Efficient Parallel Framework for Adaptive

EFFICIENT PARALLEL FRAMEWORK FOR ADAPTIVE VIDEO STREAMING IN CLOUD SOCIAL TV

R.Ramya, G.Thendral, Department of Computer Science and Engineering, Panimalar Engineering College, [email protected],[email protected].

Mrs.T.Kalaichelvi M.E., (PhD) Department of Computer Science and Engineering, Panimalar Engineering College, [email protected]

Abstract-The evolution of cloud social TV has enabled the users to have integrated media experiences across different devices.The main aim of this paper is to design a system for video teleportation.Here we propose a multi-screen social TV system over a Cloud-Centric Media Network (CCMN). In particular, one can easily migrate video session back and forward among different devices, with intuitive human-computer interactions.To achieve this we formulate a Markov Decision Problem, to balance the trade-off between the migration cost and the content transmission cost. We then present a more practical Q-learning approach to make online decisions.Finally we investigate this cloud clone migration problem, aiming to minimize the monetary cost on operating video teleportation.

Index Terms—Cloud clone, cost minimization, markov decision process, Q-learning, social TV.

I] INTRODUCTION interactions. As a result, a seamless multiscreen social experience is achieved. Recently TV experiences have dramatically The enabling technology for video changed in which the user has a unified teleportation is to instantiate a virtual media experiences.The user can teleport machine in the cloud as a cloud clone [2] for ongoing video sessions between different each user.1 Specifically, each cloud clone devices at any point of time irrespective of represents one user, serving as his proxy in the user’s location with an affordable cost the cloud, to manage all the associated without any service interruption. devices and session information. In addition, In response to this trend, we have designed the cloud clone also provides video and implemented a multi-screen social TV transcoding [4] and advertisement insertion system [1], [2] over a Cloud-Centric Media function to the original video streaming, Network (CCMN) [3]. It offers video achieving a personalized multi-screen social teleportation as its salient feature. In TV experience for end-users in a scalable particular, one can easily migrate video and flexible manner. session back and forward among different devices, with intuitive human-computer In this paper, our contributions are multi- node along the path is indexed as according folder, including: to its hop distance to the source. Note the  We formulate the cost-minimization cloud clone can neither locate at the user problem as a Markov Decision side nor the content source. To minimizing Process by adopting a Markov chain the operational cost on support the video to model the user watching behavior teleportation feature we are using html5. across TV and smartphone. The objective is to minimize the IV] SYSTEM OVERVIEW & monetary cost of operating the video PROBLEM FORMULATION teleportation service, by migrating the cloud clone to the best place, as System Architecture: In Fig. 1, we present a the user shifts his device. systematic end-to-end view of our cloud  Secondly we propose the Q-learning multi-screen social TV system.Specifically, method, which learns the user the system is built upon CCMN [3], which behavior from the history and makes provides on-demand media services, online decisions at each time slot. including content distribution,media This approach is more practical in processing and content adaption. End users real system. with different devices are connected via residential gateways and access networks II] EXISTING WORK to the cloud. They can request a live or on- demand TV program from an IPTV source The cloud clone provides video through the cloud via virtual overlay content transcoding and advertisement insertion delivery network. Under this framework, the function to the original video streaming, cloud resources can be dynamically operated achieving a personalized multi-screen social in different servers(i.e., cloud nodes) on top TV experience for end-users in a scalable of the overlay network, based on a pay-per- and flexible manner. One critical design use pricing model. Such cloud based CDN objective is to minimize the monetary cost paradigm has also been offered by a list of on operating video teleportation, potentially cloud service providers to operate media making this service affordable to the general services in a more efficient manner, than the public. traditional CDN solution [5].

III] PROPOSED SYSTEM

Cloud clone provide a highly-touted multiscreen experience via video teleportation. With this feature, one user can simultaneously operate multiple devices (e.g., TV and smart phone) and freely migrate video session back and forward from one device to another without interruption. We assume a content delivery path from the source to the end user. The path length is L , and there are overall L+1 nodes including the source node, the end user node L-1and intermediate nodes. Each Fig. 1. Multi-screen cloud social TV functionalities, such as ad insertion to architecture. support personalized multiscreen experience.We call such VM as cloud clone. Andwe assume each cloud clone can be operated at any cloud server along any overlay content delivery path. The location of cloud clone plays an important role on the cost of operating video teleportation. Specifically, the retargeted stream size changes as the user moves around or shift sessions from one device to another. It leads to the changes of the transmission cost, which further depends on the cloud clone location along the delivery path from the source to the user. Thus, there is an opportunity to reduce the operational cost by migrating the cloud clone to its best place. Fig. 2. System overview on cloud clone.

2) Key Feature: Our system offers a highly- B. System Models touted multiscreen experience via video In this subsection, we present three system teleportation.With this feature, one models to drive the problem formulation on user can simultaneously operate multiple minimizing the operational cost on support devices (e.g., TV and smartphone) and the video teleportation feature. freely migrate video session back and 1) Content Processing Model: In this paper, forward from one device to another without we assume the cloud clone performs two interruption. As a result, the users can content processing functions,including always stay connected to TV programs and advertisement insertion and video socialinteractions. transcoding.Through these processing 3) Cloud Clone: The enabling technology procedures, a content of size will be changed for video teleportation is to deploy a cloud to in the following two cases. clone as shown in Fig. 2 for each user. Case 1) : When the user is consuming the Specifically, every user is represented by a content on a bigger screen (e.g., TV), would virtual machine (VM), which serves as his be smaller than the retargeted stream size . proxy in the cloud to manage all the In this case, the cloud clone inserts a few associated devices and real-time session personalized video advertisements,by first information (e.g., ongoing programs, picking the related advertisement video information about the active device, the based on recommendation algorithms, then updated viewing history, and most recent combining the selected video ad overlay video segments, etc.). It elastically turns on with the targeted quality to the original once its represented user is online, and turns content [6]. Finally,the combined video off once the user gets offline. In addition, streaming is delivered to the end user. Such the VM also dynamically transcodes the method has become a key online original stream from the video source into monetization strategy to make profit [6]. At the one with appropriate format, resolution, the same time, some set-top boxes may not and bitrate as its user shifts the support the latest devices.Moreover, the VMalso offers other corresponding states. This model can be easily extended if there are more devices or states (e.g., when user is simultaneously using TV and phone, it is a new state). Since the collection of devices at home is limited, the number of states could be within a reasonable size. Fig. 3. User behavior on session migration Fig. 3(a) present a snapshot to capture a real by using video teleportation. use case that an user shifts his attention (a) Real user case. (b) Device switching between TV and smartphone by using model as a Markov chain. video teleportation. In particular, from the user’s perspective,he can transfer the video format (e.g., H.264 high profile) of the ongoing programs from TV to smartphone content source. Thus, the cloud clone has to by first scanning the TV screen using transcode the original content into a smartphone camera, then flipping the phone compatible one (e.g., H.264 baseline to trigger the transfer. He can also transfer profile), and pick the video ad overlay in the sessions back from his smartphone to the same format. As a result, the combined TV, by simply performing a “throw” content size could be bigger than the gesture. From the system operation aspect, original one. Note, the transmission cost for in both cases, the workflow is coordinated the cloud clone to load advertising videos is by the cloud clone.More specific details can ignored, since we assume the cloud has be found from our previous works [3], [4]. already pre-loaded all those data at each Fig. 3(b) illustrates a Markov processwith node. two states to model such user behaviors. Case 2) : When the user is consuming the Specifically, the state transition matrix is content on a smaller screen (eg.smartphone), completely determined by (the probability in would be larger than . Themain reason is which the user uses TV in both the current that, the video resolution and bitrate and the next time slot) and (the probability required by such a device are significantly in which the user uses smartphone in both lower than those of the original one. In this the current and the next time slot). case, the cloud clone transcodes the original 3) Cost Model: In supporting the cloud content into an appropriate format with clone migration, the media cloud would much smaller resolution and lower bitrate, to incur four cost components, including, fit the output. As a result,regardless of the • inserted advertisements with small Transmission Cost occurs when videos resolution and low bitrate, the retargeted are transmitted from the source to the user. stream size is still much smaller than the Note, when operating the media cloud, we original one. only need to consider the cloud network 2) User Behavior Model: We model user cost, while the cost incurred by access behavior across different devices as a networks (e.g.,broadband and wireless Markovian process, which has been widely network cost) will not be included in this adopted to characterize a variety of user work. behaviors on web browsing [7], online • Migration Cost corresponds to the social activities [8] and IPTV ineractions bandwidth cost in which the cloud clone [9].Without loss of generality, the user is migrates from one node to another within assumed to switch between two devices (i.e., the media cloud. TV and phone), thus the model has two • Computing Cost refers to the consumption very low level for each cloud clone to take of computational resources (e.g., the optimal action, because the number of CPU/GPU), when cloud clone processes possible locations along the shortest content user requests, transcodes the requested delivery path is usually small. contents into suitable format, and inserts advertisements into the original video. This cost component is a baseline cost, which is VI. CONCLUSION invariant of the cloud clone location. This paper investigated the problem on • Storage Cost is charged for keeping minimizing monetarycost via cloud clone advertising videos and user sessions. This migration in multi-screen cloud social TV cost is also a baseline cost, which is constant system. We formulated it as a Markov for different cloud clone locations.In this Decision Process,to balance a trade-off paper,we only focus on the transmission between the transmission and migration andmigration cost, and ignore other two cost. Under this framework, we first baseline costs which are independent considered a random fixed placement and an of our decision variable (i.e., cloud clone offline algorithm to obtain an upper and location). lower bound for the optimal cost. We then proposed a semi-online algorithm and a Algorithm:Online Algorithm with Q- more practical Q-learning method. We use Learning both simulated data and real user traces to Input: current system state at time slot all evaluate all the four algorithms.The results learned from Experience Replay phase indicated, up to 25% monetary cost Output: optimal policy at each time slot compared with the random fixed placement 1:for to do can be saved in typical use scenarios, by 2: apply optimally migrating the cloud clone. The 3: observe the new state , and the cost cost savings can be affected by the delivery 4: update path length, the VM migration size and the 5:end for user behavior pattern. Moreover, we also 6:return as the optimal policy found the optimal cloud clone location is This online algorithm with reinforcement either at the nearest or the furthest node to learning is the most practical one comparing the user. Those insights would offer with others. First, in real system, only the operational guidelines to deliver cost history of user behavior is accessible while effective multi-screen social TV services either the exact user traces in future or just over CCMN, potentially easing its adoption. the precise behavior model for each Our multi-screen social TV system has been individual user is difficulty or even implemented on top of a private cloud at impossible to obtain in advance. More Nanyang Technological University.It has importantly, this method adopts a model- been exposed to over 200 students for an free reinforcement learning technique. It internal trial.The next step is to deploy it to a indicates that even the transition probability vendor-neutral cloud provider(e.g., Amazon may not be necessarily deterministic. EC2) to achieve the large-scale deployment. This approach can still work when the Second, we plan to explore the possibility of transition probability is a random variable implementing other cloud application that its probability follows any distribution frameworks (e.g., one cloud clone for model. Second, the complexity of the multiple users). Finally, we will try to solve algorithms keeps at a more complicated problems (e.g., MDP with [5] F. Chen, K. Guo, J. Lin, and T. La Porta, constraints) in related scenarios. “Intra-cloud lightning: Building cdns in the cloud,” in Proc. IEEE REFERENCES INFOCOM, 2012, pp. [1] Y. Jin, X. Liu, Y. Wen, and J. Cai, 433–441. “Inter-screen interaction for session [6] T. Mei, J. Guo, X.-S. Hua, and F. Liu, recognition and transfer based on cloud “AdOn: Toward contextual centric media network,” in Proc. 2013 IEEE overlay in-video advertising,” Multimedia Int. Symp. Circuits Syst., 2013, pp. 877–880. Syst., vol. 16, no. 4–5, pp. [2] Y. Jin, T. Xie, Y. Wen, and H. Xie, 335–344, Aug. 2010. “Multi-screen cloud social TV:Transforming [7] Ş. Gündüz and M. T. Özsu, “A web page TV experience into 21st century,” in Proc. prediction model based ACM MM,2013, pp. 435–436. on click-stream tree representation of user [3] Y. Jin, Y. Wen, G. Shi, G. Wang, and A. behavior,” in Proc. ACM Vasilakos, “CoDaaS: An experimental SIGKDD, 2003, pp. 535–540. cloud-centric content delivery platform for [8] V. S. Tseng and K. W. Lin, “Efficient user-generated contents,” in Proc. IEEE mining and prediction of user 2012 Int. Conf. Comput., Netw., Commun., behavior patterns in mobile web systems,” 2012, pp. 934–938. Inf. Softw. Technol., vol. [4] Y. Zhang, C. Yan, F. Dai, and Y. Ma, 48, no. 6, pp. 357–369, 2006. “Efficient parallel framework [9 ] V. Gopalakrishnan and R. Jana et al., for h. 264/avc deblocking filter on many- “Understanding couch potatoes: core platform,” IEEE Trans. Measurement and modeling of interactive Multimedia, vol. 14, no. 3, pp. 510–524, usage of IPTV at large Jun. 2012. scale,” in Proc. ACM IMC, 2011, pp. 225– 242.