Traffic Scheduling For Cloud Data Centers

Recently,as the services,such as big data,web search,cloud com-puting,develop rapidly,data center infrastructures and technologies face many new challenges.Supporting these services require a large number of servers to process data in parallel,and the network needs to provide high bandwidth for servers.Thus,the bandwidth and network latency of the data center directly determines the performance of services.However,due to the special topologies of the data center and the diversity of services,the traditional traffic scheduling mechanisms have poor performance in data centers and cause many problems.For example,due to multi-path,load imbalance can lead to path congestion;since data centers serve multiple services at the same time,it is difficult to meet the latency requirement of latency-sensitive applications;due to inefficient coflow scheduling,the coflow completion time of big data applications is too large.Therefore,this dissertation studies the traffic scheduling problems in cloud data cen-ters and has the following contributions:1.A Flow Distribution-Aware Load Balancing(FDALB)mecha-nism is proposed.By leveraging the good scalability of distributed schemes and good performance of centralized schemes,FDALB re-duces the control overhead of centralized control and increases the scalability.The simulation results show that FDALB can reduce the network latency and avoid link congestions.2.A Traffic Prediction based Flow Scheduling(TPFS)scheme is designed.It is hard to timely identify latency-sensitive flows for information-agnostic flow scheduling.Thus,based on the observa-tion that flow size variation of one type of application is small,this dissertation proposes TPFS to predict flow size according to the type of applications and to adjust the thresholds of priority queues more accurately.The testbed evaluations show that TPFS can reduce the completion times for latency-sensitive applications.3.A congestion-aware coflow scheduler,SkipL,is proposed.To solve the agnostic to network congestion,this dissertation designs an end-host based congestion detection algorithms to quickly detect the re-maining bandwidth.Furthermore,based on the remaining bandwidth,this dissertation proposes a congestion-aware bandwidth allocation al-gorithm.The evaluation results show that SkipL can timely detect the link congestion and measure the remaining bandwidth of paths.4.A Multiple-attributes-based Coflow Scheduling(MCS)mecha-nism is proposed.The large coflow completion time is mainly caused by head-of-line blocking and improper priority thresholds.Thus,this dissertation designs a coflow scheduling mechanism MCS to sched-ule coflows according to the coflow width,length,and other infor-mation.As a result,MCS can avoid head-of-line problem and more accurately schedule coflows.The evaluation results show that MCS can efficiently reduce the coflow completion time of small coflows.