Quality-Driven Proactive Caching of Scalable Videos Over Small Cell Networks

Transcription

1 6 th International Conference on Mobile Ad-Hoc and Sensor Networks Quality-Driven Proactive Caching of Scalable Videos Over Small Cell Networks Zhen Tong, Yuedong Xu, Tao Yang,BoHu Research Center of Smart Networks and Systems, School of Information Science and Engineering, Fudan University State Key Laboratory for Novel Software Technology, Nanjing University, P.R. China {7, ydxu, taoyang, Abstract The explosion of mobile video traffic imposes tremendous challenges on present cellular networks. To alleviate the pressure on backhaul links and to enhance the quality of experience (QoE) of video streaming service, small cell base stations (SBS) with caching ability are introduced to assist the content delivery. In this paper, we present the first study on the optimal caching strategy of scalable video coding (SVC) streaming in small cell networks with the consideration of channel diversity and video scalability. We formulate an integer programming problem to maximize the average subjective quality of SVC streaming under the constraint of cache size at each SBS. By establishing connections between subjective quality and caching state of each video, we simplify the proactive caching of SVC as a multiple-choice knapsack problem (MCKP), and propose a low-complexity algorithm using dynamic programming. Our proactive caching strategy reveals the structural properties of cache allocation to each video based on their popularity profiles. Simulation results manifest that the SBSs with caching ability can greatly improve the average quality of SVC streaming, and that our proposed caching strategy acquires significant performance gain compared with other conventional caching policies. I. INTRODUCTION Mobile video service is witnessing a tremendous growth nowadays. As forecasted by Cisco, the video traffic will increase -folds between 5 and, accounting for 75% of the total mobile data traffic by []. To deal with this phenomenon, the industry is advocating the deployment of small cell base stations (SBS) to enable higher density spatial reuse of radio resources. However, a drawback of this approach is the huge expenditure to connect all the SBSs to the core network with fast backhaul links []. Meanwhile, video contents requested by users exhibit significant similarities, thus causing a large amount of redundant traffic []. All these factors jointly propel the concept of caching at the edge of networks [], in order to reduce the traffic load of backhaul and to improve the end-to-end video delivery. Distributed caching over wireless access points has attracted a lot of attentions in the past several years. The authors in [5] [6] introduced the concept of FemtoCaching that equips femto-cells with high storage capacity to store popular files. When a user requests a file, it may be served by a nearby femto-cell with higher throughput and lower delay if this file is cached. Otherwise, the user may be served by the macro- BS through a low-rate long-range wireless link. A recent trend is to coalesce content caching with physical layer features. In [7], the author presented a content placement algorithm to minimize the average file down loading delay from a cluster of cooperative BSs equipped with cache facilities. When the requested file is stored in multiple cache-enabled BSs, these clustering BSs can perform cooperative beamforming to support the transmission of high-quality video streaming [8]. Authors in [9] propose a caching policy over small cell networks based on maximum ratio transmission and zeroforcing beamforming, and their objective is to maximize the average downlink throughput of all the files. In this paper, we study a novel distributed caching problem of scalable videos over small cell networks. We are motivated by the recent prevalence of adaptive video streaming over HTTP (DASH) using scalable video coding (SVC) []. SVC provides temporal, spatial and quality (SNR) scalability in three dimensions, thus each video is encoded into multiple layers consisting of a basic layer and several enhancement layers []. The basic layer yields the minimum video quality and each level of enhancement layer provides incremental quality to the lower layers. When delivering SVC streaming over small cell networks, the content placement problem becomes much more challenging. There is a dilemma between the video quality and the overall cache hit ratio. Caching one video in multiple SBSs brings the channel diversity gain, therefore users will receive the video at a higher throughput, however the overall hit ratio will be dragged down due to the redundant cache expenditure. On the other hand, caching a video with more layers (i.e., enhancing the video scalability), fine-granularity bitrate switching points are provided, thus the video is robust to channel variation during the transmission. Above all, the diversity and scalability improvement of certain videos are at the cost of cache space, to be specific, the overall hit ratio, the diversity and scalability of other videos. To allocate cache resource efficiently for each video over small cell networks, a series of interesting questions arise: i) which videos should be cached, ii) what is the caching diversity and the bitrate of each video (i.e., the channel diversity and the video scalability), so as to optimize the average video quality, given the constraint of cache capacity? To resolve these issues, we formulate an integer programming problem to maximize the average quality of SVC streaming under the constraint of cache size at each SBS. With appropriate simplifications, the original caching allocation problem is transformed into a multiple-choice knapsack problem (MCKP) solved by dynamic programming. We /6 $. 6 IEEE DOI.9/MSN.6. 9

2 demonstrate the optimal caching allocation for each video, and simulation results manifest that the proposed algorithm significantly improves the average quality of SVC streaming compared with the baseline algorithms. To summarize, the major features of our work are as follows. We utilize the video subjective quality as a metric to optimize the content placement over small cell networks, which is different from other metrics such as transmission delay and average throughput in the closely related works. Two different delivery schemes are considered, namely the opportunistic transmission and the cooperative transmission. Our model provides a general solution for SVC video caching under these communication schemes. We reduce the SVC caching problem to MCKP and demonstrate the optimal combination of channel diversity and video scalability for each video. Such optimization information is of crucial importance for the cooperation between operators and content providers. The rest of this paper is organized as follows. Section II demonstrates the system model of video transmission, content placement and SVC video quality. In Section III, we formulate the proactive caching problem of SVC streaming and present the proposed algorithm. Simulation results are provided in Section IV. Finally, we conclude this paper in Section V. II. SYSTEM MODEL A. Video Transmission Model Consider one typical video streaming user in the range of a macro cell base station (MBS). As shown in Fig., each user is located at one cluster which is composed of many short range SBSs with caching ability. According to the prediction from content providers, some popular videos are prefetched in the local cache during the off-peak intervals, and these SBSs with cache ability will assist the MBS for content delivery. Consider the scenario where the requested video is already cached in the SBS cluster. Let N = {,,...,N} represent the set of SBSs in the neighborhood of a typical user. We denote the set of candidate SBSs which have cached video i as Ω i, where Ω i N. As the Fig. shows, we have Ω i = {,, } in cluster. After acquiring the set Ω i of potential SBSs, the user will select appropriate SBSs to deliver the requested video. In this paper, two transmission schemes are considered, namely the opportunistic transmission (OT) mode and the cooperative transmission (CT) mode. In another scenario, considering that the SBS is abundant in cache capacity but lacking in backhaul capcity, thereby, if the requested video is not cached in SBSs cluser, then the MBS will respond to the user requests [5]. As shown in Fig., the requested video i is not found in cluster (i.e, Ω i = ), consequently the MBS will fetch the video from remote servers and deliver it to users with extra backhaul delay. B. Content Placement Model Suppose that the video library contains a set of M = {,...,M} videos with the relevant popularity distribution denoted by P = {p,...,p M }. Without loss of generality, we SBS Server Delay Internet SBS User MBS SBS SBS SBS SBS SBS User Cluster SBS Cluster Fig.. Video transmission and placement model. Cluster: video cached in SBS cluster; Cluster: video served by MBS through backhaul. assume that the video popularity follows p p M, i.e., sorted in the descending order. Define s i = {, } N, where s i,n =indicates that the i-th video is cached in the n-th SBS, and s i,n =otherwise. We have l -norm s i = Ω i, which represents how many SBSs have cached the i-th video, in other words, the channel diversity of the i-th video. SVC steaming allows multiple operation points (OP), where a sub-stream of a certain bitrate is extracted and combined by different layers. For any video i M, let R OP = {R,...,R l,...,r L } indicate the different bitrate at OP l, where R < <R l < <R L. Therefore, we define a -tuple (s i,r i ), where r i R OP, as the caching state of the i-th video. The state (s i,r i ) of the i-th video provides the caching details: the i-th video is cached in the SBSs whose s i,n =with the bitrate r i. C. Quality Model of SVC Streaming According to the subjective tests of scalable video, it is observed that the influence of frame rate (FR) and quantization stepsizes (QS) on user perception is separable []. For a given sub-stream with the bitrate of r (at a certain OP), there exists an optimal combination of temporal, spacial and quality layers, which will maximize the subjective video quality. Assuming that a scalable video is encoded at a maximum bitrate of R max, [] has derived a normalized Rate-Quality model of scalable videos: Q(r) =e α( r R max ) β +α () where α, β are model parameters, and r R max. This exponential function implies that the marginal utility for video quality decreases as the video bitrate increases. In this paper, we will consider the mean opinion score (MOS) as the metric of video quality. The MOS is expressed as a single number in range of to 5, where the numbers indicate that: : Bad; : Poor; : Fair; : Good; 5: Excellent. Thereby, the normalized video quality at the bitrate of r can be scaled as Q(r) =+ Q(r). () We will use () as the Rate-Quality function to measure the video quality at different bitrates. 9

3 III. QUALITY-DRIVEN CACHING STRATEGY In this section, our goal is to explore the optimal caching allocation strategy that maximizes the average video quality over small cell networks. First, We will make connections between the caching allocation model and the video quality model, then the problem of maximizing the average video quality will be formulated. Finally, We analyze the problem and propose a proactive caching algorithm. A. Outage Probability of Video Transmission In each cluster, suppose users are served by a certain multiple access method, thus we neglect the influence of interference among different users. We assume that the received SNRs from the MBS and SBSs follow Rayleigh fading, and the probability density function (PDF) of the received SNR is given by P (x) = ( SNR exp x ) () SNR where SNR is the average received SNR. Denote the average SNR of SBSs and MBS as and υ respectively. In general, we have >υ, because the received power from SBSs decays less within the short range cluster. Users can be either served by the SBS with the best channel quality or by several candidate SBSs through cooperative transmission. ) Opportunistic Transmission Mode: for a given video, it will be transmitted by the SBS with the best channel quality from the candidate SBSs having cached the requested file, thus the received SNR can be represented as max = max n Ω SNR n () where Ω is the set of candidate SBSs. Assume that Ω = k, where k =,...,N is the number of candidate SBSs, according to () and (), the PDF of max is given by f OT (x; k) = k ( )[ ( )] k x x exp exp. (5) Therefore, under the opportunistic transmission mode, the outage probability of decoding a video with bitrate r can be written as P OT (k, r) =P{W S log( + x) <r} [ )] = exp ( r k W S (6) where the W S is the average bandwidth allocated to a user from the selected SBS. ) Cooperative Transmission Mode: assume that there are k nearby SBSs having cached the requested video, and these SBSs will be coordinated and perform cooperative transmission. We assume that there is no interference among these cooperative SBSs, and these SBSs have full CSI of the channel. For given channel gains of these coordinated SBSs h =[h,...,h k ], the outage probability [] of a video with bitrate r can be written as P { W S log ( + h ) <r } } = P { h < r W S. (7) Under the independent Rayleigh fading, h is the sum of k independent exponential variables and can be derived as Gamma distribution. Its probability density function (PDF) is given by f CT (x; k, ) = xk exp( x). (8) Γ(k) Combining (7) and (8), for a given video encoded at the bitrate r and cached in k SBSs, the outage probability of cooperative transmission can be derived as k P CT (k, r) = n= n! exp ( θ )( ) n θ (9) where θ = r W S is the SNR threshold for videos encoded with r bitrates. According to (6) and (9), the outage probability of the i-th video in the SBS cluster under different transmission modes can be expressed as { Pout(s S POT ( s i,r l )= i,r l ), OT mode () P CT ( s i,r l ), CT mode where s i is the caching position of the i-th video in the SBS cluster, and R l is the user requested video bitrate at layer l. B. Average Video Quality Over Small Cell Networks SVC provides enhancement layers which are successively refinable, and some bitrate switching mechanisms [] allow the server or client to adapt to the channel variation. In general, the SVC player will first decode the lower layers, and when the downlink throughput reaches the biterate threshold of a higher OP, then the supplementary enhancement layers will be decoded. As a consequence, the average video quality of the i-th video under caching state (s i,r i ) can be derived as q S (s i,r i )=[ Pout(s S i,r i )] Q(r i ) + [Pout(s S i,r l+ ) Pout(s S i,r l )] Q(R l ) () R l <r i which represents that when the downlink bitrate falls in the interval of [R l,r l+ ), then the user will decode the video at the OP l successfully. In another scenario, if the requested video is not cached in the SBS cluster, then it will be delivered by the MBS through backhaul with extra RTT (Round Trip Time) delay. For the adaptive video streaming, video players select bitrate version based on the chunk level, which means that videos are requested segment by segment in case of the bandwidth variation. To decode the video chunk smoothly and avoid the rebuffering, the actual transmission slots should not exceed the video chunk duration, which indicates that T slot T chunk. 9

4 Due to the extra RTT delay caused by backhaul, the actual utilization of the wireless resource will be dragged down, thus the utilization ratio can be bounded by T slot RTT T slot T chunk RTT T chunk = D B () where D B is the maximum utilization ratio of the channel due to the backhaul delay. Similar as (6) and (), the outage probability of decoding OP l with the bitrate R l, and the average video quality provided by the MBS are given by and Pout(R M l )=P{D B W M log( + x) <R l } = exp Rl D B W M () v q M = [ Pout(R M L )] Q(R L ) + [Pout(R M l+ ) Pout(R M l )] Q(R l ). () R l <R L Based on () and (), the average video quality of all the videos over the macro cell and small cell networks can be written as M Q MS = p i [(s i )q S (s i,r i )+(s i =)q M ] (5) i= where (s i )and (s i =)indicate whether the i-th video is cached in the SBS cluster. C. Problem Formulation and Analysis Our objective is to maximize the average video quality by designing the proactive caching policy. Denote by T = {t,t,...,t M } the set of video durations. The optimal content placement problem for scalable videos can be formulated as: P : maximize (s i,r i) subject to Q MS (6) M r i t i s i,n C, n =,...,N (6.a) i= s i {, } N, i =,...,M (6.b) r i R OP, i =,...,M (6.c) where (6.a) is the cache size constraint of each SBS. The optimization problem P is an integer programming, which is in general very complicated to be solved. The computational complexity grows exponentially with regard to the diverse video durations, bitrate levels and the number of SBSs. To determine the optimal caching state (s i,r i ) for each video, it is necessary to make certain reasonable simplifications. First, we assume that all the videos have the same duration denoted by T. This assumption does not change the original problem in (6), because for a video longer than T,itcan be regarded as several videos with the same duration T and popularity. Next, Eq. (6) and P indicate that it is the number of candidate SBSs which have cached the requested video that will determine the average video quality. Let n i = s i denote the number of SBSs that store the i-th video. Consequently, we only need to determine the caching state (n i,r i ) instead of (s i,r i ), thereby the average quality q S (s i,r i ) is replaced by q S (n i,r i ). With above operations, the vector constraints on the cache size can be merged together as a big cache with volume of NC. Since we take all the SBSs as an entirety and skip the process of allocating videos to the specific SBS, there may exit a few videos which can not be cached in the residual space of a single cache. However, even one SBS cache with TB can store hundreds videos, thus the number of these outliers can be ignored compared to those cached videos. Moreover, owing to the scalability of SVC, these exceptional videos can be cached at a lower bitrate. According to the above analysis, the problem P evolves into determining the caching state (n i,r i ) of each video, where n i {,,...,N} and r i {R,R,...,R L }. The total number of combinations of n i and r i is NL, however some combinations are obviously inefficient, actually the efficient caching state must obey the following criteria. Proposition. For an arbitrary video, the efficient combinations of n i and r i should satisfy the following conditions q M <q S (n (),r () ) q S (n (),r () )... q S (n (V ),r (V ) ) n () r () n () r ()... n (V ) r (V ) (7) where the superscripts represent the indexes of efficient combinations. Denote CS = {(n (),r () ),...,(n (V ),r (V ) )} as the set of efficient caching state, and let C = {n () r (),...,n (V ) r (V ) } and Q = {q S (n (),r () ),...,q S (n (V ),r (V ) )}, which are calculated according to CS, denote the set of cache consumption and the set of average video quality respectively. Proof: Assume an arbitrary caching state (n,r ) which satisfies that q S (n,r ) q M. Apparently, the cache consumption n r is not a necessity, because the video under caching state (n,r ) can be served by the MBS with a higher average quality, and then the saved cache resource will be utilized to enhance the quality of other videos. Similarly, consider a caching state (n,r ) which satisfies that q S (n (k),r (k) ) q S (n,r ) q S (n (k+),r (k+) ). If the cache consumption satisfies n (k) r (k) n r n (k+) r (k+), then (n,r ) is an efficient state. However, when n r > n (k+) r (k+), then combination (n (k+),r (k+) ) is superior to (n,r ) with less cache consumption and higher video quality, consequently (n,r ) is not an efficient state. On the other hand, when n r <n (k) r (k), then the state (n,r ) is more efficient than (n (k),r (k) ), therefore the state (n (k),r (k) ) will be substituted by (n,r ). After acquiring the efficient caching state, problem P is reduced to selecting a caching state from CS for each video based on its popularity rank. An important question is that what is the optimal cache allocation scheme for each video, and thus we have the following proposition: 9

5 Proposition. For the optimality of problem P, the videos with higher popularity will be endowed with more cache space so as to achieve higher average quality, which can be described as q S (n,r)... q S (n i,ri )... q S (n M,rM) n r... n i r i... n Mr M (8) where (n i,r i ) is the optimal caching state of the i-th video. Proof: Consider the i-th video and the j-th video, whose popularity satisfy that p i >p j.ifn i r i <n j r j, then we can alway get a higher accumulative quality by swapping the cache space of the video i and the video j. The difference value between the new and the old accumulative quality satisfies that (p i p j )(q(n j,r j ) q(n i,r i )) >. The allocated cache space of each video falls along with the decrease of video popularity. Based on Proposition and Proposition, the problem P can be transformed to P : maximize ˆm, (n i,r i) subject to ˆm i= ˆm i= p i q S (n i,r i )+ n i r i N C M i=ˆm+ p i q M (9) (9.a) ˆm {,...,M} (9.b) (n i,r i ) CS, i =,..., ˆm (9.c) where C is scaled by C/(R T ). The solution to problem P determines: i) ˆm: videos from to ˆm will be cached in the SBS cluster; ii) (n i,r i ): the optimal caching state of each video. D. Proposed Algorithm Problem P can be viewed as a multiple-choice knapsack problem (MCKP) [5]. The M videos with various popularities are viewed as different items, and each video has V choices of candidate caching state considering the channel diversity and the video scalability. According to (9), let F (m, v) denote the overall video quality under the state (m, v), where m is the number of candidate videos, v is the total cache size of the SBS cluster. The iteration relation of F (m, v) and F (m,v) can be described as F (m, v) = max{f (m,v)+p m q M, F (m,v n m r m )+p m q S (n m,r m )}, (n m,r m ) CS, for all cached videos. () Whether the m-th video is cached in the SBS cluster or not depends on the average quality reward brought by adding the m-th video. If the reward of caching the m-th video does not surpass the reward brought by MBS, then the m-th video will be served by the MBS instead of being cached in the SBS cluster. The MCKP can be solved by dynamic programming using the iteration relation () in pseudo-polynomial time. We present Algorithm to determine the optimal ˆm and caching state (n i,r i ) of each video, and the running time of Algorithm is O(ˆmN CV ). Algorithm Determine ˆm and (n i,r i ) by solving MCKP According to Proposition, determine the efficient caching state CS, the relevant video quality set Q, and the consumption set C for m =to M do for v =to N C do F (m, v) =max{f (m,v)+p m q M, F (m,v n m r m )+p m q S (n m,r m )}, where state (n m,r m ) CS Update caching state of each video end for if video m is served by MBS, given cache size of N C then ˆm = m Videos beginning with ˆm are served by MBS return ˆm and (n i,r i ), where i =,..., ˆm end if end for ˆm = M return ˆm and (n i,r i ), where i =,..., ˆm TABLE I SPECIFICS OF SVC STREAMING Resolution Suggested Bitrates (Mbps) Operation Points 9 8., 7. R, R ,.8 R 8, R ,. R 6, R ,.6 R, R R 76. R IV. SIMULATION RESULTS In this section, we present numerical results of the qualitydriven proactive caching of scalable videos over small cell networks. Assume that the content provider offers a library of M = candidate videos, and the popularity of these videos follow Zipf distribution with shape parameter s =.8. The parameters for measuring the video quality are set as α =.6, β =.66 according to []. We set the average video duration to T = hour, and the specifics of SVC streaming are elaborated in Table I following suggestions from []. The video chunk duration is T chunk =second, and the RTT delay between the remote server and the MBS is 5 ms. We demonstrate three sets of experimental results as below. First, the structure of optimal caching strategy is highlighted, revealing the optimal caching state of each video. Next, we illustrate the tendency of average video quality and cache hit ratio under the different number of SBSs and size of cache space. Finally, we compare the performance of our caching strategy with other reference strategies. A. Caching State of Scalable Videos The averaged SNRs and the allocated bandwidths for a MBS user and a SBS user are set to υ =db, W M =MHz, and =db, W S =5MHz respectively. We consider several 9

6 Caching diversity 5 5 N=, C=TB N=, C=TB N=5, C=TB Bit rates (Mbps) 8 6 N=, C=TB N=, C=TB N=5, C=TB Video quality (MOS) 5.5 N=, C=TB N=, C=TB N=5, C=TB Caching diversity (a) Channel diversity for each video (b) Bitrate scalability for each video (c) Average quality of each video Fig.. Caching state of SVC streaming under opportunistic transmission mode. =db, W S =5MHz; υ =db, W M =MHz 5 5 N=, C=TB N=, C=TB N=5, C=TB (a) Channel diversity for each video Bit rates (Mbps) 8 6 N=, C=TB N=, C=TB N=5, C=TB (b) Bitrate scalability for each video Video quality (MOS) 5.5 N=, C=TB N=, C=TB N=5, C=TB (c) Average quality of each video Fig.. Caching state of SVC streaming under cooperative transmission mode. =db, W S =5MHz; υ =db, W M =MHz scenarios where the number of SBSs is N =or N =5in one cluster, and the cache size of each SBS is C =or TB. The caching state of each scalable video under different transmission modes are demonstrated in Fig. () and Fig. (). The detailed analysis of caching state under opportunistic transmission scheme is offered as follows. Fig.(a) represents the degree of caching diversity (i.e., channel diversity) for different videos which are sorted in descending orders according to the relevant popularity prediction. The top one hundred videos are repeatedly cached in more than one SBS, which can provide diversity gain for these hottest videos to achieve a higher throughput. Meanwhile, we notice that the most unpopular videos are not cached. Fig.(b) demonstrates the bitrates (video scalability) of each video. The popular videos are allocated with higher bitrates, thus possessing more OPs to adapt to channel variation and achieving better average video quality. It is worth observing that the caching diversity falls much more faster than thte video scalability as the popularity decreases. Synthesizing the caching state (n i,r i ) of each video represented in Fig.(a) and Fig.(b), we plot the average quality of each cached video in Fig.(c). The caching state of each video under cooperative transmission mode are represented in Fig. (), whose results are similar as Fig. (). A bit difference is that the channel diversity of the hottest videos under cooperative transmission mode is slightly lower than those under opportunistic transmission mode, since the cooperative transmission mode can provide higher success delivery probability compared to the opportunistic transmission mode. B. Tendency of the average video quality and Hit Ratio Now we proceed to explore the tendency of the overall video quality and hit ratio under different transmission schemes. The channel condition is the same as that of the previous experiments. Fig.() shows that even one SBS equipped with TB cache will boost the average video quality compared to the scenario where there is only MBS. Owing to the additional diversity gain brought by multiple SBSs, higher average quality can be provided by adding the number of SBSs in the cluster than simply increasing the cache space of a single SBS. Due to the heavy tail property of Zipf distribution, in spite of the top or videos cached in more than one SBSs, the majority of videos are cached in one SBS, consequently the overall hit ratio mainly depends on the total cache size of the SBS cluster, which is illustrated in Fig.(5). The tendency of OT and CT mode poses the strong similarity. The average video quality under CT mode is slightly higher than that under OT mode, whereas the overall hit ratio of CT mode is a bit lower compared to that of OT mode. C. Comparison with Other Caching Strategies Next, we compare the proposed caching strategy with two baseline caching schemes. The simulation results under opportunistic transmission mode are represented, and the results under cooperative transmission mode are similar. We consider a representative circumstance where SBSs reside in the neighborhood of a user, and the cache size of each SBS is TB. The first reference caching strategy duplicates the most popular videos in each SBS (DMP), thus the cache 95

7 Average video quality (MOS) 5.5 N =, OT N =, OT N = 5, OT N =, CT N =, CT N = 5, CT BS only 5 Cache size (TB) Fig.. Comparison of average video quality under different transmission schemes. OT: opportunistic transmission, CT: cooperative transmission. Hit ratio N =, OT N =, OT N = 5, OT.5 N =, CT N =, CT N = 5, CT. 5 Cache size (TB) Fig. 5. Comparison of overall hit ratio under different transmission schemes. OT: opportunistic transmission, CT: cooperative transmission. Average video quality (MOS) OPT DMP.8 DMP 7. DMP. MHR.8 MHR 7. MHR. BS only 9 5 SNR of the small cell cluster (db) Fig. 6. Comparison of the average video quality between baseline caching strategies and the proposed strategy (OT mode, N =,C =TB). content of any SBS is identical. The other baseline scheme only caches one copy of each video in the SBS cluster in order to maximize the overall hit ratio (MHR). We consider three levels of bitrates for the baseline strategies, i.e.,.8, 7., and. Mbps respectively. Fig. 6 demonstrates the average video quality of these caching strategies under various SNRs of the SBS cluster. We observe that DMP outperforms MHR in the low SNR scenarios, but has poor performance compared with MHR in the high SNR scenarios. Our caching strategy is superior to baseline strategies in all the situations. The reason lies in that it considers the diversity gain of multiple caching for hot videos, and adjusts the cached bitrates adaptively for videos with different popularity. V. CONCLUSION This paper investigated the proactive caching strategy for maximizing the average quality of SVC steaming over small cell networks. Caching videos in collocated small cell base stations may not only reduce transmission range, but also bring channel diversity gains. The content placement problem of SVC streaming is relaxed as a multiple-choice knapsack problem (MCKP), given the channel diversity and the video scalability. Simulation results demonstrate that the SBSs with caching ability can greatly improve the average quality of SVC streaming, and that our algorithm achieve higher average video quality compared with two baseline caching strategies ACKNOWLEDGMENT This work is supposed by Natural Science Foundation of China (No. 6), Shanghai Pujiang Program (No. PJ), Huawei Innovative Research Program (HIRPO), and Open project of State Key Laboratory for Novel Software Technology, Nanjing University (KFKT6B). REFERENCES [] Cisco visual networking index: Global mobile data traffic forecast update, 5-, whitepaper, Cisco, Feb 6. [] V. Chandrasekhar, J. G. Andrews, and A. Gatherer, Femtocell networks: a survey, IEEE Commun. Mag, vol. 6, no. 9, pp , Sept 8. [] S. Woo, E. Jeong, S. Park, J. Lee, S. Ihm, and K. Park, Comparison of caching strategies in modern cellular backhaul networks, in MobiSys. ACM,, pp. 9. [] M. Zink, K. Suh, Y. Gu, and J. Kurose, Characteristics of youtube network traffic at a campus network - measurements, models, and implications, Comput. Netw., vol. 5, no., pp. 5 5, Mar. 9. [5] K. Shanmugam, N. Golrezaei, A. G. Dimakis, A. F. Molisch, and G. Caire, Femtocaching: Wireless content delivery through distributed caching helpers, IEEE Trans. on Inf. Theory, vol. 59, no., pp. 8 8, Dec. [6] N. Golrezaei, K. Shanmugam, A. G. Dimakis, A. F. Molisch, and G. Caire, Femtocaching: Wireless video content delivery through distributed caching helpers, in INFOCOM, Proc. IEEE, March, pp [7] X. Peng, J. C. Shen, J. Zhang, and K. B. Letaief, Backhaul-aware caching placement for wireless networks, in GLOBECOM, 5 IEEE, Dec 5, pp. 6. [8] A. Liu and V. K. N. Lau, Cache-enabled opportunistic cooperative mimo for video streaming in wireless systems, IEEE Trans. on Signal Processing, vol. 6, no., pp. 9, Jan. [9] W. C. Ao and K. Psounis, Distributed caching and small cell cooperation for fast content delivery, in MobiHoc 5, Proc. ACM, 5, pp [] M. Grafl, C. Timmerer, H. Hellwagner, W. Cherif, and A. Ksentini, Evaluation of hybrid scalable video coding for http-based adaptive media streaming with high-definition content, in WoWMoM, Symp. IEEE, June, pp. 7. [] H. Schwarz, D. Marpe, and T. Wiegand, Overview of the scalable video coding extension of the h.6/avc standard, IEEE Trans. on Circuits and Syst. for Video Tech., vol. 7, no. 9, pp., Sept 7. [] H. Hu, X. Zhu, Y. Wang, R. Pan, J. Zhu, and F. Bonomi, Proxy-based multi-stream scalable video adaptation over wireless networks using subjective quality and rate models, IEEE Trans. on Multimedia, vol. 5, no. 7, pp , Nov. [] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. New York, NY, USA: Cambridge University Press, 5. [] Y. Sánchez de la Fuente, T. Schierl, and Hellge, idash: Improved dynamic adaptive streaming over http using scalable video coding, in MMSys, Proc. ACM,, pp [5] D. Pisinger, A minimal algorithm for the multiple-choice knapsack problem. Euro. J. of Operational Research, vol. 8, pp. 9,