Research On Performance And Efficiency Optimization Technologies In Data Center Networks

As the rapid development of cloud computing,cloud platforms and cloud applications have deeply impacted people's lifestyle.Data centers,as the core infrastructure of cloud computing,have received extensive attention from academia and industry in recent years.In a data center,the data center network(DCN)plays the backbone role in the aspects of communication,computing and storage.Therefore,this paper mainly studies how to solve these three aspects of problems by using performance and efficiency optimization technologies in DCN.In communication,multicast communication includes many important applications,such as serializable session/data synchronization,large file copying,MapReduce,the operating system/application upgrade.Existing multicast schemes have redundant links;In computing,virtual machine placement has a direct impact on the operation of the cloud platform and the service quality of cloud tenants.The existing schemes do not take into account both the correlations of virtual machines and the service level agreement of tenants.In storage,redundant traffic is a sort of non-negligible traffic,which can be eliminated by caching mechanism.In this thesis,the research contents mainly include multicast routing mechanism,virtual machine placement,and caching mechanism.At last,the architectures that are suitable for performance and efficiency optimizations are showed by comprehensive analysis and comparison for data center network architectures.Major research results include:1.With the 60 GHz millimeter-wave wireless technology,NEMO,an efficient multicast routing mechanism is proposed,including both NEMO-Group and NEMO-Cluster algorithms.The expected multicast performance improvement of the two algorithms is theoretically analyzed,and the upper and lower bounds of wireless link reduction are estimated.The simulation results show that NEMO can reduce traffic by15%-40% compared with the existing DCN multicast scheme,and improve throughput by 10%-40%.2.On one hand,a correlation-aware virtual machine placement scheme is proposed,including resource utilization prediction model and three correlation-aware algorithms of virtual machine placement.The simulation results show that the proposed virtual machine placement algorithms can improve the resource utilization by 15%-30% compared with the existing algorithms,and satisfy the service level agreement defined by the tenant.On the other hand,a bandwidth-aware virtual machine placement problem is formalized.A bandwidth balancing virtual machine placement algorithm based on communication locality is designed.The simulation results show that the algorithm is superior to the greedy algorithm that only considers communication locality and the longest processing time algorithm that only consider s bandwidth balancing.3.Piche,an efficient caching mechanism is presented,including a network protocol,node cache management mechanism,cache sharing mechanism and cache placement mechanism.The extensive simulations and Amazon EC2 experimental results show that the Piche is scalable and suitable for many different topologies.It can reduce around 40% redundant traffic on average,and only increase about 10% runtime.4.DCN architectures are comprehensively analyzed and compared to implement performance and efficiency optimizations better.According to the literature of the last decade,the DCN architectures are comprehensively analyzed,such as connection type,circuit layout,interconnection equipment,and network features.Qualitative and quantitative analyses are performed on the aspects of communication,computing and storage in data center networks to indicate the architectures that are proper for performance and efficiency optimizations.The indexes includes scalability,total cost,energy efficiency,bisection bandwidth,network diameter and so on.At last,the architectures with wireless antennas and optical devices are also discussed.