular networks: impact on QoS and network resources,” in Proc. of
40% of users. IEEEISCC2010,Riccione,Italy,IT,2010.
[11] H. Zhang, Y. Zheng, M. Khojastepour, and S. Rangarajan, “Scalable
videostreamingoverfadingwirelesschannels,”inProc.ofWCNC2009,
V. CONCLUSIONS Budapest,HU,Apr.2009.
[12] M. Me´dard and A. Sprintson, Network Coding: Fundamentals and
In this paper we presented a viable method for the incorpo- Applications,ser.AcademicPress. Elsevier,2012.
rationoftheUEP-NCschemeintoLTE-Anetworksasameans [13] A. Tassi, I. Chatzigeorgiou, and D. Vukobratovic´, “Resource Allo-
cation Frameworks for Network-coded Layered Multimedia Multicast
of improving the reliability of a H.264/SVC video multicast
Services,”IEEEJ.Sel.AreasCommun.,2014.
service.Furthermore,weproposedanovelresourceallocation [14] D. Vukobratovic´, C. Khirallah, V. Stankovic´, and J. Thompson, “Ran-
frameworkthataimstoimprovetheservicecoveragewithare- dom Network Coding for Multimedia Delivery Services in LTE/LTE-
Advanced,”IEEETrans.Multimedia,vol.16,no.1,pp.277–282,Jan.
ducedresourcefootprint.Inspiredbyafundamentaleconomics
2014.
principle, the goal of the proposed optimization framework is [15] 3GPP TR 36.814 v9.0.0 (Release 9), “Evolved Universal Terrestrial
achievedbymaximizingthesystemprofit-costratio.Inspiteof RadioAccess (E-UTRA);Furtheradvancements forE-UTRAphysical
layeraspects,”2010.
the natural complexity of the proposed modelling, we defined
[16] S.Schwarz,J.Ikuno,M.Simko,M.Taranetz,Q.Wang,andM.Rupp,
a novel heuristic strategy that was used to efficiently derive “Pushing the limits of LTE: A survey on research enhancing the
a good quality resource allocation solution of the considered standard,”IEEEAccess,vol.1,pp.51–62,2013.
[17] “Video Traces 2 for Terminator.” [Online]. Avail-
problem.Numericalresultsshowedthattheproposedresource
able:http://trace.eas.asu.edu/videotraces2/cgs/cif/Terminator1 G16B15
allocation framework ensures a service coverage which is up DQP15 2EL 48 33 18/
to 2.5-times greater than that of the considered conventional [18] “Video Traces 2 for Gandhi.” [Online]. Available: http://trace.eas.asu.
edu/videotraces2/cgs/cif/Gandhi G16B15 DQP10 3EL 48 38 28 18/
MrT strategy.
[19] S. Sivanandam and S. Deepa, Introduction to Genetic Algorithms.
Springer,2007.
ACKNOWLEDGEMT [20] K. Deep, K. Pratap Singh, M. Kansal, and C. Mohan, “A real coded
genetic algorithm for solving integer and mixed integer optimization
This work is part of the R2D2 project, which is supported problems,”AppliedMathematicsandComputation,vol.212,no.2,pp.
by EPSRC under Grant EP/L006251/1. Collaboration of the 505–518,2009.
authors was facilitated by COST Action IC1104 on Random
Network Coding and Designs over GF(q).