Research On The In-network Caching And Evaluation In Smart Identifier Networks

The current Internet originates in the Arpanet,which is designed to transfer and share data among wired and trusted terminals at the end of the 1960s.With its rapid development,the scale of its users and applications has continued to expand.The original designs have been severely challenged and faced with many problems that are difficult to solve.Recently,countries around the world have conducted relevant researches on the future network architecture and tried to fundamentally solve various major problems of the existing Internet.The National Engineering Laboratory for Next Generation Internet Technologies(NGIT)of Beijing Jiaotong University proposes the SINET(Smart Identifier NETworks)architecture,which can effectively improve network security,mobility,programmability,and scalability.Meanwhile,SINET introduces in-network caching,then effectively improving the user access experience.In-network caching can greatly increase the efficiency of link bandwidth and reduce user access delay,which is one of most important research topics in SINET.To promote the development of caching mechanisms,this dissertation comprehensively studies the SINET architecture,proposes an SINET caching framework,and summarizes caching mechanisms' performance indicators and evaluation methods.On this basis,to reduce caching redundancy and improve the caching efficiency and hit ratio,a caching mechanism based on family collaboration is proposed.To accelerate the throughput of caching system,a hierarchical high-speed caching system is designed.In addition,to simplify the SINET evaluation experiments and promote its research and development,a distributed emulation system for SINET is proposed.The main works and innovations of this thesis are as follows:(1)The SINET architecture and researches are surveyed,and an SINET caching framework is proposed.First,the SINET architecture is deeply studied.Its mechanism principle and research works are comprehensively analyzed,including the SINET reference model,design principles,state-of-the-art,and challenges.Then,a scheme of SINET caching framework is proposed.Caching performance indicators and evaluation methods are summarized.All of these establish the foundation for works in the subsequent chapters.(2)To achieve the maximum caching revenue,a family-based collaboration caching mechanism is proposed,named FCache.On the basis of the actual network topology,FCache can automatically and dymanically create caching function families(groups).Each caching function family performs the maximum revenue caching algorithm to maximize the efficiency of the caching system.The FCache prototype system based on an NetFPGA 10G board is implemented.The single-point throughput experiments show that FCache can well support high-throughput'caching scenarios.At the same time,the multi-point evaluation experiments are done with a self-developed testbed.The results show that FCache can effectively improve the cache hit rate(15%-30%in typical scenarios)and reduce the access delay.(3)To improve the throughput of caching systems,a hierarchical high-speed caching system is designed,named HCaching.Its core idea is that,utilizing fully the locality of network traffic and DRAM read-write characteristics,it uses a small amount of SRAM as the buffer of the main memory DRAM,then accelerating DRAM read-write throughput,and improving the overall speed and capacity of the caching system.HCaching scheme includes a large-capacity storage organization optimization scheme,a high-speed prefetching mechanism,and an efficient bloom-filter indexing technique for caching.It is evaluated by a self-developed discrete event simulator based on NS-3.The results show that its throughput is 3-10 times faster than current mainstream cache mechanisms.(4)Based on the above two self-developed caching evaluation methods,an evaluation system for SINET is designed and implemented,named EmuStack.The system adopts a cloud computing system architecture and uses the "test-as-a-service"design concept.It can support large-scale,full-stack protocol network performance evaluation.EmuStack uses lightweight computing virtualization technology to achieve an experimental scale of 100 times the physical hosts.The system supports the dynamic control of virtual link delay,delay jitter,bit error rate,etc.The control range of link delay is from 1ms to 200s,and the control range of link error rate is 10-4 to 10-9.Meanwhile,it supports multi-tenancy,dynamically orchestrating technology,and verification of integration of discrete,virtual and physical experiments.In particular,to support the evaluation of high-speed such as 100Gbps caching throughput,an NS-3-based evaluation module is proposed,whose public version is open-source.