MEC-Assisted Flexible Transcoding Strategy for Adaptive Bitrate Video Streaming in Small Cell Networks

Chunyu Liu,Heli Zhang,*,Hong Ji,Xi Li

1 School of Information and Communication Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China.

2 Key Laboratory of Universal Wireless Communications,Ministry of Education(Beijing University of Posts and Telecommunications),Beijing 100876,China.

Abstract:Adaptive bitrate video streaming (ABR)has become a critical technique for mobile video streaming to cope with time-varying network conditions and different user preferences.However,there are still many problems in achieving high-quality ABR video streaming over cellular networks.Mobile Edge Computing(MEC)is a promising paradigm to overcome the above problems by providing video transcoding capability and caching the ABR video streaming within the radio access network (RAN).In this paper,we propose a flexible transcoding strategy to provide viewers with low-latency video streaming services in the MEC networks under the limited storage,computing,and spectrum resources.According to the information collected from users,the MEC server acts as a controlling component to adjust the transcoding strategy flexibly based on optimizing the video caching placement strategy.Specifically,we cache the proper bitrate version of the video segments at the edge servers and select the appropriate bitrate version of the video segments to perform transcoding under jointly considering access control,resource allocation,and user preferences.We formulate this problem as a nonconvex optimization and mixed combinatorial problem.Moreover,the simulation results indicate that our proposed algorithm can ensure a low-latency viewing experience for users.

Keywords:mobile edge computing;adaptive bitrate video streaming;flexible transcoding strategy;ADMM

I.INTRODUCTION

Mobile Video streaming is growing rapidly and has the potential to dominate mobile data consumption today.According to the Cisco report,Video streaming is expected to account for 82% of mobile traffic by 2022.Ensuring a good video viewing experience is critical;despite a lot of effort,delivering high-quality video streaming to users over a cellular network remains a considerable challenge under the existing mobile network with limited capability[1,2].

ABR streaming is able to ensure a good viewing experience by dynamically adapting video quality according to the time-varying network conditions and different user preferences [3–8].In adaptive video streaming,a video is divided into equal-duration segments,and each video segment is encoded into multiple bitrate versions/quality levels.Thus,different bitrate versions of the video segments can be dynamically generated and transmitted to the users according to network conditions and user preferences.The higher the bitrate version/quality level of the video segment,the more bits need to be encoded,the more bandwidth resources are required for the process of video transmission.Although ABR technology can provide the appropriate bitrate version of video segments to users,video streaming transmission still faces many problems,such as back-haul congestion and long latency due to fetching video streaming from the remote Internet server traverses a long-distance.

Mobile edge computing (MEC) is a promising paradigm cope with the above problem by pushing the computing and storage to the network edges(e.g.,the small base station),in the proximity of wireless users[9–14].When the MEC servers cache the required bitrate version of the video segments,the demands from users to the same content can be accommodated easily without duplicate transmissions from original content servers that can release the huge backhaul burdens and enable a low-latency connection to users [15].When the MEC servers did not cache the requested bitrate version of the video segment but cached the higher bitrate versions of this video segment,the MEC servers could perform transcoding to obtain the requested version of this video segment instead of fetching the video segment from the remote Internet servers[16].Formulating a MEC-assisted ABR video streaming framework can not only cope with the different user requests and the time-varying network conditions but also reduce the delay and backhaul congestion caused by fetching the video from the remote Internet network.

However,the proposed framework faces several challenges.Firstly,since the caching resources at each MEC server are limited,it is impossible to cache all bitrate versions of each video segment.The effect of video caching is constrained by the caching ability of the MEC server.Secondly,the transcoding dependency among different video versions,increases the complexity of a traditional cache placement strategy where different video segments are treated independently.Thirdly,although performing transcoding to obtain the bitrate version of the video segment requested by the user can reduce caching pressure and improve caching efficiency on the MEC server,transcoding a large number of videos simultaneously can quickly deplete the available computing resources on the MEC server and increase the transcoding delay of all users.Leveraging MEC caching and transcoding ability,we propose a flexible transcoding strategy to further improve the efficiency of video caching and reduce the transcoding delays of all users which beyond what could be achieved using traditional caching approaches.

There are many works that have been done to enhance the performances of the video services by combining the MEC technology and the ABR video streaming.[17–19]focused on the caching placement and video delivery mechanism without considering the dependency between different bitrate versions of the same video segment.[20]introduced transcoding technology into video streaming without considering whether transcoding is appropriate.For example,there is a video segment request for 1Mbps,and the higher bitrate version of this content in the MEC server is 10Mbps.If the MEC server chooses to transcode the higher bitrate version to the request bitrate version,there will be at least 9Mps to deal with,which is much greater than the data size of retrieving from the remote server and may lead to a larger transcoding delay.

In contrast to the previous works,we focus on how to set a flexible transcoding strategy to decide whether to transcode the high bitrate video version to the desired bitrate video version to guarantee the low-latency video streaming service based on optimizing the cache placement strategy (selecting the appropriate bitrate version of the video segments to be cached on the MEC server).With this strategy,many factors are jointly considered,including limited resources,access control,caching placement,and user preference.To this end,we can summarize the main contributions of this paper as follows:

• We formulate a MEC-assisted ABR video streaming framework to provide viewers with lowlatency video streaming services under the timevarying network condition and different user preferences.We develop our problem as a joint optimization of video caching,video transcoding,and video transporting,subject to the caching and transcoding ability constraints at each MEC server and the limited wireless resource.

• We propose a flexible transcoding strategy to determine whether to respond to users’ video requests by transcoding or not.With this strategy,we can select the appropriate bitrate version of the video segments to perform transcoding to reduce all users’ transcoding delay.Besides,by setting this transcoding strategy,the cache strategy is also optimized that enable the MEC server caching the more proper bitrate version of the video segment.

• We extend the concept of video caching from segment level to bitrate version level.Consider the process of video requesting and responding in three different cases,which are direct hit case,transcoding hit case,and cache miss case.

1.1 Related Work

In general,the researches on ABR video streaming mainly focused on video cache placement,the combination with MEC,bandwidth prediction,and Learning-based approaches.

In the aspect of video cache placement:Paper[21]proposed a cooperation sub-channel allocation and caching placement scheme in the NOMA-assisted cloud radio access network with the maximum fronthaul capacity constraint to improve the network sum rate.To maximize the user revenues,the authors in paper[22]proposed a joint user association and resource allocation algorithm in heterogeneous networks with device-to-device (D2D) communications.In emerging heterogeneous cellular networks,paper [23]proposed a cluster-centric scheme with joint consideration of caching and transmission policy to reduce the energy consumption and data traffic of the backhaul.However,the benefits of video transcoding have been utilized in none of the above works.

In the aspect of the combination with MEC:Paper[24]proposed a MEC based video caching mechanism,where only cache the highest bitrate version,which can be transcoded to the requested lower bitrate version by using MEC’s computing capability.In the paper[20],the authors performed the cooperative feasible resource allocation and the joint task scheduling to maintain a low network and minimize the total delay.Paper [3]proposed a joint transcoding and caching framework that supports ABR video streaming in networks with MEC.Paper[25]proposed an effective approach with joint consideration delivery and transcoding based on MEC and blockchain to achieve a decentralized content market among untruthful parties.However,the above work does not fully consider the scenario of the multi-user and multi-base station,and the joint optimization of access decision and transcoding strategies.Paper [26]proposed a MEC enhanced ABR video streaming delivery scheme in the scenario of multiple users and base stations,which combines video segment placement and ABR streaming technology together,but the transcoding strategies are not considered in this paper.

In the aspect of bandwidth prediction:Paper [6]modeled rate adaptation for ABR streaming as a control problem:the past network bandwidth and the amount of video in the playback buffer are monitored by the video player to determine the bitrate version of the current video segment.Paper[27]proposed an online ABR algorithm to balance the contradiction between fast adaptive and smooth bitrate by selecting bitrate based on a multi-step prediction of the future state of the system.Paper[28]proposed an off-line algorithm to solve the problem of non-convex optimization in the case of non-random bandwidth prediction.Paper[29]proposed a new efficient video bitrate version selection scheme to solve inaccurate bandwidth estimation in wireless networks.

In the aspect of Learning-based approaches:Paper[4]designed a deep reinforcement learning(DRL)model to extract significant features from the flight status information,and the ABR algorithm is automatically learned through the training process to adapt to the constantly changing bandwidth capacity of the unmanned aerial vehicle (UAV).Paper [30]proposed a sequential reinforcement learning method for ABR tiles based on 360 video streaming,which makes ABR decisions only by observing the quality of experience(QoE)performance of the results of past decisions.Paper [31]presented an ABR algorithm that combines the deep learning method and the traditional bufferbased algorithm.

1.2 Article Organization

The rest of our paper is organized as follows.In section II,we presented the system model included the communication model,caching model,computing model,and delay model.In section III,we formulated this problem as a non-convex optimization and mixed combinatorial problem.In section IV,we presented the solution to this problem.In section V,we simulated and discussed our scheme.Finally,we conclude this paper in Section VI.

II.SYSTEM MODEL

In this section,we propose a system model to minimize all users’ total delay of transmission and transcoding with limited resources by using the technologies of video transcoding,caching placement,access control,resource allocation,and so on.

There areMSBSs andUUEs randomly distributed in the small cell network,as shown in Figure1.For presentation,we denote the SBS set asM={1,2,3,···,m,···,M}and the UE set asU={1,2,3,···,u,···,U}respectively.Each SBS is equipped with a MEC server with different computing and caching capabilities.All SBSs can connect to the remote Internet server via backhaul links to request the original contents and the transmission rate between the remote Internet server and them-th SBS isrm.

We assume that the user requested a set of video segments,denoted byF={1,2,···,f,···,F}.Each segmentfcontainstseconds in playtime and can be encoded to L different bitrates ranging from the highest bitrate L to the smallest bitrate 1,denoted byL={1,2,···,l,···,L}.MEC servers not only can cache partial segments depending on the caching policy but also can transcode a higher bitrate version segment to a lower bitrate version of the same segment by utilizing their computing capability.In our model,when users request video segments according to their preferences,there are three possible ways to obtain the requested video segment:1)When the required bitrate version of the video segment exactly exists in the MEC server,the SBS will deliver video segment to the user directly.2) If the MEC server caches the higher bitrate version of the video segment instead of the required bitrate version,the MEC server can transcode the cached higher version into the requested version or request the requested version from the remote Internet server through backhaul links,which depends on the transcoding strategy.3)If the MEC server caches neither the exact version of the video segment nor the higher version,SBS will obtain the requested video segment from the remote cloud and then deliver it to the user.

We take a quasi-static scenario and assume that the user remains the same during video delivery,but that it may change during different periods.In the transmission process,the channel does not change,and the perfect instantaneous channel state information can be obtained.

2.1 Communication Model

In this section,we present the communication model in the small cell network based on NOMA,which mainly concentrates on the down-link transmission.Here,the access decision of UEuis denoted asxmu ∈{0,1},which depicts not only whether to access but also where to access.When UEudecides to access the SBSm,we havexmu=1;otherwise,we havexmu=0.So the access strategies can be denoted asX={xmu | m ∈M,u ∈U}.Particularly,it satisfies:

Figure1.System Model.(1)The SBSs in this network are equipped with MEC Servers.(2) Each of the UE requests some video segments from the near SBS.

which means that one user can only access to one SBS at the same time.

In the small cell network with downlink NOMA,because all the users served by the same SBS share the same spectrum,there exists intra-cell interference.Also,to avoid inter-cell interference,we assume the proximate SBSs cannot use the same frequency band with others.Due to the signals from different SBSs have different channel gains,the received signals can be sorted in descending order.

According to the descending order,the interference can be reduced and the signals can be decoded by applying SIC.Therefore,the signal-to-interference-plusnoise ratio (SINR) between SBSmand useruis expressed as follow:

wherepmuis the transmission power of SBSmto useru,σ2is the variance of additive white Gaussian noise(AWGN),and|gmu|2is the channel gain for SBSmconnecting to usern.

We assume thatBis the total frequency band,andbmudenote the occupied frequency band of useruserved by SBSm.The users served bym-th SBS can share the same frequency band to request the video segments by utilizing NOMA,Thus,the transmission ratermubetween SBSmand UEucan be given by:

Letdenote the maximum frequency band occupied by SBSmfor serving its users.Since the allocated frequency band is not more than the total bandwidth in wireless access links,corresponding bandwidth constraints can be formulated as:

Moreover,the power that the SBSmassigns to the users should be less than it’s maximum powerpm,max.Then the power constraint is formulated as:

2.2 Caching Model

In this network scenario,we assume that the maximum storage capacity of MEC server inm-th SBS isZm.The caching strategies can be denoted asA=,whereindicates whether cache the video segmentfwith versionlby SBSm.And=1 represents that the video segmentfwith versionlis cached in them-th SBS’s MEC server,otherwise,=0.

For simplicity,we suppose that each video segment containstseconds in playtime.Hence,the size of segmentfwith versionldenotes astRfl(bytes),whereRflis the bitrate of the segmentfwith the versionl.

The total size of video segments cached in each SBS’s MEC server cannot exceed it’s maximum caching capacityZm,which is expressed by the following formula:

Moreover,for segmentf,we use0,1}to show whether there are higher bitrate versions than versionlcached inm-th SBS’s MEC server.And=1 represents that the video segmentfwith higher bitrate versions than versionlare cached atmth SBS’s MEC server;otherwise,=0.Thus,andshould comply with the following constraints:

2.3 Computing Model

We letdmudenote the computing ability that MEC servermassigned to UEu.So the computing resource allocation strategies can be denoted asD={dmu |m ∈M,u ∈U}.The computing ability allocated to users by the one MEC server cannot exceed the max computing ability owned by the MEC server itself.So,it satisfies:

whereDmdenotes the maximum computing capability of them-th SBS.

Takef-th segment for example,the smaller the bitrate difference value between the higher version and the desired version,the less time it is likely to take to process the transcoding task.We assume that the versionl+is the version with the smallest bitrate difference value from versionl,where=min=1}.We consider that the bitrate of thel-th segment with versionlisRfl,andRfl+ηmfldenotes the bitrate of thef-th segment with versionl+

Thus,the transcoding delay that the MEC server transcodes thel+-th version into thel-th version can be denoted as:

2.4 Delay Model

Due to partial video segments are cached at the servers of SBSs,there are three possible cases to provide video segments to the users,including direct hit case,transcoding hit case,cache miss case.Next,we will introduce the three video segment delivery ways and the delay of these ways.

Direct hit case:When the required bitrate version of the video segment exactly exists in the MEC server,the users can obtain the video segment from MEC servers directly.

In the direct hit case,when them-th SBS’s MEC server transmit the versionlof the segmentftouth UE,there is only transmission delay.From what has been discussed above,we can get the formula of transmission delay as follows:

thus,the total delay of the direct hit model can be denoted as:

Transcoding hit case:If the MEC server caches the higher bitrate version of the video segment instead of the required bitrate version,the MEC servers can choose to transcode the cached higher bitrate version into the requested bitrate version or to request the required bitrate version from the remote Internet server,which depends on the transcoding strategy.

Therefore,arranging a transcoding strategy is a sensible way to reduce the delay.We usecmfl ∈{0,1}to indicate whether the MEC server of SBSmto obtain the segmentfwith versionlby transcoding.

Transcoding strategy is effective only when the higher bitrate versions of the requested video segment are cached in the MEC server,so,cmfl ≤hmfl.

When the MEC servers choose to transcode,the total transcoding delay that them-th SBS deliver thelth version off-th segment to theu-th user includes transmission delay and transcoding delay,which can be denoted as:

When the MEC servers choose not to transcode,the total delay is equal to the total delay of the cache miss model like follow.

Cache miss case:If the MEC servers cache neither the exact bitrate version nor the higher bitrate version of the requested video segment,it will respond users by getting the video segment with the requested bitrate version from the remote cloud.That will bring great back-haul delay,which can be denoted as:

Therefore,the total cache miss delay that them-th SBS deliver thel-th version off-th segment to theu-th user includes transmission delay and back-haul delay,which can be denoted as:

III.PROBLEM FORMULATION

In order to achieve efficient transcoding and provide viewers with low-latency video streaming services,we transform the optimization problem into federated access decisions,bandwidth allocation,caching placement,and transcoding strategy problems.

According to the above analysis,we denote the total delay for theu-th user obtaining thel-th version of thef-th segment fromm-th SBS as:

Both the popularity of the segment and the users’preference of different versions are taken into account in our network scenario.In general,the video segment setFis sorted in descending order of popularity.qfis the probability of users requesting video segmentf,which follows the Zipf distribution.

Different users have different preferences for different versions of the video segment.The bitrate setLis sorted in ascending order of popularity.We letjufldenote theu-th user’s preference of the segmentfwith bitrate versionl.Therefore,the probability of the userurequesting for the segmentfwith the bitrate versionlcan be denoted as follows:

whereγfis the Zipf parameter,which is related to different video segments.

To minimize the total delay of the content delivery,we now formulate the optimization problem as follows:

ConstraintC1 ensures that one user can only access one SBS at the same time.ConstraintC2 ensures that the spectrum resource allocated to the SBSs is less than the network’s total spectrum resource.ConstraintC3 ensures that the computing resource allocated to the users is less than the maximum computing capability of each MEC servers.ConstraintC4 ensures the power that the SBS assigns to the users should be less than it’s maximum power.ConstraintC5 ensures that the caching resource allocated to the users is less than the maximum storage capacity of each MEC server.ConstraintC6 ensures that the video segment’s higher version has been cached in the MEC server when performed the transcoding strategies.Tfuis the maximum tolerable of useruto request video segmentf,constraintC7 guarantees that the delay for the users to get the video segmentfis less than his maximum tolerance delay.

IV.SOLUTION

In order to solve the original problem efficiently,this section transforms the original problem into a convex optimization problem and proposes a distributed algorithm based on the Alternating Direction Method of Multipliers (ADMM) and Simulated Annealing (SA)to solve the problem.

4.1 Problem Reformulation

Firstly,we relax the binary variablexmuto the continuous variable,that isxmu ∈[0,1].Considering the interference between the users served by the same SBS,the problem is still difficult to solve.In order to reduce the complexity of the problem (17),an additional interference constraint is introduced,which is expressed as follows:

whereIm,thrdenotes the maximum intra-cell interference of any user served by SBSm.In general,adding a constraint tends to cause the feasible region to shrink.By adjusting the value ofIm,thr,SBS can control the interference in its cell to improve system performance.

SubstitutingIm,thrinto the user’s data transfer rate formula(3),we derivethus,Based on the revised transmission rate,an estimate of the transmitted power can be obtained as below:

wherermu,maxis the maximum available transmission rate of useruserved by SBSm,bm,minis the minimum available frequency band of SBSm.

Then,we definexmubmu=xmudmu=due to the loss of definition whenxmu=0,it is not a one to one mapping.Whenxmu=0,we can makebmu=0,dmu=0,to ensure one-to-one mapping during the variable substitution process between{xmu,bmu,dmu}and{xmu,},takingbmufor example:

Thus,we transform the original problem into the following problem:

In addition,we prove the convexity of the optimization problem based on the variable substitution above.Since the structures likeare well-known convex function(Quadratic-linear fraction function),the optimization problem can be proved to be a convex problem.takingfor example,where the variables can be converted to the following formSo,we can conclude that under the fixing caching and transcoding strategies (AandC),the form of the objective function in(21)is a linear sum of convex problem.Therefore,the optimization problem is a convex problem.

4.2 Distributed Solution for Joint Access Control and Resource Allocation

Since the variables{xmu}andin the problem can affect all SBSs,they are global variables,which are not independent in the problem (21).To apply ADMM to this problem,introduces the local copy of the global variables and obtain a distributed feasible solution on each SBS by decoupling the original problem.

For SBSm,we introduce the new variables,={| e ∈M,m ∈M,u ∈U}and={|e ∈M,m ∈M,u ∈U}as the local information,thus we can get:

As the local variation of the problem,={|u ∈U}represents computing resource allocation of the SBSm.So,we define Φm=() to represent the feasible local variables of the SBSm,and define Ω to represent the constraint set of the new objective function.

We introduce the penalty functionϖ(Φm),if the Φmbelongs to the constraint set Ω of the new objective function Φm ∈Ω,we can getϖ(Φm)=0.Otherwise,ϖ(Φm)=+∞.With these notations above,the problem(21)is equivalent to:

The objective function is separable between SBSs,and each SBS can determine its local variables by using the local information.The Lagrangian of the augmented problem by utilizing ADMM is given by:

whereαm={αemu}andβm={βemu}are the vectors of the Lagrange multipliers,and the penalty parameter isR++.In order to solve the above problems (25),we introduce the ADMM algorithm.Next we will introduce local variables,global variables,and Lagrange multipliers respectively.

Local variables:

whereιdenotes the iteration times of the ADMM process.

Since the updating process of Φmof each SBS is independent,and we can decouple the problem intoMindependent subproblems.We can update the local variables by solving the problem as follow:

Since the above problem is convex,we can solve it by using a convex optimization algorithm.Then the decision of each SBS can be broadcasted to other SBSs.

Global variables:

Because the quadratic regular term is added in the augmented Lagrangian,the above problems are strictly convex and unconstrained quadratic problems.Let the gradient of x and ¯b to zero.we can get the following results:

We can derive:

Lagrange multipliers:

After receiving the updated local variables{Φm},and global variables{x}andfor each SBS,we can obtain the Lagrange multipliers by calculating directly at each iteration as follows.

When the caching variablesand transcoding variablesare fixed,the original problem is a convex problem with strong duality.When the number of iterations tends to infinity,the proposed ADMM based algorithm satisfies the convergence.Thus,the corresponding rational stopping criterion is given as follows:

whereξpri >0 andξdual >0 indicate the primal feasibility and dual feasibility conditions.

Since we have relaxed the binary variablesxmuinto continuous variables,we need to recover the user access controlxmuafter solve the above problem.

Letιmaxdenote the maximum number of iterations,the iteration process of ADMM-based solution algorithm is concluded in Algorithm 1.

?

4.3 Solution for Joint Caching Decision and Transcoding Strategies

After fixing the feasible strategiesX,D,andB,we reduce the problem Γ2 to the 0,1 planning problem with variablesAandCas follows:

Next,we will address the problem (43) to reduce total user latency.Problem (43) is a NP-hard problem,which has enormous complexity at the aspect of computation.So,we propose a scheme based on the simulated annealing(SA)algorithm to obtain the best solutionAandC.

The simulated annealing algorithm is a probabilistic algorithm based on the principle of solid annealing,which can be used to solve combinatorial optimization problems.By combining our optimization problem with the physical annealing process,we conclude the solving process of our problem based on SA as follows:

Firstly,generating the random initial feasible solutionAandC,setting max temper,and then computing the value of the objective function under the fixedX,BandD.

Secondly,generating the neighboring solutionA∗andC∗,and calculate the value of the objective function by them.

Thirdly,calculate the difference between new total delay and current total delay ∆T=T∗−T,if ∆T>0 ore−∆T/ι1>rand,the new neighboring solution is accepted.Otherwise,the new solution is not accepted.

Letι1maxdenote the maximum number of iterations,the iteration process of SA-based solution algorithm is concluded in Algorithm 2.

?

4.4 Two-stage Overall Iterative Algorithm

There are five optimization variables in the original problem,so it has high algorithm complexity.In order to reduce the algorithm complexity and get the optimal solution of the original problem,we solve the original problem in two steps.The algorithm 3 can be summarized as:

?

4.5 Algorithm Complexity Analysis

In the network scenario considered in this paper,there areMSBSs,Uusers,andFvideo segments.Each video segment hasLbitrate versions.Then,there are(4MU+2MFL)variables in the optimization problem Γ1.If the centralized algorithm is adopted,the algorithm complexity of finding the global optimal solution isO((4MU+2MFL)ϕ) withϕ >0,whereϕ=1 means a linear algorithm,ϕ >1 denotes a polynomial time algorithm.When the caching and transcoding strategies update,the whole problem can be broken down into solving local optimization problems (27) at each SBS by using ADMM algorithm,which has a maximum problem size ofU,and its algorithm complexity isO((U)ϕ) withϕ >0.In the global variables updating,wherekΨis the number of iterations required for updating global variables,the algorithm complexity isM × U × kΨ.In the Lagrangian multiplier updating,we usekιas the number of iterations required in problem(36)and(37),so the computational complexity isU×kι.In the caching and transcoding process,wherekι1is the number of iterations required for updating caching and transcoding strategies,the algorithm complexity isM×F×L×kι1.Thus,the sum complexity in each iteration isO(Uϕ+M×U×kΨ+U×kι+M×F×L×kι1)=O((U)ϕ,withϕ >0.Assuming thatkι2is the number of iterations required for algorithm convergence,the total computational complexity isO(kι2(U)ϕ).Therefore,compared with the centralized algorithm,the proposed distributed algorithm has lower complexity.

V.SIMULATION RESULTS AND DISCUSSIONS

This section provides numerical results to verify the performance of our proposed model.We assume that there areM=10 SBSs andU=15 UEs randomly deployed in a 120×120m2area.The number of video segments isF=1000,and each segment hasL=4 bitrate version.Moreover,we set the caching capacity relative to the total size of the video library is 10%-60%.

To evaluate the performance of our proposed approach,we selected the following four typical schemes for comparison:(1) The SBSs no caching the video segments,users can only request video segments from the remote Internet servers,as labeled as “Without caching”;(2)The SBSs cache some video segments of different bitrate versions,but cannot transcode these bitrate versions into another bitrate versions for the same segment on the fly,as labeled as “Without transcoding”,namely,users can only request video segments from the remote Internet server or MEC servers that cache the requested video segment;(3)The SBSs will carry out transcoding as long as the edge servers cache the higher bitrate version than the request version for the same video segment,regardless of whether transcoding will cause higher delay,as labeled as“Without transcoding strategy”;(4)The same proposed approach with orthogonal frequency division multiplexing(OFDM),as labeled as“OFDM”.

As Figure2 shows,for all schemes,with the iteration increasing,the total delay decreases fast in the beginnings,and then it shows a convergence and remains at an almost constant value.In addition,we can see from the iteration diagram that compared with other schemes,and our proposed scheme keeps the lowest total delay during the iteration.

If we do not cache video segments on the edge server,we do not need to consider caching and transcoding strategies.So,The “Without caching”curve is not considered in Figure3 and Figure4.

Figure3 illustrates the total delay versus the MEC servers’ capability of computing.With the increase of computing capability,the users’ delay in obtaining the video segment is decreased greatly.This is mainly because the more computing resources are allocated to the user,the less delay is required for the user to transcode.Meantime,it is noted that the total delay of the scheme “Without transcoding strategies”is longer than the total delay of other schemes with transcoding strategies obviously.Thus,comparing to other schemes,the proposed approach has a lower total delay of users.

Figure2.Total delay versus iteration times.

On the other hand,as shown in Figure4,With the increasing cache size of SBSs,the users’delay in obtaining the video segment are decreased greatly.This is mainly because the larger the cache size of the server,the more video segments can be cached on the server,thus avoiding the transmission delay between the remote Internet server and the SBS.

Figure3.Total delay versus computing capability.

Figure4.Total delay versus caching capability.

Figure5 illustrates the total delay versus the transmission power of the SBS.With the increase of transmission power,the users’delay in obtaining the video segment is decreased greatly.This is mainly because the higher the transmission power of the base station,the higher the transmission rate between the base station and the user,and the smaller the transmission delay.Furthermore,comparing to other works,the proposed scheme has a lower delay,seen from the simulation.

Figure6 illustrates the total delay versus the transmission rate between the remote Internet server and SBSs.With the increase of the transmission rate between the remote Internet server and SBSs,the users’delay of all schemes is decreased.When the transmission rate between the remote Internet server and SBSs is high,it is possible to obtain video segments from the remote Internet server with a smaller delay than video transcoding.

Figure5.Total delay versus the transmission power of SBS.

Figure6.Total delay versus the backhaul capability of the SBS.

Figure7 illustrates the total delay versus the total bandwidth.With the increase of the total bandwidth,the users’delay in obtaining the video segment is decreased greatly.This can be accounted for:when the total bandwidth is high,the transmission delay between UEs and SBSs will descend.

Figure7.Total delay versus the total bandwidth.

In Figure8,we compared our proposed algorithm with the algorithms in the paper[3]and the paper[16],our algorithm has the lowest delay.In a note,when the MEC server’s computing resources are small,the flexible transcoding strategy we proposed can better allocate computing resources,so the delay decreases faster with the increase of computing resources.

Figure8.Total delay versus computing capability.

VI.CONCLUSION

In this paper,we propose a MEC-assisted flexible transcoding strategy to achieve adaptive bitrate video streaming.Specifically,we cache the proper bitrate version of the video segments at the edge servers and select the appropriate bitrate version of the video segments to perform transcoding under joint considering access control,resource allocation,and user preferences.In addition,access control and user preference are taken into account.In the future works,we will cooperatively take into account the access control,transcoding strategy,and the effective content caching placement for video streaming to optimize the QoE of users and improve the network performance.

ACKNOWLEDGEMENT

This work was supported by National Natural Science Foundation of China (No.61771070) and National Natural Science Foundation of China(No.61671088).