Service Feature-Aware Resource Scheduling In CDN Network

Content delivery network(CDN)is an important architecture for Internet data distribution.The CDN deploys the files requested by users to the edge cache server that is close to users so that the user can directly download the required files from the edge cache server.As a result,the CDN reduces the traffic pressure on the backbone network and reduces the time delay for users to download files.The main service scenarios of CDN include two categories.The first is the current video file distribution,and the second is to provide executions for emerging artificial intelligence applications that use deep neural network(DNN)models.CDN not only provides file distribution service,but also provides computing service.The service of CDN has gradually evolved from traditional file distribution to integrated file distribution and computing.With the growth of CDN service volume,the improvement of CDN performance is facing resource bottlenecks,which mainly come from the cache resource and the bandwidth resource that distributes files to users.This thesis will design corresponding resource scheduling mechanisms to improve the resource utilization of CDN and the quality of experience of users according to the resource requirements of different CDN services.The main contribution of this thesis is as follows.(1)This thesis proposes a joint optimization model of content-aware CDN load balancing mechanisms,which directs different video requests to different edge cache servers and decides the video bitrates for each request,to maximize the overall user experience under the constraint of server bandwidth resource.The basis behind the joint optimization model is that the user's experience of a video is a non-linear increasing function of bitrate,and the parameters of the user experience function are different for different videos.Therefore,when the CDN distributes videos,the overall user experience can be improved by jointly planning the direction decision of each edge cache server and the bitrate of each video.This thesis then designs a greedy heuristic algorithm to solve the optimization model.The algorithm first treats all edge cache servers as a virtual ideal server to obtain the optimal bitrate of each video.Then the video is iteratively assigned to an edge cache server according to the optimal bitrate.Experimental results show that the proposed load balancing algorithm can improve user experience by 3%-70%compared with the existing algorithms.(2)This thesis proposes a cache optimization model for artificial intelligence applications,which selects and caches the DNN model used by artificial intelligence applications into the limited GPU memory of the edge cache server to minimize the average response time of applications.The basis behind this optimization model is that an artificial intelligence application may concurrently use multiple DNN models to achieve its final function,and a DNN model may be used by multiple artificial intelligence applications,only when all the DNN models required by an application are cached in the GPU,the application can obtain short response time.Therefore,using the association relationship between the artificial intelligence applications and the DNN models to cache can reduce the overall response time.This thesis then designs a greedy heuristic algorithm,which iteratively selects applications and caches the associated DNN models in the GPU memory.Experimental results show that the proposed caching algorithm can improve the cache performance by about 10% in most cases.