Throughput And Latency Optimization For Applications In Data Center Networks

In the past few years,more and more companies have their own data centers.Some companies such as Amazon,Microsoft,and Google even provide data center service.In this case,how to use cheap and common network devices to provide low-latency,high-bandwidth services to the applications of data center is an important problem.TCP is the most widely used transport protocol in the Internet.Although TCP is widely used in data centers,the traditional TCP transport protocol has following problems:(1)TCP s congestion window will be halved when there is congestion in the network and this will lead the network utility low.(2)Queue delay of TCP is too long to satisfy the latency demands of applications.(3)TCP is a fair allocation method and fair allocation can not meet the demands of applications on bandwidth and latency.(4)TCP can not meet the requirements of task level transmission.As a result,it is necessary to optimize the transport of applications from flow level and task level.Also,the different bandwidth and latency demands of applications should also be taken into consideration.Based on these,we make the following contributions in this thesis:(1)We propose flow transfer model and task optimization model.With the ECN-based flow level model,parameters of rate control methods can be analyzed.Based on the task-level model,we propose Idealized Weighted Coflow Completion Time Mini-mization(IWCCTM)problem.To solve IWCCTM problem,we propose a 2-approximate optimal method to solve.(2)For the optimization of flow level,we propose the load adaption protocol-LPD and the flow duration time rate control protocol-FDRC.As most policies lose efficacy under heavy congestion,LPD advocates that both network condition and deadline should be taken into consideration when designing rate control algorithm.With the principle"more load,more differentiation",LPD works well even when the network is under heavy load.Due to the traffic is the mixture of deadline sensitive and short flows,modern methods can not meet the requests of the two kind flows simultaneously.FDRC proposes to use flow duration time when designing congestion control methods.As a result,both the percentage of flows missing deadline and flow completion time reduce.(3)For the optimization of task level,we propose the centralized schedule methods D-Target and Yosemite.With the information of flows and for the problem of TCP transportation as well as random source selection,D-Target can reduce average File Access Time(FAT)significantly.As the scheduling methods ignore the importance of coflows which have poor performance.Yosemite considers both the network condition and the importance of coflows,as a result,the performance of important coflows improves.(4)We design and evaluate Fly Transfer,a system that can schedule the transfer of flows as well as tasks.We deploy Fly Transfer at the platform of openstack and other real world platform.Evaluations show that Fly Transfer can optimize the transfer of tasks and flows with small overheads.We also test the overhead of Fly Transfer in this thesis.