High Throughput Low Latency Congestion Control Method In Data Center Networks

Since cloud services are more sensitive to flow completion time(FCT),researchers have proposed many congestion control mechanisms to reduce the average flow completion time in data center networks.Theoretically,the Short-est Remaining Time First(SRTF)scheduling policy can achieve the minimum FCT.However,existing data center congestion control mechanisms either do not achieve global-optimal SRTF or are difficult to be deployed.The devel-opment of datacenter applications leads to the need of microsecond latency for end-to-end communication.As a result,RDMA is becoming prevalent in data-center networks to mitigate the latency caused by the slow processing speed of the traditional software network stack.However,existing RDMA congestion control mechanisms are either far from optimal in simultaneously achieving high throughput and low latency or in need of additional in-network function support.In this thesis,we analyze the challenges of approximating global-optimal SRTF scheduling in a distributed transport control protocol.Then a credit-based distributed data center congestion control mechanism,Helm,is proposed.Helm solves the challenges by combining the proposed dynamic priority assign-ment and flow-size-based rate control algorithms at receivers and thus achieves near global-optimal SRTF without modifying commodity switches.Theoretical analysis shows that the performance of Helm is close to SRTF.Besides,exten-sive simulations are conducted and the results show that Helm reduces the mean and tail FCT by up to 62%and 75%respectively compared with Homa in an oversubscribed data center network.Besides,by leveraging the observation that most congestion occurs at the last hop in datacenter networks,we propose RCC,a receiverdriven rapid con-gestion control mechanism for RDMA networks that combines explicit assign-ment and iterative window adjustment.Firstly,we propose a network conges-tion distinguish method to classify congestions into two types,last-hop conges-tion and innetwork congestion.Then,an Explicit Window Assignment mecha-nism is proposed to solve the last-hop congestion,which enables senders to con-verge to a proper sending rate in one-RTT.For in-network congestion,a PID-based iterative delay-based window adjustment scheme is proposed to achieve fast convergence and near-zero queuing latency.RCC does not need additional innetwork support and is friendly to hardware implementation.In our evalu-ation,the performance of RCC is 9?79%better than ExpressPass,DCQCN,TIMELY,and HPCC.