Research And Implementation On Optimization Of Network Traffic In Cloud Computing

Development of Cloud Computing needs well-built network infrastructure of Datacenter.Optimization of intra-Datacenter traffic is the key of building Datacenter for improving Quality of Service(Qo S).The main body of optimization of traffic is reducing the Flow Completion Time(FCT)and this paper aims to reduce FCT by load balancing mechanisms.This paper researches load balancing mechanisms on two most typical traffic in Datacenter: normal traffic and distribute deep learning traffic for reducing FCT.This paper leverages their characteristics and proposes two brand new load balancing mechanisms to overcome problems of existed techniques.New mechanisms are both proved to be effective by simulating on network simulator NS3.The main contributions of this article are listed below:This paper proposes a new load balancing granularity: “flowlet-packet-mixing” and designs a Flow-aware and Mixed Granularity(FAMG)load balancing based on“flowlet-packet-mixing” granularity.FAMG combines good features of packet granularity and flowlet granularity and overcomes the challenging of mixing the two different granularities.FAMG routes elephant flows per-flowlet while routes mice flow per-packet.The phenomenon of two granularities coexisting in network which enhances the robustness in asymmetric topology and gets more theoretical performance in symmetric topology.In addition,FAMG focuses on reducing FCT of delay-sensitive mouse flows.This paper proposes Tailor,a load balancing mechanism based on hashing network-layer information to reroute flows,which avoid flow stragglers since performance bottleneck of distributed deep learning traffic is tail FCT.Tailor finds out flow stragglers by recording the ratio of sent amount and reroutes these flows to less congested link to chase up other flows which reduces tail FCT.Results show that tail FCT decreases about 10%～50% compared to popular load balancing mechanisms.Tailor accelerates training time of distributed deep learning by reducing tail FCT.This paper implements a large-scale Datacenter load balancing simulation module for simulating realistic traffic generation to get correct results and decreasing simulation complexity by redeveloping on NS3.Meanwhile,RPS、Let Flow、FAMG、Tailor and other load balancing mechanisms are implemented in our simulation module.This module can covers the whole work flow in load balancing including generating different network topologies,simulating traffic by some kind of random process on traffic in different distribution.This module provides a convenient researching tool for ourselves and other researches.In summary,this paper makes contribution to optimize traffic in cloud by researching datacenter load balancing.