Research On Distributed Machine Learning Oriented Optical Network Control Mechanism

With the continuous development of information technology and the use of computing accelerators such as GPU,computing performance has been greatly improved,which also means that a large amount of data is transmitted through the network every minute.As for computing tasks,the bottleneck is transferred from computing to communication.Distributed machine learning is a kind of common high performance computing tasks,when the amount of training data is very big,mass communication between nodes may lead to the emergence of network bottlenecks,the existing network is a kind of general design,network structure rigidity,unable to provide flexible support for distributed machine learning ability,it is difficult to dynamically adjust the network adapter with distributed machine learning.Distributed in order to solve the underlying network is difficult to perceive adaptation of machine learning,the network bottleneck resulting in performance degradation problem,this thesis research and design for distributed optical network control mechanism of machine learning,natural flexibility and ability to use optical network SDN technology centralized control characteristics of network resources,aiming at the performance requirements of distributed application of machine learning,adaptive adjustment,according to the current network status of the network designed to be flexible and controllable,can be in the direction of the machine learning needs fit closely distributed in exchange pattern,topology,adjust the allocation of resources,etc,to improve their application performance.The thesis for distributed machine learning has carried on the experimental analysis of network bottlenecks,this type of application based on SDN optical network control mechanism was designed and implemented,including the choreographer and the network controller design,custom nets yuan YANG model design,the design of the north and south to deal,finally set up the experimental platform is related to verification,proved that the distributed optical network control mechanism of machine learning oriented feasibility and validity.In addition,this thesis in view of the scene at the edge of the computing and cloud computing,designs and realizes a kind of computing and network resources scheduling choreographer,the use of computing resources and network resources information for joint scheduling,avoid the local hot spots and unbalanced load,and so on and so forth which led to the decrease of the quality of the user service.Finally,an experimental platform was built to verify the effectiveness of the proposed method.