Adaptive Bitrate Video Streaming for Wireless nodes: A Survey
11 Adaptive Bitrate Video Streaming forWireless nodes: A Survey
Kamran Nishat,Omprakash Gnawali, Senior Member, IEEE and Ahmed Abdelhadi, SeniorMember, IEEE
Abstract —In today’s Internet, video is themost dominant application and in addition tothis, wireless networks such as WiFi, Cellu-lar, and Bluetooth have become ubiquitous.Hence, most of the Internet traffic is videoover wireless nodes. There is a plethora of re-search to improve video streaming to achievehigh Quality of Experience (QoE) over the In-ternet. Many of them focus on wireless nodes.Recent measurement studies often show QoEof video suffers in many wireless clients overthe Internet. Recently, many research papershave presented models and schemes to op-timize the Adaptive BitRate (ABR) basedvideo streaming for wireless and mobile users.In this survey, we present a comprehensiveoverview of recent work in the area of Internetvideo specially designed for wireless network.Recent research has suggested that there aresome new challenges added by the connec-tivity of clients through wireless. Also thesechallenges become more difficult to handlewhen these nodes are mobile. This surveyalso discusses new potential areas of futureresearch due to the increasing scarcity of wire-less spectrum.
Index Terms —Video, Wireless, WiFi, Cel-lular, Spectrum Sharing.
I. Introduction V IDEO is the most frequent type of trafficon today’s Internet [1, 2]. It is importantfor services like Youtube, Netflix, and Facebookto deliver a high Quality of Experience (QoE)
Kamran Nishat, Omprakash Gnawali and AhmedAbdelhadi are with the University of Houston,Houston, TX 77004, USA (e-mail:[email protected],[email protected], [email protected])Manuscript received XXX during video streaming to sustain revenues [3]and user engagement [4]. Most Internet videodelivery services like Twitch, Vimeo, Youtubeuse Adaptive BitRate (ABR) to deliver high-quality video across diverse network conditions.Many different types of ABR are implementedin recent years [5, 6] to optimize the quality ofthe video based on different inputs like availablebandwidth and delay. Recently, Pensieve, a neu-ral adaptive video streaming platform developedby MIT [5], it uses deep reinforcement learning(DRL) [7, 8], and outperforms existing ABRs.One of the major challenges will occur inthe near future with 5G wide deployment whenmany devices share the unlicensed spectrum,such as [9–11] .Video stream applications canbe optimized for these resource critical scenar-ios with the introduction of edge device basedfeedback to the Reinforcement Learning (RL)based ABR running on the server. These edgedevices will collect data by spectrum sensing andthen allocate the spectrum for the next timeslot. This allocation will be transmitted to theABR server in addition to the feedback from themobile client.
A. Motivation: Why we need ABR solutions forWireless Networks?
Video streaming over wireless/mobile nodesnow accounts for more than 70% of Internettraffic, and it is still growing with a phenomenalrate [1]. Massive deployments of LTE basedcellular networks has also played a vital role inthis. LTE supports peak down-link bitrate of 300Mbps, almost 10 times more than over 3G [12]. a r X i v : . [ c s . N I] J u l TABLE I
LIST OF COMMONLY USED ACRONYMS IN THIS PAPER
Acronym ExplanationDASH Dynamic Adaptive Streaming over HTTPABR Adaptive BitRateQoE Quality of ExperienceDL Deep LearningRL Reinforcement Learning5G 5 th Generation mobile networksLTE Long-Term EvolutionMIMO Multi-Input Multi-OutputIoT Internet of ThingsHTTP HyperText Transfer ProtocolDRL Deep Reinforcement LearningMAC Media Access ControlMDP Markov Decision ProcessDQL Deep Q-LearningCNN Convolutional Neural NetworksOFDMA Orthogonal Frequency-Division Multiple AccessOFDM Orthogonal Frequency-Division MultiplexingHAS HTTP Adaptive StreamingPoPs Point-of-PresenceCDN Content Delivery NetworkMPTCP Multi-Path TCPEC Edge ComputingSILP Stochastic Integer Linear ProgramMEC Mobile Edge ComputingMNOs Mobile Network OperatorsKPIs Key Performance IndicatorsDNN Deep Neural NetworkSDN Software Defined NetworksDRNN Deep Recurrent Neural Network
However, most of the studies show QoE is stillunsatisfactory [13].New applications of video over mobile clientare getting popular [14]. Exponential growth ofIoT based networks will increase these inno-vative scenarios, with the applications like on-line object detection [15, 16] and energy efficientscheduling [17]. Many new applications applymachine learning algorithms like deep learning[18] on video streams on resource constraint mo-bile devices. These applications introduce newchallenges and opportunities for Internet videoecosystem.
B. Prior Survey Articles
Having established the importance of ABR al-gorithms optimizing QoE for wireless and mobileclients in particular, in this paper, we are review-ing existing models and algorithms in this area. While, there exist previous surveys, in the areaof Internet video and optimizing applications forwireless networks in our opinion there are nonewhich focuses on mobile video streaming algo-rithms. Previous surveys like Seufert et al. [19]and Bentaleb et al. [20] discuss different ABRalgorithms in general and related influence fac-tors. In another survey by Juluri et al. [21], theydiscussed tools and measurement methodologiesfor predicting QoE of online video streamingservices. Similarly in [22], authors provide a sur-vey of QoE models for ABR applications. Kuaet al. [23] focuses on rate adaptation methodsfor Internet video in general, provides a com-prehensive review of video traffic measurementmethods and a set of characterization studiesfor well-known commercial streaming providerslike Netflix, YouTube, and Akamai. The surveyin [24] discusses the growing popularity of deep learning (DL) based techniques to solve differentwireless network problems. They discuss theapplications of DL methods for different layersof the network, but do not include Internetvideo and its challenges in particular. Seufertet al. in [25] focused on video quality metricsand measurement approaches that are relatedto HTTP based adaptive streaming. Similarly,Barakabitze et al. [26] focused on techniques ofmaintaining QoE in emerging types of networksbased on SDNs and NFVs. They do discussQoE for multimedia application in LTE and5G networks but more with the context andopportunities related to SDN/NFV.Bentaleb et. al. in their survey [27] describesa many recent paper related to ABR in detail.Their main focus is a scheme classification basedon the unique features of the adaptation logic ofABR algorithms.Our survey is unique from others in threekey aspects: (1) It is focused on clients con-nected to the internet using wireless technologieslike WiFi or cellular network. In all previoussurveys none focused on different schemes andtheir challenges of designing ABR specificallyfor wireless networks. (2) We provide an in-dept survey of schemes using machine learning ingeneral and RL in particular to optimize QoE ofvideo for wireless nodes. (3) We provide manyopen challenges in designing ABR for futurewireless networks.
C. Contributions of This Survey Article
In this article, a comprehensive survey on theproposed ABR algorithms for wireless networksis presented. The contributions of this reviewpaper are summarized as follows. Towards thisend we present in this paper a review of theproposed ABR algorithms for wireless networks. • We present and classify the existing worksrelated to ABR for wireless networks. In thispaper, we provide an overview of the currentstate of the art in the field of Internet videoin wireless networks. • We identify some important directions of fu-ture research. We present some area where upcoming new standards and their adapta-tions will create challenges for existing ABRalgorithms. We discuss suggestions for futuredesign of ABR algorithms. • We review many open-source implementationsof different ABR algorithms. And we presenttheir differences and comparisons.A list of acronyms used throughout the articleis presented in Table I. The rest of this paperis organized as follows. Section II presents theoverview of the Internet video delivery ecosys-tem. It also introduces basics of machine learn-ing techniques used in the papers discussed inthis survey. Section III surveys the bitrate adap-tation algorithms for wireless networks. This sec-tion is divided in different subsection accordingto the type of algorithms. Section IV presentsdifferent open challenges in the area of ABRfor wireless. In the section V discusses differentopen-source implementations of ABR algorithmsand also different dataset available for experi-ments. Finally, Section VI provides concludingremarks.
II. Wireless ABR Video Streaming: AnOverview
A. Why video on wireless is different?
Internet video systems are designed to copewith the inherent variability in network con-ditions. Media players at the client implementABR algorithms [9, 28, 29]. There are a vari-ety of protocols like MPEG-DASH [30], Ap-ple HLS [31], Microsoft Smooth Streaming [32],Adobe HDS [33] etc. that adopt HTTP basedadaptive video streaming. These protocols arecalled Dynamic Adaptive Streaming over HTTP(DASH). In these schemes, server splits eachvideo into multiple segments with uniform play-back time (typically 1 to 10 seconds). After-wards, the server encodes these segments intomultiple copies with different discrete encod-ing bitrate levels having different sizes. Beforea DASH video session starts, a client obtainsthe available bitrate map from the server. Todownload each segment, the client needs to send an HTTP request to the server, and specify thebitrate level it prefers for that segment.Most of the content publishers in today’sInternet serve their videos from some popu-lar content delivery networks (CDNs). TheseCDNs have point-of-presence (PoPs) in manydifferent geographical and network domains. Byusing their PoPs and their peers, CDNs reducethe cost of serving videos and join times, eachvideo is delivered over an ISP network. Mostof these Internet service providers (ISPs) havetwo parts, the core network and the radio-accessnetwork (e.g., cellular network) as shown inFigure 1. User devices are connected with theISP via wireless technology like WiFi or LTE.However, these solutions render unsatisfactoryperformance in WiFi or LTE networks.ABR algorithms work by (a) chopping thevideo into chunks, each of which is available ata range of bit rates; then (b) choosing which bitrate to fetch a chunk at based on conditionssuch as the amount of video the client hasbuffered and the recent throughput of the net-work. These ABR algorithms are implementedin video clients. Hence, on a mobile client theymust be energy efficient in addition to all otherproperties like computational efficiency and op-timize QoE for the user.One of the first ABR algorithms was de-signed in model predictive control (MPC) [34]. Itpredicts throughput of future chunk downloadsusing the historic data of recently downloadedchunks. The predicted value of throughput isused to select the bitrate for the future chunkssuch that optimizes the QoE function. MPC hasa look-ahead window of 5 future chunks. Thereis an aggressive version of this algorithm calledFastMPC which directly uses the throughputestimate obtained using a harmonic mean pre-dictor.On the other hand there are algorithms likeBuffer Occupancy based Lyapunov Algorithm(BOLA) [35] which uses buffer occupancy toselects bitrate. BOLA solves an optimizationproblem to select optimal bitrate. BOLA is abuffer-based algorithm used in Dash.js [36]. In contrast to MPC, it does not employ throughputprediction in making decisions. It tries to avoidre-buffering by maintaining a minimum bufferthreshold. This threshold can be used to makethis algorithm conservative or optimistic aboutthe future bitrates.Recently, RL and other machine learning tech-niques are used to design ABR algorithms [37–39]. Using RL Pensieve [5] was able to outper-form the state-of-the-art. Oboe [6] presented anABR algorithm which performs an automatictuning of configuration parameter values foreach network state independently. This allowsOboe to give better QoE than Pensive.
B. Why Reinforcement Learning is popular forABR design?
There are many schemes based on learning-based approach to solve problems in networks ingeneral [40]. Some of them are focused towardsapplying RL. In the past, it has been noted thatRL is very suitable to be applied to many com-puter network problems. RL is quite a naturalway to model an optimization control problem[41].There are two main entities in a RL problemFigure 2,
Agent and
Environment . Agent ob-serves the state s i of the environment at eachinterval and then choose an action a i . Agentreceives a reward r i for his action which can bepositive or negative. The goal of an agent in RLis to maximize the cumulative reward defined asfollows: V π ( s ) = E " ∞ X i =0 γ i r i | s = s Where π is the policy function π i ( s i , a i ). Itgives the probability distribution of the currentstate and action. While the γ is the discountfactor for the future reward. Hence, an RL agentlearns optimal policy to maximize its rewards.To learn the optimal policy most of the RLapplications use Q-learning.In Q-Learning, each pair of state and action( s, a ) is mapped to a value under a policy π . This Fig. 1. Mobile video player architecture value is the expected total reward of taking anaction a in a state s . Q π ( s, a ) = E " ∞ X i =0 γ i R ( s t , π ( s t )) | s = s, a = a The goal of Q-learning algorithm is to findthe policy to maximize this function. In DRL anagent is learning this optimal policy using a deepneural network (DNN). So, in DRL approximatevalue functions called deep Q learning functionis used. This function is learned by gradientmethod used in deep learning. Here the agentinteracts with the environment like in RL anduses its reward as the training input for the deepneural network. The goal during training of thisDNN is to optimize its parameters. Hence, itselects actions that can result in the best futurereturn.Pensive [5] was the first paper to use DRL indesigning ABR. Pensive model it using DRL soit can be independent of the assumptions takenby the designer of ABR schemes. In Pensivealgorithm, they defined the QoE as the rewardof the ABR algorithm working as an agent inthe DRL. They modeled their generic QoE basedreward function as follows.
Fig. 2. Overview of Deep Reinforcement Learning
QoE = n X i =1 q ( R i ) − µ i X i =1 T i − n − X i =1 | q ( R i +1 ) − q ( R i ) | Where the function q is an increasing functionof bitrate selected R i for the interval i so,higher the bitrate higher the reward. The secondterm depends on the time taken to buffer thatsegment T i . This will penalize the reward forany re-buffering required before playing the nextsegment. Last segment of the reward function ispenalizing for lack of smoothness. If the bitrateof the video change from the previous segment, then user will observe some lack in smoothness.Pensive [5] is used in many similar research fordifferent variants of the problem. There are twomain types of the RL based on the training. Firstis model based RL and the other is model freeRL.
1) Model-free Reinforcement Learning:
Model-free RL learns directly from theexperiences while in training. The statesand transition probabilities of the underlingMarkov decision process (MDP) are unknown.In ABR, we have no prior model of QoEdependence on the different state variables.Model-free RL learns in more interaction withthe environment as compared to the model-based RL Figure 3. It is free from the biases ofthe supervised training data.
2) Model-based Reinforcement Learning:
Inthis type of RL we have a prior model ofthe system MDP. This will reduce the time oflearning and cost in-terms of interactions withthe environment, on the other hand, it requiredesigner to provide or learn the model beforethe start of the training. RL training phase willonly optimize or refine this model. In this case, ifthere are inaccuracies in the model, then this willlead to degradation in quality of final output.In model-based RL, policy is learned throughsupervised learning. Then planning over thelearned model is done in second phase. Conse-quently, model-based algorithm uses a reducednumber of interactions with the real environ-ment during the learning phase 3. So, learningcan be much faster because there is no need toget the feedback from the environment. On thedownside, however, if the model is inaccurate,we risk learning something completely differentfrom the reality.
III. Different types of ABR algorithms
There are different types of research in thisarea. Some have designed ABR algorithms formobile nodes to incorporate the movement ofthe nodes, while others focused on the resourceconstraints like spectrum and energy of the videoclient.
Fig. 3. Model-free and Model-based ReinforcementLearning
Mobile Network Operators (MNOs) offer dif-ferent packages to increase their number ofusers. In some packages, they offer not to countcertain services like Facebook, WhatsApp andNetflix toward monthly data quota. But theylimit the rate of the users toward those ser-vices. Traditional ABR based services does notaccount for this rate limiting. Here traditionalthroughput maximization based ABRs will notperform well. Zero-rated QoE [42] proposed anovel approach which uses the collaboration ofthe content provider and MNOs. They designedan ABR which focuses on improving QoE inthese special scenarios. They implemented theirapproach in a simulated environment and per-formed evaluation with the baseline.
A. Machine Learning based ABR schemes
Comyco [43] is another study using theLearning-based ABR. They discuss a few weak-nesses of previous RL-based ABR algorithms.Their measurement study shows that the qualityof video presentations is not always maximizedby QoE metrics based on only video bitrates, re-buffering times and video smoothness. They pro-posed
Imitation Learning instead of supervisedlearning can address these weaknesses.Reinforcement learning is also applied to makevideo streaming appear more smooth to the user.
TABLE II
A SUMMARY OF RELATED PAPERS AND MAIN IDEA.
Paper Main Idea Is ABR? For mobile?
FESTIVE [34] Stateful ABR with randomized chunk scheduling to avoid synchronization biases (cid:55) (cid:88)
Pensive [5] DRL is used to optimize the QoE (cid:55) (cid:88)
Oboe [6] Combine offline and online tunning of the parameters (cid:55) (cid:88) pstream [12] Improves the QoE by taking advantage of the PHY information of LTE networks (cid:88) (cid:88)
MP-DASH [44] Use MPTCP to schedule (cid:88) (cid:88)
Wi-Fi Goes to Town [45] Improves the QoE during high speed handovers (cid:88) (cid:55)
HotDASH [46] Use DRL to detect user specific important part of video to improve their quality (cid:55) (cid:88)
Zero-rated QoE [42] Collaboration of MNOs and content providers to improve QoE for rate limited users (cid:88) (cid:88)
QARC [47] Imrove the perceptual quality of the video instead of traditional QoE metrics (cid:55) (cid:88)
Bursttracker [13] Find the bottleneck in the video streaming over the LTE network (cid:88) (cid:55)
Qflow [48] Used both Model based and Model free DRL (cid:88) (cid:88)
NAS [49] A deep neural network based ABR (cid:88) (cid:88)
IncorpPred [50] Incorporate cellular throughput prediction to improve ABR (cid:88) (cid:88)
Comyco [43] Proposed imitation based Learning instead of supervised learning (cid:55) (cid:88)
Jigsaw [51] 4K video streaming (cid:88) (cid:55)
TransPi [52] Introduced hardware-assisted video transcoding for Wireless (cid:88) (cid:88)
CASTLE [53] Client schedular to minimizes Load and Energy at the same time (cid:88) (cid:88)
ACAA [54] Incorporate user’s subjective viewing information to improve ABR (cid:55) (cid:88)
LinkForecast [55] Bandwidth prediction for LTE network (cid:55) (cid:88)
QUAD [56] Reduce the bandwidth usage while maintaining high QoE (cid:88) (cid:88)
In [57], the authors design an optimizationproblem for ABR to exploit power control overmultiple sub channels at the transmitter in sucha way that video quality remains smooth.It pe-nalizes both for buffer underflow and overflow.Then, they mapped this constraint optimizationproblem into a MDP. The MDP is solved usingreinforcement learning techniques.It is challenging to implement heavyweightARB techniques in resource constraint mobiledevices. PiTree [58] introduced the idea of us-ing lightweight decision trees to simplify thecomplex and heavyweight neural network basedtechniques. PiTree give a highly scalable frame-work to convert complex ABRs into decisiontrees. They also provide some theoretical upperbound on the optimization loss during the con-version.Challenges of Video streams with high qualityare increased in the case of remote drone pilot-ing. The study in [59] discusses these challenges.It suggests decreasing the coupling of differentfunctional blocks. They proposed to use edge-computing elements in addition to adapting fornetwork conditions.QFlow [60] paper used reinforcement learningto perform one-way adaptive flow prioritizationat the edge network. QFlow argued currentlink are application agnostic in their schedul-ing. By making these links intelligently adaptfor different type of traffic leads to a betterQoE of the video streaming. QFlow borrowedconcepts of network level priority queues fromsoftware defined networks (SDNs) and applyit to PHY/MAC layer using Software-DefinedRadios.QFlow uses RL to optimize the QoE for videoby adapting configurations. QFlow uses bothmodel-free and model-based RL approaches.According to their evaluation, RL based ap-proach not only improves the QoE but alsoresults in better buffer state and lower stall dura-tion. The survey [40] provides a comprehensivesurvey of deep learning-based techniques usedin different wireless networking scenarios. It alsohighlights some potential applications of DL to networking, like in network security and userlocalization.In QARC [47] they have designed a rate con-trol algorithm that is focused on the perpetualquality of the video. The perpetual quality, ofthe video is defined as how many objects are inthe image and how bright or dark it is. For lowperceptual quality parts of the video we can savethe bandwidth and delay by requesting videoat the low quality. While for high perceptualquality, parts should be downloaded at a higherbitrate. This can be achieved by lower sendingrate and latency. QARC also use DRL to trainthe neural network to predict the future videoframes based on the perceptual quality of theprevious frames. It employs two-fold training ofDRL one for prediction of perceptual quality andthe second one using A3C based asynchronoustraining technique to train the actual RL algo-rithm. They did trace driven analysis of theirtechniques and compare it with Google Hangoutand compound TCP.
B. Egde computing based ABR systems
Increasing demands of lower network delayand higher data transmission rate are gettingdifficult meet from traditional ABR systems.Recently, edge computing (EC) Figure 4 basedoptimizations to ABR systems have been pro-posed to meet these challenges. In the paper [61],authors present some of the challenges and lim-itations of the current ABR applications. Theyproposed an edge computing based solution toaddress these challenges and limitations. QFlow[60] is an Edge computing based ABR systemusing ML based model to optimize QoE.ShareAR [62] is a multi-user augmented re-ality (AR) system which uses edge nodes tooptimize QoE for the user. The main challengein multi-user AR is the communication betweenAR platforms. There are no prior work involvingdata transmission in between AR devices andtheir impact of the QoE. In multi-user AR, de-vices can have different fields-of-view. They needto render their respective FoVs. In their system,they overcome these challenges and implemented
Fig. 4. Introduction of MEC to improve the ABR based video streaming a prototype of the system using two Androiddevices and an edge server.In FlexStream [63] they leveraged the SDNfunctionality to get the benefits of centralizedmanagement of distributed components. Herethey use wireless edge device like AP as SDNcontroller. They have implemented their systemas a light weight controller. In there evaluation,they showed FlexStream can achieve appropriatebandwidth distribution.Blockchain technology is used by decentral-ized peer-to-peer video streaming systems tomonetize using smart contracts. In these newvideo streaming systems, content creators, con-sumers and advertisers can communicate witheach other without the help of a trusted thirdparty. There some challenges in designing thesesystems like processing and publishing of thevideo content in these systems. In [64], theypropose using edge computing servers to of-fload these computationally intensive tasks.They proposed to employ edge servers throughdistributed block-chain based incentive mecha-nisms.Mobile Edge Computing (MEC) is gettingpopular to provide low-latency ABR service. One way to decrease the latency of the systemis to use MEC servers for video caching. Tranet al. [65] investigates a novel caching schemeusing multi-server MEC systems. Their systemsuse two timescales. They formulated stochasticinteger linear program (SILP) to integrate thesetwo timescales of long-term caching and short-term video retrieval mode. By using simulationsthey showed the effectiveness of their system byreducing access delay and increasing cache hitratio.PrivacyGuard [66] is a system designed anddeveloped to obfuscate the activities of sensitiveIoT and mobile applications from attacks overWiFi network. They have implemented a pro-totype this systems on Android mobile devicesthat to apply application level traffic shapingand IP-sec tunneling schemes.
C. ABR with different optimization goals
In Wi-Fi Goes to Town [45], implement aWiFi based hotspot network using picocell sizeaccess point networks along the road to sup-port vehicular communication over high speed.They implemented optimized version of IEEE802.11k and 802.11r standards. Although their main focus is not video streaming but most oftheir evaluation is done over video. Their schemeprovides more reliable video stream for highspeed mobile client. Also it improves the QoEmetrics for the video like rebuffed ratio.Sengupta et al. in HotDASH [46] focused onimproving the video quality for the specific userrequirements. In most of the video streamingsituations there is some content of the videowhich is more important for the user. Theyuse DRL to detect that part of the video andthen the requirement, therefore, is for a videostreaming strategy to into account the contentpreferences of the users. So, ABR will be awareof the high-priority temporal content. ABR tryto pre-fetch those high priority parts of the videoat much higher bitrate. HotDASH maximizesthe content preferences of users, in addition tooptimal use of bandwidth. They implementedtheir scheme in dash.js and compared it with thesix baseline algorithms like FESTIVE, FastMPCand PENSIVE.Most of the ABR algorithms are not designedwith the consideration of data consumption. Butmost cellular customers have limited data intheir monthly data plan. According to [67] aver-age U.S. cellular customer has only 2.5 GB permonth data plan, while one hour high definition(HD) video on mobile require 3 GB data. QUAD[56] focuses on reduce the bandwidth usage whilemaintaining high QoE for the user. Their schemeis also energy efficient because it requires todownload less amount of data.QUAD introduced a novel Chunk Based Fil-tering (CBF) approach which leverages two fun-damental tradeoffs of video quality and bitrates.First, higher bitrate leads to diminishing re-turn in terms of video quality. Second, dif-ferent chunks have different impact on videoquality. Their scheme selects chucks to maxi-mize the QoE while keeping the data consump-tion minimal. QUAD implemented its schemein both dash.js and ExoPlayer and performevaluations. They compared their approach withRobustMPC and PANDA.At the same time MP-DASH [44] take a dif- ferent approach to optimize video quality overmobile devices. Their focus is on leveraging theavailability of multi-path in many common mo-bile devices like cell phones. They have WiFicard and LTE modem at the same time. So, inmany cases it is possible to use LTE opportunis-tically. They used Multi-Path TCP (MPTCP) toimplement their approach. It prefers WiFi overLTE when at home. The evaluations were donein both controlled setting and in the wild. Traceof throughput and RTT of the WiFi networks at33 locations in the US for the evaluation. MP-DASH is impelled using GPAC. FESTIVE andBBA is implemented over the multi-path sched-uler. In [68] they implemented traffic offloadingbuild on the MP-DASH approach for generalapplication.ACAA [54] is a scheme focused towards se-mantic information of video content. RecentlyABR researchers are designing with user’s sub-jective viewing information to improve the QoEof the video specific to the user requirement.ACAA use the research on video affective con-tent analysis. It incorporated individual userpreferences into the bit-rate adaptation decisionsto improve the QoE. Identify the user relevantparts of the video and then assign bit-rate bud-get according to it. They compared their schemewith BBA and buffer-based adaptation (BBA),and model predictive control (MPC). ACAAimplemented their scheme with the DASH clientin accordance with [69] to perform trace-drivenevaluation platform with python 2.7. D. Measurement of different ABR schemes
Many papers study performance of differentABR algorithms and make a comparative study.Some of them performed active measurement[70] [71] while other perform passive measure-ments [72]. There are others like [73] and Puffer[74] made a database of different ABRs. Duaneet al. developed the Waterloo SQoE-III database[73]. This database provides a subjective evalu-ation of different QoE models and ABR algo-rithms. SQoE-III evaluated Rate-based, AIMD,Dynamic Adaptive Streaming algorithm, etc in their paper. According to their evaluation 5out of 6 models are quite close in terms ofperformance. In addition to the experimentalevaluation of other ABR techniques, Puffer [74]developed a live TV streaming website. Thisprototype website has attracted over 100,000users across the Internet. This system works as arandomized experiment; one set of ABR schemesis randomly assigned to each session.Haung et al. [71] is the first study performedin this area. In their study, they perform ameasurement study of three popular video ser-vices Hulu, Netflix, and Vudu. [75] studied anddiscussed the impact of QUIC on QoE of thepopular ABR. It also discussed how can existingABRs leverage the potential benefits of QUIC.The study in [76] is a general measurementstudy of application performance in the rapiddeployments of LTE networks. Data traces hasbeen used from different major LTE providers.VideoNOC [70] is an passive measurement studyof Video QoE for Mobile Network Operators(MNOs). VideoNOC presented an approach toassess the QoE for different MNOs using ob-jective metrics for video quality. To get an ob-jective estimate of QoE metric they collectedHTTP/S traffic in the core of the LTE network.VideoNOC performed many efficient and scal-able cross-layer analytics over these logs.Recently, a third-party based system to is de-signed to evaluate and understand the behaviorof different closed source ABR based streamingservices [77]. Channel State Information (CSI)can also able to understand the behavior ofABRs in the presence of traffic encryption. E. 4K and 360-degree video streaming
4K videos are now getting increasingly com-mon. New applications like virtual reality (VR)and augmented reality (AR) will make 4K ex-tremely important in the coming future. Theseapplications do not only require high resolu-tion but also very low latency. In there rawform 4K video stream requires more than 2Gpbsphysical data rate. Currently, IEEE 802.11adbased WiGig card are commodity wireless cards supporting these data rates. These devices workin 60GHz spectrum. In this range, transmis-sion is highly sensitive to mobility. There canbe drastic change in the throughput for minormovement. In the case of blocking, the line ofsight throughput might be affected and reachesto zero.In the presence of these large throughputvariations, traditional video codecs like H.264and HEVC become infeasible. To overcome theselimitations, layered video codecs are used byJigsaw [51]. It uses scalable video coding (SVC)which is an extension of the H.264 standard.The study in [51] use fast encoding schemes andimplement it using new layered video codingmethods.Panoramic video is another emerging appli-cation of video streaming, it is known as 360 ◦ video. Platforms like Facebook and YouTubealso support them. Flare [78], presented a practi-cal prototype of a 360 ◦ video streaming solutionusing commodity devices. This study predictsfuture behavior of the user to fetch only therelevant portions of the video to cover the viewof the user, which enables Flare to reduce thebandwidth usage of the system significantly.This viewport-adaptive 360 ◦ streaming is anestablish technique. Flare is the first completeworking implementation on commodity mobilephone. It uses a online machine learning (ML)algorithm to predict head movement of the userwhich changes the users’ future viewpoint.Another 360 ◦ video streaming system is pre-sented in Rubiks [79]. Rubiks discusses differ-ent challenges of implementing tile-based videostreaming techniques used in different imple-mentations to predict field of view (FoV) ofthe user. In resource constraints of commoditysmartphones, it is not possible to meet with therequirements these tile-based systems.Rubiks uses HEVC to implement tile-basedstreaming instead of H.264 [80] used by previoussystem. HEVC [81] has a built-in tiling schemeto encode video data. Their system can streamdifferent parts of the video at different bitrates.This allows them to download tiles in different quality according to their probability of viewing.Managing the amount of data downloaded at theclient it can control the decoding time. Decodingtime increases substantially for 8K videos onmobile device.In the paper [82], DRL is used to implementa panoramic video streaming system. Their sys-tem used DRL to optimize QoE using a broadset of features. Here their focus is on two mainchallenges of 360-degree video. First, there isa large number of time-variant features whichneeded to be adapted to achieve a reasonablequality. Second, QoE metrics are also differentfor different scenarios. Zhang et al. uses DRLto find an optimization model. This model findsthe best rate allocation scheme for differentscenarios.Tang et al. in [83] presented a promising ap-proach to improve QoE for the user in a 360-degree video streaming system. Their focus ison a streaming a newly generated 360-degreevideo. In this case, there is no historical view-ing information available which can be used topredict user viewing behavior. In these scenarios,there are additional challenges of learning FoVpatterns online and also the lengths of theseFoV segments are also unknown in advance. Theauthors present OBS360 algorithm which is anonline bitrate selection algorithm to optimizethe user’s QoE in 360-degree video. OBS360algorithm is able to learn user’s FoV preferenceand also the time-varying downloading capacityof the user.Perfecto et al. [84] they discussed Immer-sive virtual reality (VR) applications. Theseapplications require achieving motion-to-photon(MTP) delays, which are defined as end-to-endlatency of 15-20 milliseconds. Providing 360 ◦ video with these delay guarantees is quite achallenge. They applied a deep recurrent neuralnetwork (DRNN) to predict the upcoming tiledFoV. In addition to that, they exploit millimeterwave (mmWave) multicast transmission at thephysical layer to improve the efficiency of thesystem. F. Video stream in vehicular networks
Video streaming in vehicular networks is evenmore challenging [85]. There are different typesof communications in V2X networks. In recentyears, many papers try to address these chal-lenges [86, 87]. Some of them use Edge network-based caching techniques and other exploredlearning-based approaches to optimize videostreaming.Recent papers explore the use of differentmachine learning-based approaches to optimizevideo streaming in wireless networks [88–91].Some of them focus on the presence of IoTdevices on the same spectrum, others optimizefor energy-efficiency. In [92], they perform an ex-perimental analysis of 10 widely deployed ABRs.Their measurement shows none of the deployedABRs focus on available bandwidth and someleave a large fraction of available network ca-pacity unused.
G. Optimizing video during handovers
Wi-Fi Goes to Town [45] was one ofthe first research approaches to implement aperformance-tuned version of the IEEE 802.11rand 802.11k fast handover protocol. It try touse Picocells to increase the capacity of thenetwork, which results in higher spectral effi-ciency and throughput. In [45] focus is not onvideo delivery. Focus of [45] is on maximizingthroughput that can often lead to lower QoEfor video clients. Wang et al. [93] propose areal-time handover protocol called mmHandoverfor a 5G network working in mmWave. In thepast, there have been many efforts to designmechanisms to predict handovers [94–96]. Theseschemes try to predict handovers to improve theQoS of different traffic on mobile devices.
H. Understand the network bottleneck to im-prove ABR
In their paper BurstTracker [13] focused onLTE networks. BurstTracker identify issues af-fecting the QoE of the video streaming perfor-mance. BurstTracker is able to identify a sur- Fig. 5. Resource Block and Resource Element in LTEnetworks prisingly different bottleneck. BurstTracker un-derstand the scheduling pattern of the LTE basestation. Their focus is to identify the occupancyof each user’s download queue. If this queue isnot empty for one scheduling cycle then accesslink the bottleneck link. It means that data wasin the queue of LTE base station and was notbeing delivered in the next cycle. If there areso many cycles like that, they will decrease theQoE for the video.One of the main challenges, is that user queueinformation is not available at the client. So,this approach designed a method to estimatethis information. The approach observes thatif a user is selected for transmission and itsqueue is full, then the LTE base station sched-uler allocates the complete millisecond durationresource block (RB). As shown in the Figure5 a RB is the smallest unit of resources thatcan be allocated to a user. It consists of 180kHz in frequency while 6-7 OFDM symbols intime. In frequency it is further divided into 12sub-carriers of 15 kHz each. Using these insightsBurstTracker is able to find out most of timesuser queue becomes empty before the completeallocation of the user queue. This suggests thatbottleneck is not at the base station radio link.According to BurstTracker, most of the largeLTE network providers use split-TCP middleboxes. Due to TCP slow start used by middle boxes for TCP connections, these middle boxesintroduce serious performance bottleneck.PiStream [12] has determined total downloadresources allocated to the user on a LTE basestation. BurstTracker is able to estimate it atthe user and then use it to improve the estima-tion of bitrate. Pistream assumes in case of thebottleneck it is at the base station.
I. ABR algorithms with cross-layer optimiza-tions
In the ground breaking paper [12] PiStreamthe first presented a challenge faced by DASHplayers in LTE. LTE bandwidth is very high,around 10x than its predecessor 3G. Despite thishigh bandwidth video clients does not performwell. PiStream motivated this problem with ameasurement study which shows how a DASHclient behaves in the LTE network. DASH usesa throughput estimator to predict the data rate.But in the LTE network this estimate is mostlyan underestimate. This leads DASH selecting alower quality for the future. PiStream observethat in the LTE networks all the bandwidthinformation for the access link is known to theLTE network. The approach takes advantageof the Physical layer information to get anaccurate estimate of the bandwidth. PiStreamimplemented their scheme using SDRs at thePhysical layer and using GPAC player as theopen source client. In their evaluations, theycompared their scheme with FESTIVE, BBA,and PANDA.Recently, Raca et al. [97] designed a approachto address the challenges to ABR video inthe challenging environment of cellular network.They observed in cellular network radio channel,conditions and load on the cell are continu-ously changing. In addition to that, there is gapin the time scale among different componentsof the system. Transport layer protocols reactat the granularity of hundreds of milliseconds,while radio channel changes at the fraction ofmillisecond. One the other side, base stationcan allocate resource in a bursty manner. MLbased approach is used to predict throughput of the mobile devices in a LTE network. ThereML model is learned on cellular trace data.Their approach shows the importance of radiolevel metrics in the video streaming applications[98]. Their technique is implemented in Androidvideo player (ExoPlayer) and performed evalua-tion on a real testbed.In [99], a dataset for 5G measurements is pre-sented. These measurements are performed on aa major Irish mobile operator. In this datasetall the key performance indicators (KPIs) forclient-side cellular metrics and throughput arecollected. Dataset is collected with two differ-ent mobility patterns of the user driving andstatic. Also, stream content is generated fromAmazon Prime and Netflix streaming device.In this dataset, GPS based location informa-tion is also present. This data is generatedfrom Android network monitoring application,G-NetTrack Pro. This application can run overa non-rooted Android phone.In addition to the real dataset, a second syn-thetic dataset is also presented in [99] . This dataset is generated from ns-3 [100] using a large-scale multi-cell 5G/mmwave framework.One of the first cross-layer ABR algorithmfor wireless network was CrystalBall algorithm[101]. It is a two step algorithm to predict avail-able bandwidth. Main ABR algorithm is basedon it. In this study, the effects of the predictionquality on the accuracy of the ABR. CrystalBallshows with error mitigation techniques somelevel of prediction error can be tolerated.Another similar technique is CQIC [102] topredict TCP throughput using Radio level in-formation in smart phones.Yue et al. [55] they showed even a trace of500 data points can be used to develop anaccurate model of LTE link level predictionsusing a ML model. LinkForecast presented anextensive measurement study to understand thebandwidth allocation algorithm of current cel-lular networks. Most of them allocate a fairproportion of the available bandwidth. Availablebandwidth is calculated using the observationsof the recent past throughput and link condi- tions. Using this measurement study as motiva-tion which shows benefits of sharing applicationlayer throughput to the lower-layer. LinkFore-cast explore the idea of sharing lower-layer linkinformation to the application.LinkForecast designs a ML based frameworkto predict link bandwidth in real time. Thisframework combine both upper and lower layersinformation for future prediction. According totheir evaluation this technique is not only moreprecise but also lightweight and insensitive tothe training data.In [97] authors investigate further on the ob-servation of high variability of network condi-tions in cellular radio access networks. Publiclyavailable data sets [103] are used to learn amachine learning model. The data set is veryrich in terms of parameters and different mobil-ity patterns like static, walking, car and trainetc. Random forest based technique is selectedas their ML model. They used random forest(RF) as model for prediction. By increasing thenumber of trees and using mean of all trees inthe RF as the predicted value avoid over-fitting.Parameters used in [97] model are availablethrough Android Debug Bridge (ADB) APIs.Implementation is done on ExoPlayer and per-formed evaluation on a real testbed. In thereevaluation, they show the effectiveness of theirsystem by improving all QoE metrics.Chen et al. [107] authors proposed ABR cansave energy during video stream if it considerthe context of streaming. It means if user iswatching a video in a room may have completelydifferent QoE requirements as compared to auser on a moving vehicle. Using traces they havemodeled the impact of vibration level in additionto video bitrate on the QoE and signal strengthon power consumption. They designed an opti-mal algorithm using an optimization problem tominimize energy consumption.Raca et al. in [106] presents the effects ofthe highly dynamic wireless communication ondifferent application. Here they provide evidenceof PHY layer metrics are used in assigning re-sources to the users in the cellular network. ABR Adaptive Bitrate Streaming (ABR) ABR based on ML[58, 59] Based on DL[10, 11, 18, 40, 47]Based on RL[104]Egde computing based ABR systems[60, 62, 65] Using SDN[63, 66]ABR with different optimization goals Improve QoE for vehicles[45, 86, 87, 93]Optimize for user data-plan[67]Chunk Based Filtering[56]Optimizing video during handovers[45, 93, 96]Video stream in vehicular networks[85–87, 105]4K and ◦ video streaming 4K video streaming[51] ◦ video streaming[78, 79, 82–84]Cross-layer optimizations for Video streaming[55, 97, 106, 107] Fig. 6. Major categories of ABR for wireless networks in the literature is used as an example application to explain theadvantages of using AI based techniques to learnaccurate throughput prediction using PHY layerlevel metrics in cellular networks.We classify the different ABR for wirelessschemes into five main categories see Figure 6.These categories are based on techniques usedto design ABR or different types of objectivesused in the design of the algorithms. Similarlythe reviewed major approaches are summarizedalong with the references in Table II.
IV. Open challenges and opportunities
A. Improving video for developing world clients
Mobile phone adoption is even more explosivein developing countries. It has reached more than 98% recently [108]. But most of these de-vices are low end devices with slow network like2G [109]. But video is the most dominant typeof traffic in these parts of the world. This createsan even higher level challenge for Internet videoproviders [110].
B. Providing effective ABR for developing world
New services like Web Light and Facebook’sFree Basics service are introduced to improvethe Internet quality and availability. Now, Freebasics service is expanded to over 60 countries[111] across select cellular service providers. Butthese services do not handle video elegantly.Both of these services replacing videos with animage [112]. In future, demand for better quality Internetvideo will increase more for these low resourcedeveloping world clients. It is a challenge toprovide even bare minimum service to theseclients [108]. But one can use techniques likeOboe in [6]. There are some proposed schemeswhich incorporate device level characteristics toimprove the selection of bitrate [113].In the developing world, vast amount of dataaccess and high speed both are a luxury. Most ofthe communication is through text-based medialike posters and flyers. Most of the users in thedeveloping regions are illiterate and resource-constraint in terms of poor connectivity andlittle exposure to the technology. Access to com-puter and laptops is also very limited. Mostof the Internet access is through low-end cell-phone with a low-end camera. There has beenmany novel applications designed to solve dif-ferent developing world specific problems. Manyof these applications depends on video basedsolutions [114, 115]. Video streaming patterns ofcommunity health workers in Africa is studiedin [116].They demonstrated the effectiveness ofhealth videos and also presented lessons forprojects seeking to use multimedia content inrural setting.AudioCanvas [117] is an application createdfor rural developing regions. It can be used bytelephone as an audio information system. Thissystem enable rural users to interact directlywith their pictures and receive narration or de-scription.In a recent paper [113] authors have studiedeffects of memory pressure on video streamingapplications. Their experiments suggest QoE ofvideo streaming is significantly affected by theselecting higher nitrates in a low memory cell-phone. On the other hand [109] presents a com-prehensive measurement study of cell phonesused by users in developing world. Dataset usedin [109] has less than 1% of the cellphone users inthe developing world have more than 2GB mem-ory in there devices. This creates a challengeof designing specialized ABRs for developingworld. In their measurements Nexus 5 phone with2GB memory led to frequent video playercrashes when it plays a 1080p video. To achievehigh QoE under low-memory scenarios they pro-posed an new scheme DAVS which adapt theplayback buffer size based on conditions basedon device memory pressure.
C. Optimize the spectral efficiency for video traf-fic
The consistent exponential growth of videotraffic will increase in the future with the adventof 5G based IoT devices. WiFi alliance hasrecently approved WiFi-6 [118] (802.11ax) withthe focus of high density WLANs. These newchanges will lead to even more congestion in theavailable spectrum specially in the unlicenseddomain. There have been many recent studiesto understand and optimize wireless spectrumsharing between different technologies like LTEor 5G based cellular networks and WiFi in unli-censed bands [9, 28, 29, 119]. Some of them useda machine learning-based approach to optimizespectrum sharing [10]. But most of these papersare optimizing at the network level. This leadsto many lost opportunities specifically related tovideo streaming [88].
D. Improving the QoE in the presence of networkhandovers
Some recent measurement studies [105] showsthat current policies of cellular carriers are notoptimized, especially during handovers. Thesepolicies do not consider cell load informationduring handover resulting in degradation of ap-plication performance. In a 5G small cell, han-dovers will be more frequent. Figure 7 shows atypical scenario of small cell based network in5G vehicles in these small cells will experiencefrequent hand-offs. It will be critical for videostreaming applications to perform well duringthese handovers. Fig. 7. Presence of Small Cells and frequent handovers in 5G
V. Open-source implementations ofABR algorithms and datasets
There are many open-source video playersavailable on the Internet. But dash.js and Ex-oPlayer [120] are the most popular in the in-dustry and research. ExoPlayer is developedby Google as the first Android-based mobileDASH player. Many research paper have usedit as their reference to implement their ABRalgorithm. The other popular implementationis dash.js [121]. It is developed by DASH In-dustry Forum which is supported by most ofthe major players in Internet video industry likeAkamai [122]. Previously GPAC [123] was alsovery popular in research for prototyping. It isnow called MP4Client. Both GPAC and DASHare implemented in JavaScript. In comparisonto all the open-source implementations, dash.jsencapsulates the standard and best practices. Itis easy to customize and there is an Akamaireference implementation also available onlinewhich makes it easy to test. There are librariesavailable to use this for trace-driven analysis.Video providers wishing to use DASH oftenuse the reference client dash.js to build theirown video players. Table III shows open-source
TABLE III
Summary of open-source ABR implementations
Implementation Corresponding papersDash player Pensive [5],Oboe[6],HotDASH [46],Bursttracker [13],QUAD [56] and NAS[49]GPAC video player [44] and Pstream [12]ExoPlayer IncorpPred [50] andQUAD [56]Trace driven QARC [47], Comyco[43], and ACAA[54] implementation used in different papers. It isevident that most popular implementation inresearch is dash.js. Even in wireless networkbased evaluations [13, 56] it is more commonlyused due to its flexibility and acceptance. Thereare hundreds of different ABR algorithms im-plemented using these open-source players. Oneof the popular one is of Pensive and their datatraces [124]. Recently NAS [49] implementationis also available open-source [125].There are two open-source data sets available[103, 126] for trace-based analysis and to trainmachine learning algorithms. They are available in multiple encoding. These encoding rates arecomparative to the large CDN provider likeHulu, YouTube and Netflix.The testbed used to collect measurementsin [126] is based on “MP4Client”, a multime-dia player based on GPAC [123]. They utilisedvery well known animated videos like ElephantDreams, Big Buck Bunny etc. Their files areobtained as 920x1080 YUV files.Similarly the second dataset [126] is a traceof 4G dataset which is composed of key perfor-mance indicators (KPIs) from two major Irishcellular providers. They are collected with dif-ferent mobility patterns like static, car, bus andtrain. It has a very diverse range of throughputfrom 0 to 173 Mbit/sec. These traces are gener-ated from a well-known Android network moni-toring applications. There are few limitations ofthis dataset first all of its samples are of 1secduration. Also, there is no GPS information inthis dataset. To supplement the limitations ofthe real dataset there is repository of syntheticdataset. VI. Conclusion
In this paper, we surveyed different key tech-niques in the area of Internet video for wirelessnetworks. It was observed that many approachesused cross layer communications on the clientsto improve the re-buffering, quality switchingand encoding related impairments of the mobilevideo. It is important to note that with theupcoming deployments of WiFi-6 and 5G newchallenges will arise which require us to rethinkthe implementation of ABR algorithms to ad-dress these challenges.
References
ACM IMC, 2012 ,2012. [4] V. S. Florin Dobrian, “Understanding the impact ofvideo quality on user engagement.” in
SIGCOMM,2011. , 2011.[5] R. N. Hongzi Mao and M. Alizadeh, “Neural adap-tive video streaming with pensieve,” in
ACM SIG-COMM 2017. , 2017.[6] Z. Akhtar, Y. S. Nam, R. Govindan, S. Rao,J. Chen, E. Katz-Bassett, B. Ribeiro, J. Zhan,and H. Zhang., “Oboe: auto-tuning video abr algo-rithms to network conditions.” in
ACM SIGCOMM2018 , 2018.[7] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lil-licrap, T. Harley, D. Silver, and K. Kavukcuoglu,“Asynchronous methods for deep reinforcementlearning,” in
Proceedings of The 33rd InternationalConference on Machine Learning , ser. Proceedingsof Machine Learning Research, M. F. Balcan andK. Q. Weinberger, Eds., vol. 48. New York, NewYork, USA: PMLR, 20–22 Jun 2016, pp. 1928–1937.[8] V. Mnih, D. S. Koray Kavukcuoglu, A. A. Rusu,J. Veness, M. G. Bellemare, A. Graves, M. Ried-miller, A. K. Fidjeland, and G. Ostrovski, “Human-level control through deep reinforcement learning.”in
Nature, 2015. , 2016, pp. 1928–1937.[9] S. Mustafa, K. A. Alam, B. Khan, M. H. Ullah, andP. Touseef, “Fair coexistence of lte and wifi-802.11in unlicensed spectrum, a systematic literature re-view,” in
ACM ICFNDS 2019. , 2019.[10] M. Alsenwi, I. Yaqoob, S. R. Pandey, Y. K. Tun,A. K. Bairagi, L. Kim, and C. S. Hong, “Towardscoexistence of cellular and wifi networks in un-licensed spectrum: A neural networks based ap-proach.” in
IEEE Access 2019 , 2019.[11] M. A. et al., “Towards coexistence of cellular andwifi networks in unlicensed spectrum: A neuralnetworks based approach,” in
IEEE Access 2019 ,2019.[12] X. Xie, X. Zhang, S. Kumar, and L. E. Li,“Pistream: Physical layer informed adaptive videostreaming over lte,” in
ACM MobiCom 2015 , 2015.[13] A. Balasingam, M. Bansal, R. Misra, K. Nagaraj,R. Tandra, S. Katti, and A. Schulman, “Detectingif lte is the bottleneck with bursttracker,” in
ACMMobicom 2019 , 2019.[14] X. Ran, H. Chen, X. Zhu, Z. Liu, and Jiasi, “Chen.deepdecision: A mobile deep learning frameworkfor edge video analytics.” in
In Proc. IEEE Interna-tional Conference on Computer Communications,2018 , 2018.[15] Y. Yuan, X. Liang, X. Wang, D.-Y. Yeung, andA. Gupta., “Temporal dynamic graph lstm foraction-driven video object detection.” in
In Proc.IEEE International Conference on Computer Vi-sion (ICCV) 2017 , 2017.[16] K. C. Zongqing Lu, Noor Felemban and T. L.Porta., “Demo abstract: On-demand informationretrieval from videos using deep learning in wirelessnetworks.” in
In Proc. IEEE/ACM Second Inter-national Conference on Internet-of-Things Designand Implementation (IoTDI) , 2019. [17] Q. Zhang, M. Lin, L. T. Yang, Z. Chen, and P. Li.,“Energy-efficient scheduling for real-time systemsbased on deep q-learning model.” in IEEE Trans-actions on Sustainable Computing, 2017. , 2017.[18] S. Rallapalli, H. Qiu, A. Bency, S. Karthikeyan,R. Govindan, B. Manjunath, and R. Urgaonkar.,“Are very deep neural networks feasible on mobiledevices.” in
IEEE Transactions on Circuits andSystems for Video Technology, 2016 , 2012.[19] S. E. M. Seufert, M. Slanina, T. Zinner, T. Hoßfeld,and P. Tran-Gia, “A survey on quality of experi-ence of http adaptive streaming,” in
IEEE Com-mun. Surveys Tuts., 2015. , 2015.[20] A. Bentaleb, B. Taani, A. C. Begen, C. Timmerer,and R. Zimmermann, “A survey on bitrate adap-tation schemes for streaming media over http,”in
IEEE Communications Surveys Tutorials 2019 ,2019.[21] V. T. P. Juluri and D. Medhi, “Measurement ofquality of experience of video-on-demand services:A survey,” in
IEEE Commun. Surveys Tuts., 2016. ,2016.[22] N. Barman and M. G. Martini, “Qoe modeling forhttp adaptive video streaming–a survey and openchallenges,” in
IEEE Access, 2019 , 2019.[23] J. Kua, G. Armitage, and P. Branch, “A survey ofrate adaptation techniques for dynamic adaptivestreaming over http,” in
IEEE CommunicationsSurveys Tutorials, 2017 , vol. 19, no. 3, 2017, pp.1842–1866.[24] Q. Mao, F. Hu, and Q. Hao, “Deep learning forintelligent wireless networks: A comprehensive sur-vey,” in
IEEE Communications Surveys Tutorials ,2019.[25] M. Seufert, “A survey on quality of experience ofhttp adaptive streaming,” in
IEEE Communica-tions Surveys Tutorials 2015 , 2015.[26] A. A. Barakabitze, N. Barman, A. Ahmad, S. Zad-tootaghaj, L. Sun, M. G. Martini, and L. Atzori,“Qoe management of multimedia streaming ser-vices in future networks: A tutorial and survey,”in
IEEE Communications Surveys Tutorials, 2020 ,vol. 22, no. 1, 2020, pp. 526–565.[27] A. Bentaleb, B. Taani, A. C. Begen, C. Timmerer,and R. Zimmermann, “A survey on bitrate adap-tation schemes for streaming media over http,” in
IEEE Communications Surveys Tutorials, 2019 ,2019.[28] L. Bertizzolo, E. Demirors, Z. Guan, and T. Melo-dia, “Cobeam: Beamforming-based spectrum shar-ing with zero cross-technology signaling for 5g wire-less networks,” in
IEEE Infocom 2020 , 2020.[29] S. Zheng, T. Han, Y. Jiang, and X. Ge, “Smartcontract-based secure spectrum sharing in multi-operators wireless communication networks,” in
ArXiv 2020 , 2020.[30] I. Sodagar, “The mpeg-dash standard for multime-dia streaming over the internet,” in
IEEE Multi-media, 2011.
In Proceedingsof the ACM Conference on Special Interest Groupon Data Communication, SIGCOMM, 2015. , 2015.[35] R. U. Kevin Spiteri and R. K. Sitaraman., “Bola:Near-optimal bitrate adaptation for online videos.”in
In Proceedings of the IEEE International Con-ference on Computer Communications, INFO-COM, 2016 , 2016.[36] DASH Industry Forum. [Online]. Available: https://github.com/Dash-Industry-Forum/dash.js[37] L. T. Federico Chiariotti, Stefano D’Aronco andP. Frossard., “Online learning adaptation strategyfor dash clients.” in
In Proceedings of the Interna-tional Conference on Multimedia Systems, MMSys2016. , 2016.[38] M. Claeys, S. Latré, J. Famaey, T. Wu, W. V. Leek-wijck, and F. D. Turck., “Design and optimisationof a (fa)q-learning-based http adaptive streamingclient.” in
Connection Science, 26(1):25–43, 2014. ,2014.[39] J. C. Virginia Martín and N. García., “Design,optimization and evaluation of a q-learning httpadaptive streaming client.” in
IEEE Transactionson Consumer Electronics, 62(4):380–388, 2016 ,2016.[40] C. Zhang, P. Patras, and H. Haddadi, “Deep learn-ing in mobile and wireless networking a survey,” in
IEEE COMMUNICATIONS SURVEYS and TU-TORIALS 2019 , 2019.[41] K. Arulkumaran, M. P. Deisenroth, M. Brundage,and A. A. Bharath, “A brief survey of deep rein-forcement learning,” in
IEEE SIGNAL PROCESS-ING MAGAZINE, SPECIAL ISSUE ON DEEPLEARNING FOR IMAGE UNDERSTANDING ,2017.[42] A. Ahmad and L. Atzori, “Mno-ott collaborativevideo streaming in 5g: The zero-rated qoe approachfor quality and resource management,” in
IEEETransactions on Network and Service Management2020 , vol. 17, no. 1, 2020, pp. 361–374.[43] T. Huang, C. Zhou, R.-X. Zhang, C. Wu, X. Yao,and L. Sun, “Comyco: Quality-aware adaptivevideo streaming via imitation learning,” in
ACMMM 2019 , 2019.[44] B. Han, F. Qian, L. Ji, and V. Gopalakrishnan,“Mp-dash: Adaptive video streaming overpreference-aware multipath,” in
Proceedingsof the 12th International on Conferenceon Emerging Networking EXperiments andTechnologies , ser. CoNEXT ’16. New York,NY, USA: Association for Computing Ma- chinery, 2016, p. 129–143. [Online]. Available:https://doi.org/10.1145/2999572.2999606[45] Z. Song, L. Shangguan, and K. Jamieson, “Wi-figoes to town: Rapid picocell switching for wirelesstransit networks.” in SIGCOMM 2017 , 2017.[46] S. Sengupta, N. Ganguly, S. Chakraborty, andP. De, “Hotdash: Hotspot aware adaptive videostreaming using deep reinforcement learning,” in , 2018, pp. 165–175.[47] H. Tianchi, Z. Rui-Xiao, Z. Chao, and S. Lifeng,“Qarc: Video quality aware rate control for real-time video streaming based on deep reinforcementlearning,” in
Proceedings of the 26th ACM Inter-national Conference on Multimedia , ser. MM ’18,2018.[48] B. Rajarshi, B. Archana, R. Desik, R. Mason,S. Srinivas, K. Dileep, M. R. K. P., and D. Amogh,“Qflow: A reinforcement learning approach to highqoe video streaming over wireless networks,” in
Mobihoc ’19 , 2019.[49] H. Yeo, Y. Jung, J. Kim, J. Shin, and D. Han,“Neural adaptive content-aware internet video de-livery,” in ,2018.[50] D. Raca, A. H. Zahran, C. J. Sreenan, R. K.Sinha, E. Halepovic, R. Jana, V. Gopalakrishnan,B. Bathula, and M. Varvello, “Incorporatingprediction into adaptive streaming algorithms:A qoe perspective,” in
Proceedings of the28th ACM SIGMM Workshop on Networkand Operating Systems Support for DigitalAudio and Video , ser. NOSSDAV ’18. NewYork, NY, USA: Association for ComputingMachinery, 2018, p. 49–54. [Online]. Available:https://doi.org/10.1145/3210445.3210457[51] G. Baig, J. He, M. A. Qureshi, L. Qiu, G. Chen,P. Chen, and Y. Hu, “Jigsaw: Robust live 4k videostreaming,” in
The 25th Annual International Con-ference on Mobile Computing and Networking , ser.MobiCom ’19, 2019.[52] J. Yoon and S. Banerjee, “Hardware-assisted, low-cost video transcoding solution in wireless net-works,” in
IEEE Transactions on Mobile Comput-ing, 2020 , 2020.[53] J. Lee, J. Lee, Y. Im, S. Dhawaskar Sathya-narayana, P. Rahimzadeh, X. Zhang,M. Hollingsworth, C. Joe-Wong, D. Grunwald, andS. Ha, “Castle over the air: Distributed schedulingfor cellular data transmissions,” in
Proceedingsof the 17th Annual International Conference onMobile Systems, Applications, and Services , ser.MobiSys ’19, 2019.[54] S. Hu, M. Xu, H. Zhang, C. Xiao, and C. Gui,“Affective content-aware adaptation scheme on qoeoptimization of adaptive streaming over http,”in
ACM Transactions on Multimedia Computing,Communications, and Appli. 2019 , 2019.[55] C. Yue, R. Jin, K. Suh, Y. Qin, B. Wang, andW. Wei, “Linkforecast: Cellular link bandwidth prediction in lte networks,” in
IEEE Transactionson Mobile Computing , 2018.[56] Y. Qin, S. Hao, K. R. Pattipati, F. Qian, S. Sen,B. Wang, and C. Yue, “Quality-aware strategies foroptimizing abr video streaming qoe and reducingdata usage,” in
Proceedings of the 10th ACM Mul-timedia Systems Conference , ser. MMSys ’19, 2019.[57] C. Ye, M. C. Gursoy, and S. Velipasalar, “Powercontrol for wireless vbr video streaming: Fromoptimization to reinforcement learning,” in
IEEETransactions on Communications , 2019.[58] Z. Meng, J. Chen, Y. Guo, C. Sun, H. Hu, andM. Xu, “Pitree: Practical implementation of abralgorithms using decision trees,” in
Proceedings ofthe 27th ACM International Conference on Multi-media 2019 , ser. MM ’19, 2019.[59] J. Molina, D. Muelas, J. E. L. de Vergara, andJ. J. G. Aranda, “Network quality-aware architec-ture for adaptive video streaming from drones,” in
IEEE Internet Computing. 2020 , 2020.[60] R. Bhattacharyya, A. Bura, D. Rengarajan, M. Ru-muly, S. Shakkottai, D. Kalathil, R. K. P. Mok, andA. Dhamdhere., “Qflow: A reinforcement learningapproach to high qoe video streaming over wirelessnetworks.” in
ACM Mobihoc 2019 , 2019.[61] K. Bilal and A. Erbad, “Edge computing for inter-active media and video streaming,” in , 2017, pp. 68–73.[62] X. Ran, C. Slocum, M. Gorlatova, andJ. Chen, “Sharear: Communication-efficient multi-user mobile augmented reality,” in
Proceedingsof the 18th ACM Workshop on HotTopics in Networks , ser. HotNets ’19. NewYork, NY, USA: Association for ComputingMachinery, 2019, p. 109–116. [Online]. Available:https://doi.org/10.1145/3365609.3365867[63] I. Ben Mustafa, T. Nadeem, and E. Halepovic,“Flexstream: Towards flexible adaptive videostreaming on end devices using extreme sdn,”in
Proceedings of the 26th ACM InternationalConference on Multimedia, 2018 , ser. MM ’18.New York, NY, USA: Association for ComputingMachinery, 2018, p. 555–563. [Online]. Available:https://doi.org/10.1145/3240508.3240676[64] M. Liu, F. R. Yu, Y. Teng, V. C. M. Leung,and M. Song, “Distributed resource allocation inblockchain-based video streaming systems withmobile edge computing,” in
IEEE Transactions onWireless Communications, 2019 , vol. 18, no. 1,2019, pp. 695–708.[65] T. X. Tran and D. Pompili, “Adaptive bitrate videocaching and processing in mobile-edge computingnetworks,” in
IEEE Transactions on Mobile Com-puting, 2019 , 2019.[66] M. Uddin, T. Nadeem, and S. Nukavarapu, “Ex-treme sdn framework for iot and mobile applica-tions flexible privacy at the edge,” in , 2019. [67] “Ericsson. 2017. ericsson mobility report. (2017).”[Online]. Available: https://goo.gl/mjkwSH.[68] W. Hu and G. Cao, “Quality-aware traffic offload-ing in wireless networks,” in Proceedings of the15th ACM International Symposium on Mobile AdHoc Networking and Computing , ser. MobiHoc ’14,2014, p. 277–286.[69] R. S. Kevin Spiteri and D. Sparacio., “From theoryto practice: Improving bitrate adaptation in thedash reference player.” in
In Proceedings of the9th ACM Multimedia Systems Conference (MMSys2018). , 2018.[70] T. Mangla, E. Zegura, M. Ammar, E. Halepovic,K.-W. Hwang, R. Jana, and M. Platania,“Videonoc: Assessing video qoe for networkoperators using passive measurements,” in
Proceedings of the 9th ACM MultimediaSystems Conference , ser. MMSys ’18. NewYork, NY, USA: Association for ComputingMachinery, 2018, p. 101–112. [Online]. Available:https://doi.org/10.1145/3204949.3204956[71] T.-Y. Huang, N. Handigol, B. Heller, N. McKeown,and R. Johari, “Confused, timid, and unstable:Picking a video streaming rate is hard,” in
Proceed-ings of the 2012 Internet Measurement Conference ,2012.[72] I. Orsolic, D. Pevec, M. Suznjevic, and L. Skorin-Kapov, “Youtube qoe estimation based on theanalysis of encrypted network traffic using machinelearning,” in , 2016, pp. 1–6.[73] Z. W. Zhengfang Duane, Abdul Rehman, “Aquality-of-experience database for adaptive videostreaming,” in
IEEE Transactions on Broadcasting2018 , 2018.[74] F. Y. Yan, H. Ayers, C. Zhu, and K. Winstein,“Learning in situ: a randomized experiment invideo streaming,” in
USENIX NSDI 2020 , 2020.[75] D. Bhat, A. Rizk, and M. Zink, “Not so quic: A per-formance study of dash over quic,” in
Proceedingsof the 27th Workshop on Network and OperatingSystems Support for Digital Audio and Video , ser.NOSSDAV’17, 2017.[76] K. Abdullah, N. Korany, A. Khalafallah, A. Saeed,and A. Gaber, “Characterizing the effects of rapidlte deployment: A data-driven analysis,” in , 2019, pp. 97–104.[77] S. Xu, S. Sen, and Z. M. Mao, “Csi: Inferringmobile abr video adaptation behavior under httpsand quic,” in
Proceedings of the Fifteenth EuropeanConference on Computer Systems , ser. EuroSys’20. New York, NY, USA: Association forComputing Machinery, 2020. [Online]. Available:https://doi.org/10.1145/3342195.3387558[78] F. Qian, B. Han, Q. Xiao, and V. Gopalakr-ishnan, “Flare: Practical viewport-adaptive 360-degree video streaming for mobile devices,” in
Pro-ceedings of the 24th Annual International Confer-ence on Mobile Computing and Networking , ser.MobiCom ’18, 2018. [79] J. He, M. A. Qureshi, L. Qiu, J. Li, F. Li, andL. Han, “Rubiks: Practical 360-degree streamingfor smartphones,” in
Proceedings of the 16th An-nual International Conference on Mobile Systems,Applications, and Services
IEEE INFOCOM2019 - IEEE Conference on Computer Communi-cations , 2019.[83] M. Tang and V. W. Wong, “Online bitrate selectionfor viewport adaptive 360-degree video streaming,”in
CoRR 2020 , 2020. [Online]. Available: https://arxiv.org/pdf/2005.02479.pdf[84] C. Perfecto, M. S. ElBamby, J. D. Ser, andM. Bennis, “Taming the latency in multi-uservr 360-degree: A qoe-aware deep learning-aidedmulticast framework,” in
CoRR , 2020. [Online].Available: http://arxiv.org/abs/1811.07388[85] S. M. A. Kazmi, T. N. Dang, I. Yaqoob, A. Ndiku-mana, E. Ahmed, R. Hussain, and C. S. Hong,“Infotainment enabled smart cars: A joint commu-nication, caching, and computation approach,” in
IEEE Transactions on Vehicular Technology 2019. ,2019.[86] S. S. H. Khan and M. Bennis, “Enhancing videostreaming in vehicular networks via resource slic-ing,” in
IEEE Transactions on Vehicular Technol-ogy. 2019 , 2019.[87] Z. Zhang, Y. Yang, Y. H. M. Hua, C. Li, andL. Yang, “Proactive caching for vehicular multi-view 3d video streaming via deep reinforcementlearning,” in
IEEE Transactions on Wireless Com-munications, 2019 , 2019.[88] R.Ali, B.Kim, S.W.Kim, and H.S.Kim, “(relbt): Areinforcement learning-enabled listen before talkmechanism for lte-laa and wi-fi coexistence in iot,”in
Computer Communications 2020. , 2020.[89] D. R. C. Li, B. Lin, and M. Pan, “Energy-efficientproactive caching for adaptive video streaming viadata-driven optimization,” in
IEEE Internet ofThings Journal, 2020 , 2020.[90] H. Shamim, H. Zhang, C. Moussa, and K. Liu.,“Effective capacity-aware resource allocation for 5gultra-dense network with hybrid access mode.” in
RACS 2019 , 2019.[91] C. E. Thornton, R. M. Buehrer, A. F. Martone, andK. D. Sherbondy, “Experimental analysis of rein-forcement learning techniques for spectrum sharingradar,” in
Arxiv 2020 , 2020.[92] A. S. Melissa Licciardello, Maximilian Grüner,“Understanding video streaming algorithms in thewild,” in
PAM 2020 , 2020.[93] X. Wang, L. Kong, J. Wu, X. Gao, H. Wang, andG. Chen., “Mmhandover: a pre-connection basedhandover protocol for 5g millimeter wave vehicularnetworks.” in
ACM IWQoS 2019. , 2019. [94] H. Ge, X. Wen, Z. L. W. Zheng, and B. Wang.,“A history-based hand-dover prediction for lte sys-tems.” in International Symposium on Computer-Network and Multimedia Technology 2009 , 2009.[95] I. M. Bălan, B. Sas, T. Jansen, K. S. I. Moer-man, and P. Demeester., “An enhanced weightedperformance-based handover parameter optimiza-tion algorithm for lte networks.” in
EURASIPJournal on Wireless Communications andNet-working 2011 , 2011.[96] W. Luo, M. C. X. Fang, and X. Zhou., “An opti-mized handover trigger scheme in lte systems forhigh-speed railway.” in
In Proceedings of the FifthIn-ternational Workshop on Signal Design and ItsApplications in Communications 2018 , 2018.[97] D. Raca, A. H. Zahran, C. J. Sreenan, R. K.Sinha, E. Halepovic, R. Jana, V. Gopalakrishnan,B. Bathula, and M. Varvello, “Empowering videoplayers in cellular: Throughput prediction from ra-dio network measurements,” in
ACM MMSys ’19 ,2019.[98] M. Mirza, J. Sommers, P. Barford, and X. Zhu,“A machine learning approach to tcp throughputprediction,” in
IEEE/ACM Transactions on Net-working 2010 , vol. 18, no. 4, 2010, pp. 1026–1039.[99] D. Raca, D. Leahy, C. S. J., and J. J. Quinlan,“Beyond throughput: the next generation a 5gdataset with channel and context metrics,” in
ACMMultimedia Systems Conference , ser. MMSys’ ’20,2020.[100] N. Baldo, M. Miozzo, M. Requena-Esteso, andJ. Nin-Guerrero, “An open source product-orientedlte network simulator based on ns-3. in proceedingsof the 14th acm international conference on model-ing, analysis and simulation of wireless and mobilesystems,” in
MSWiM 2011 , 2011.[101] T. Mangla, N. Theera-Ampornpunt, M. Ammar,E. Zegura, and S. Bagchi, “Video through a crystalball: Effect of bandwidth prediction quality onadaptive streaming in mobile environments,” in
Proceedings of the 8th International Workshopon Mobile Video , ser. MoVid ’16. New York,NY, USA: Association for Computing Machinery,2016. [Online]. Available: https://doi.org/10.1145/2910018.2910653[102] F. Lu, H. Du, A. Jain, G. M. Voelker,A. C. Snoeren, and T. andreas, “Cqic:Revisiting cross-layer congestion control forcellular networks,” in
Proceedings of the 16thInternational Workshop on Mobile ComputingSystems and Applications , ser. HotMobile’15. New York, NY, USA: Association forComputing Machinery, 2015. [Online]. Available:https://doi.org/10.1145/2699343.2699345[103] D. Raca, J. J. Quinlan, A. H. Zahran, andC. J. Sreenan, “Beyond throughput: A 4g ltedataset with channel and context metrics,” in
Proceedings of the 9th ACM Multimedia SystemsConference , ser. MMSys ’18. New York, NY,USA: Association for Computing Machinery, 2018, p. 460–465. [Online]. Available: https://doi.org/10.1145/3204949.3208123[104] HongziMao, R. Netravali, and M. Alizadeh, “Neu-ral adaptive video streaming with pensieve,” in
SIGCOMM ’17 , 2017.[105] M. Z. Xu S., Nikravesh A., “Leveraging context-triggered measurements to characterize lte han-dover performance.” in
PAM 2019. , 2019.[106] D. Raca, A. H. Zahran, C. J. Sreenan, R. K.Sinha, E. Halepovic, R. Jana, and V. Gopalakrish-nan, “On leveraging machine and deep learning forthroughput prediction in cellular networks: Design,performance, and challenges,” in
IEEE Communi-cations Magazine, 2020 , 2020.[107] X. Chen, T. Tan, and G. Cao, “Energy-aware andcontext-aware video streaming on smartphones,”in
Proceedings of the 2016Internet Measurement Conference , ser. IMC ’16.Association for Computing Machinery, 2016.[110] T. B. Usama Naseer, “Configtron: Tackling net-work diversity with heterogeneous configurations.”in
USENIX HotCloud 2017 , 2017.[111] Free Basics Platform. [Online]. Available: https://developers.facebook.com/docs/internet-org.[112] A. Tahir, M. T. Munir, S. M. Malik, Z. A. Qazi,and I. A. Qazi, “Deconstructing google’s web lightservice,” in
Proceedings of The Web Conference2020 , 2020.[113] A. A. Jumani, F. Zafar, Z. A. Qazi, and I. A. Qazi,“Device-aware adaptive video streaming,” in
Pro-ceedings of the ACM SIGCOMM 2019 ConferencePosters and Demos , New York, NY, USA, 2019.[114] D. M. Frohlich, D. Rachovides, K. Riga, R. Bhat,M. Frank, E. A. Edirisinghe, D. Wickramanayaka,M. Jones, and W. Harwood, “Storybank: mobiledigital storytelling in a development context,” in
Proceedings of the 27th International Conferenceon Human Factors in Computing Systems, CHI2009, Boston, MA, USA, April 4-9, 2009 . ACM,2009, pp. 1761–1770.[115] N. Kumar and R. J. Anderson, “Mobile phones formaternal health in rural india.” in
In Proceedingsof the 33rd Annual ACM Conference on HumanFactors in Computing Systems. , 2015.[116] M. Molapo, M. Densmore, and B. DeRenzi, “Videoconsumption patterns for first time smartphoneusers: Community health workers in lesotho,” in
Proceedings of the 2017 CHI Conference on HumanFactors in Computing Systems , ser. CHI ’17.New York, NY, USA: Association for Computing Machinery, 2017, p. 6159–6170. [Online]. Available:https://doi.org/10.1145/3025453.3025616[117] J. Pearson, S. Robinson, and M. Jones, “Ex-ploring low-cost, internet-free information accessfor resource-constrained communities,” in
ACMTrans. Comput.-Hum. Interact., 2016 . Associa-tion for Computing Machinery, 2016.[118] “Wi-fi alliance R (cid:13) ACM Mobihoc 2019 , 2019.[120] “Google, “exoplayer.” [online]. available:.”[Online]. Available: https://google.github.io/ExoPlayer/guide.html[121] “Dashif-reference-player. [online]. available:.”[Online]. Available: http://dashif.org/reference/players/javascript/1.4.0/samples/dash-if-reference-player/[122] “Akamia player testing. [online]. available:.”[Online]. Available: http://players.akamai.com/players/dashjs[123] “Gpac 2019. gpac multimedia open source project,javascript version of gpac’smp4box tool.” [Online].Available: https://gpac.github.io/mp4box.js[124] “H. mao, r. netravali, and m alizadeh. 2017.pensieve. code repository.” [Online]. Available:https://github.com/hongzimao/pensieve[125] “Nas. code repository.” [Online]. Available: https://github.com/kaist-ina/NAS_public[126] J. J. Quinlan, A. H. Zahran, and C. J. Sreenan,“Datasets for avc (h.264) and hevc (h.265) eval-uation of dynamic adaptive streaming over http(dash),” in
Proceedings of the 7th InternationalConference on Multimedia Systems , ser. MMSys’16, 2016.
Kamran Nishat is a Postdoc-toral Fellow in the Departmentof Engineering Technology, Uni-versity of Houston, Texas, USA.He finished his PhD from Depart-ment of Computer Science, LUMS,Lahore, Pakistan. He has receivedhis B.Sc. degree in Mathematics in1999 and MCS Computer Sciencedegree, in 2001,from University ofKarachi, Karachi. Prior to join-ing LUMS, he taught in the Department of ComputerScience, University of Karachi, Karachi.He received hisMasters in computer science in 2007 from LUMS. Hisresearch domain is Networking and Systems. He haspublished in leading Networking conferences includingACM CoNEXT and IEEE Infocom.
Omprakash Gnawali is an As-sociate Professor at the ComputerScience Department of the Univer-sity of Houston, USA. He does re-search in wireless networks, cyber-security, and related technologiesand has contributed research arti-cles, open source software, and in-dustry standards. He received hisSB and MEng from MIT, PhDfrom USC, and was a postdoc atStanford.