COST-EFFECTIVE LOW-DELAY DESIGN FOR MULTI-PARTY CLOUD VIDEO CONFERENCING
ABSTRACT
Multi-party cloud video conferencing architecturehas been recently advocated to exploit rich computing andbandwidth resources in cloud to effectively improve video conferencingperformance. As a typical design in this architecture,multiple agents, i.e., virtual machines, are deployed in differentcloud sites, and users are assigned to the agents. Then, theusers communicate through the agents, and the agents mighttranscode the recorded videos given the heterogeneities adevices in terms of hardware specification and connectivity. Inthis architecture, two critical and nontrivial challenges are: (1)assigning users to agents to reduce the operational cost andthe user-to-user conferencing delay, (2) identifying best agentsto perform transcoding tasks, taking into account the heterogeneousbandwidth and processing availabilities. To address thesechallenges, we cast a joint problem of user-to-agent assignmentand transcoding-agent selection. The ultimate objective is tosimultaneously minimize the cost of the service provider andthe conferencing delay. The problem is combinatorial in naturewhich belongs to the NP-hard node assignment problems. Weleverage the Markov approximation framework and devise anadaptive parallel algorithm that finds a close-to-optimal solutionto our problem with a bounded performance guarantee. Toevaluate the performance of our solution, we implement aprototype video conferencing system, and carry out trace-drivenexperiments. In a set of large-scale experiments using PlanetLabtraces, our solution decreases the operational cost by 77% andsimultaneously yields lower conferencing delay compared to anexisting alternative.
EXISTING SYSTEM:
Multi-party Video Conferencing. Previously, P2P architecture was considered as an alternative to traditionalclient/server architecture. In , following network utilitymaximization framework, a problem of video conferencingin P2P architecture has been studied. In another recent work the authors propose a scheme to maximize video qualityunder uplink-downlink capacity constraints for peer-to-peermulti-party video conferencing. However, in P2P there is nopowerful server in the architecture which hinders executing thehigh demand tasks such as transcoding. The cloud architecturefor video conferencing has been proposed in Airlift [19] forthe first time, and it suggests to use cloud bandwidth resourcesto boost the conferencing experience. In vSkyConf thecomputational resources of the cloud is exploited for executingprocessing tasks, in addition to the dedicated cloud communicationinfrastructure. These studies assume nearest assignmentpolicy, which is not optimal in a multi-party applicationin terms of intra-cloud traffic and user-to-user conferencingdelay. Two recent studiespropose different serverselection/placement and topology control approaches to onlyminimize the latency in transcoding-free video conferencing,without taking into account the operational cost. Finally,the delay-constrained video streaming in different networkinfrastructures and applications has been studied previously,e.g., in wireless and wireline networks
PROPOSED SYSTEM:
Almost all of the existing studies that we are aware,neglect to consider the cost to the service provider andsimply adopt the nearest policy for user-to-agent assignment. To the best of ourknowledge, this study is the first that aims to improve the cloudvideo conferencing design by tackling the user-to-agent andtranscoding assignments problem in a unified combinatorialoptimization framework. The main contributions of this studyare summarized below.B We formulate the User-to-agent Assignment Problem(UAP), that tries to select the user-to-agent assignment andtranscoding assignment in a multi-party application with theobjective of simultaneously minimizing the operational cost tothe service provider and conferencing delay. The problem issubject to the capacity constraints of the diverse agents and theuser-to-user conferencing delay constraints. The problem is anonlinear combinatorial optimization problem in the categoryof NP-hard node assignment problems which are difficultto solve due to persistent dynamics in the system and largeproblem size.B We leverage the Markov approximation framework which is a technique for solving the combinatorial networkproblems in a distributed fashion. We devise an efficientparallel and iterative algorithm to solve the UAP, which runslocally in a representative agent of each session and convergesto a close-to optimal assignment. The algorithm adapts to thesystem dynamics, provides a bounded approximation gap, andis robust against the inaccurate measurements of the problemdata. In addition, we improve the convergence of the algorithmby proposing another initialization algorithm called AgRank,which is a simple scheme with low complexityWe implement a cloud video conferencing system prototypeusing Amazon EC2 platform and also carry out tracedrivenevaluation experiments using PlanetLab nodes. Theresults demonstrate the significant improvement of our solutioncompared to the existing alternatives. In a representative experimentalsetting of PlanetLab traces, our algorithm outperformsthe nearest assignment policy by reducing theoperational cost and the delay by 77% and 2%, respectively
CONCLUSIONS AND FUTURE DIRECTIONS
This study addresses the problem of user-to-agent assignmentand transcoding task assignment in cloud videoconferencing architecture. Considering the challenges of theproblem due to the underlying large-scale combinatorial problem,we devise a parallel and adaptive solution to optimizethe assignment tasks. The algorithm achieves a suboptimalsolution with a bounded performance guarantee. Observationson prototype system implementation corroborate our claim thatuser assignment is a critical design issue in cloud architecturethat can lead to a big difference in entire system performance.In addition, trace-driven simulations demonstrate that oursolution design outperforms the existing solutions in terms ofreduced delay and cost, and thus demonstrates its viability asa win-win solution for both users and conferencing servicengbr provider. Finally, in future research, a promising directionto tackle is the more general problem in which other tasks,rather than transcoding are performed at the cloud agents. Theproblem could be further generalized to consider other types ofcommunication and computation cloud resources for generalinteractive real-time multimedia applications.ACKNOWLEDGMENTThe work presented in this paper was supported in partby National Basic Research Program of China (Project No.2013CB336700) and the University Grants Committee of theHong Kong Special Administrative Region, China (Area ofExcellence Grant Project No. AoE/E-02/08 and CollaborativeResearch Fund No. C7036-15G), and the National NaturalScience Foundation of China under Grant No. 61402247, andHong Kong RGC grants 718513, 17204715, and 17225516.
REFERENCES
[1] Amazon elastic compute cloud, http://aws.amazon.com/ec2/.
[2] http://opencv.org/.
[3] Cisco VNI service adoption forecast, 2012–2017. White Paper, Febru-ary, 2013.
[4] Cisco VNI global mobile data traffic forecast update, 2013–2018. WhitePaper, February, 2014.
[5] B. Alinia, M. H. Hajiesmaili, and A. Khonsari. On the construction ofmaximum-quality aggregation trees in deadline-constrained WSNs. InIEEE INFOCOM, pages 226–234, 2015.
[6] D. G. Andersen. Theoretical approaches to node assignment. TechnicalReport, 2002.
[7] P. Bailis, A. Davidson, A. Fekete, A. Ghodsi, J. M. Hellerstein, andI. Stoica. Highly Available Trans.: virtues and limitations. In VLDB,2014.
[8] M. Bianchini, M. Gori, and F. Scarselli. Inside PageRank. ACM Trans.on Int. Tech., 5(1):92–128, 2005.
[9] H. Bobarshad, M. van der Schaar, A. Aghvami, R. Dilmaghani, andM. Shikh-Bahaei. Analytical modeling for delay-sensitive video overwlan. IEEE Trans. on Multimedia, 14(2):401–414, April 2012.
[10] S. Boyd and L. Vandenberghe. Convex Optimization. CambridgeUniversity Press, 2004.