| Content Delivery Network(CDN)is important information service infrastructure.With widely distributed edge servers and client mapping mechanisms,including traditional DNS-based and novel anycast-based mappings,it provides low-latency and reliable content service for end-users,as well as efficient means for content providers to scale quickly and gracefully.However,its performance is still subject to various constraints,which emphasizes the importance of performance management.The critical management tasks include detecting and digging root causes for unsatisfactory cases via wide-area network measurement,and further forming optimization schemes.To improve CDN performance,based on two client mapping mechanism,this dissertation takes operational commercial CDNs as real scenario to study the above challenges.The main research content and contributions are summarized as follows:1.By using a fine-grained site-granularity method to measure DNS-based CDN redirection behavior,we perform a periodical high-bandwidth measurement targeting at4600 sites delivered by dominating Chinese and global CDNs.We identify inefficient scenarios and causes in global CDN services due to inconsistent local technical and regulation policies.Specifically,Chinese CDNs show almost static dynamics at customer site granularity with region-specific services.This leads to a well-known multi-CDN selector,which uses synced coarse-grained scheduling,providing suboptimal performance.Besides,due to content delivery policy in China,some global CDNs partner with Chinese CDNs to use their edge servers for an order of magnitude better latency;otherwise they may have unstable performance: although schedule efficiency is consistently good among regions with different node density,the latency of selected server is much worse than the best available one in about 50% cases in regions with sparse node,suggesting the scheduling algorithms need to take absolute performance into consideration more strictly.2.Aiming at the largest anycast CDN Cloudflare,we conduct a global client mapping and traceroute measurement.We evaluate the mapping proximity and the impact of anycast efficiency on latency performance.We further propose a facility-granularity inefficient anycast routing pathology,and diagnose inefficient anycast paths.The results show good overall proximity.In a few unsatisfactory cases,the inefficient anycast routing could lead to 5-times inflated path latency.By quantitatively representing the scale of direct peering networks as the number of corresponding facilities,we find that several huge direct peering networks whose scale rank at top 3% have important impact on inefficient anycast paths.Most of inefficient paths are caused by inter-domain inflation——upstream ASes prefer these huge direct peering networks instead of local ones;additionally,these huge direct peering network could also have intra-domain inflation,thus deserving priority focus when troubleshooting.3.We propose a distributed hybrid scheduling model for load management to mitigate load unawareness problem in anycast CDNs.Based on anycast CDN architecture and client localization,we design a cooperative distributed algorithm ACCO;each control node decides client mapping locally based on local status and communicated status to achieve globally load management.We further propose 2 heuristic algorithms.Using measurement data as simulation settings,we evaluate the feasibility and efficiency of proposed algorithms.Results show that ACCO could converge to the solution as central method with better scalability;heuristic algorithms could further decrease response time by sacrificing a little accuracy. |
