Research On Efficient Delivery Techniques For Cloud Services

Any application or service that is hosted at a remote location and accessed via theInternet or private networks are essentially acloud service. Some cloud services willaccompany with a large number of network traffic generated in the process of deliveringto geo-distributed end-users, e.g., the online services of Web and streaming whichaccount for the vast proportion of Internet traffic. The scale of data and users of suchSaaS cloud services (hereinafter referred to as cloud services) with heavy traffic hasgrown exponentially, the end-users’ experience and the complexity of end-users’accessing hehavior have been rised. Before the advent of the concept of cloudcomputing, in order to ensure the Quality of Services (QoS) of the Web and streamingservices, Content Delivery Network (CDN), which is constructed over the “Best Effort”Internet, has been used to deliver them. However, in the case of the diversified cloudservices presenting large-scale development and a large number of Cloud services’Flash crowd Phenomenon (CFP) emerged, CDN has become increasingly difficult tomeet the demand of the elastic scalability of cloud services. The contradictions betweenimproving the CDN resource utilization and ensuring the QoS of end-users havebecome increasingly acute.On the basis of analyzing the existenced problems of the actual cloud servicesdelivery systems, This paper extracts the following problems related to the sharpconflict between the QoS ensured and the CDN resource utilization Maximized.(1)“Adaptivity” problem, usually shared resources is used to pursue maximization ofstatistical multiplexing among multi-applications (i.e., cloud services), however themethod easily leads to multiple applications be mutual interference for seizing sharedresources. CDN is difficult to adaptive regulation to resolve the decline of QoS andresource utilization for mutual interference aomong multiple applications.(2)“Economy” problem, popular contents caused by CFP exceed the available cachingability and lead to the accessing cache hit rate declined, which refers to that cacheresources has become a new bottleneck. On the condition, the traditional dynamicresources allocation method, which only considers bandwidth resources as thebottleneck parameter is difficult to ensure more end-users with guaranteed QoS underthe same resources utilization ratio.(3)“Scalability” problem, CFP leads to theaggregated resource demand exceeding the upper bound of CDN service capacityfrequently. CDN is difficult to guarantee the QoS of the end-users resulting in theoverload.(4)“fairness” problem, CDN resources are usually priority assigned to thecloud services with high concern which refers to CFP happens. Therefore, when theresource demand is approaching the upper limit of CDN service capacity, CDN isdifficult to satisfy the resource demand of cloud services with low concern and ensure the QoS of the end-users. In response to these challenges, the main researchworks and contributions are listed as the following:(1) Surge CFP-oriented deployment method for multiple applicationsSurge CFP refers to a class of flash crowd that when it happens the resourcesdemand of cloud services from far below the level of service capacity instantly jumpmuch higher than the level of the service capacity. In response to the “adaptivity”problem, this paper presents Multiple Applications Co-Exist (MACE) method whichoffers a tradeoff between the CDN resource statistical multiplex maximized andavoiding multiple applications mutual interference maximized. According to the statisticlaw of cloud services’ resources demand, MACE classified multiple applications intodifferent types and the key mechanisms resource static allocation, dynamic supplementand resource reserved mechanism etc. were proposed, which can be scheduledadaptively. MACE was applied to a real large scale CDN infrastructure as well as6-month measurement and analysis were given. Compared with traditional method ofmultiple applications sharing bandwidth, the bandwidth utilization is increased by about20%, the balance coefficient which is inversely proportional to the efficiency of thesystem statisticaly is decreased from12%to5%, and the number of complaint eventsaffecting the dependability of CSDN services caused by multiple applications’mutual-interference has been dropped to0.(2) the dynamic allocation model of edge resources associated with popular contentgenerated by CFPIn response to the “economy” problem, this paper formulated the dynamicallocation problem of edge resources into Multidimensional Facility Location Model(MFLM), considering the cache resource as the bottleneck resource. MFLM is aNP-Complete problem, a heuristic algorithm on the basis of increasing some necessaryassumptions and constraints violation were roposed in this paper. The key of thealgorithm is multi-cache coordination and Peer-assisted, which is in response to thespecial case of the dynamic construction of logial sub server overlay network forstreaming and file sharing applications etc. On the basis of the performance metrics andperformance analysis model proposed for the heuristic algorithm of MFLM, we gave acomprehensive evaluation of the performance of the algorithm.(3) Slow CFP-oriented “peak clippling” moderatelyIn response to the “scalability” problem, Dynamic Moderate Peer-assisted Method(DMPM) was presented. DMPM is mainly response to the slow growth CFP whichrefers to the resources demand increased to not less than the level of the service capacitywith slow speed. DMPM adopted time series analysis method and radial basis functionneural network to predict and decide when and how much CDN server loads need to be offloaded, and a novel P2P mechanism were designed to improve the scalability ofCDN. Furthermore, we use the real traces collected from an actual CDN entity toevaluate the accuracy of the two prediction method as well as the effectiveness of theDMPM.(4) The elastic construction method for overlay network with CFP effect restrainedMost or all of CDN resources are allocated to CFP and the resources demand ofcloud services with low concern are usually difficult to be satisfied. The phenomenon iscalled CFP effect. In response to the “fairness” problem under CFP effect, This paperused the VoD application as the research object and presented Elastic ConstructionMethod (ECM) for P2P VoD overlay based on CDN. ECM used the Peer-assistedmethod to achieve improving the scalability of CDN, since ECM ordered the peers indifferent CDN node who were accessing the same content collaboratively to constructP2P overlay. ECM solves the peer resource scarce problem through elastic scales theconstruction scope of P2P overlay. Meanwhile in order to solve the overloading of thebachbone network traffic generated by P2P streaming, ECM chosen the scale path basedon the distances among CDN nodes. From the simulation, compared with isolatedoverlay construction inside a single CDN node ECM improves the acceptance ratio ofend-users’ requests by41%while compared with pure P2P system without on the basisof CDN infrastructure, the traffic crossing backbone has been cut down obviously.In summary, for the new challenges caused by the cloud services development oflarge-scale, this paper give some researches about many key technologies related withCFP, which has important theoretical and actual significance in achieving efficientdelivery of cloud services with guaranteed QoS.