Research On Rate-Adjustment-Based Congestion Control In Lossless Datacenter Networks

The high-speed datacenter networks provide a strong backbone of infrastructure for efficient processing and transmission of massive amounts of data.With the rapid growth of link speed in datacenter networks,the queuing delay is gradually replacing the propagation delay as the dominant component of end-to-end delay and flow complete time.Datacenter congestion control mechanisms need to achieve faster convergence in severe congestion scenarios,thereby reducing congested switch buffers as quickly as possible to reduce queuing delay.The existing end-to-end congestion control mechanisms rely on different congestion signals to obtain the state of networks for rate adjustment.However,these congestion control mechanisms are either not precise enough for rate adjustment due to coarse granularity of congestion signals,which leads to long convergence periods,or difficult to deploy due to the high reliance on network hardware devices.To address the above issues,the main work of this paper is as follows:First,this paper analyzes and compares the existing end-to-end congestion control mechanisms in terms of convergence speed,congestion signal granularity and hardware dependency.Then this paper proposes an ECN-marking-ratio-based Queue Buffer Occupancy Estimating(QBOE)scheme and a Queue-building-rate Driven Rate Adjustment(QDRA)mechanism for obtaining accurate network state information with coarse-grained signals.Second,this paper proposes DQCC,a fast converging rate-adjustment-based congestion control mechanism.DQCC estimates the queue length at receivers and adjusts sending rates according to target rate at senders.DQCC drains out the queue buffer occupancy in a minimum of one control period,achieving fast convergence with de-switch dependencies.Then this paper evaluates DQCC via OMNe T++ event simulator and compares DQCC to other mechanisms.Simulation results show DQCC improves about 56%-65%in terms of convergence speed,reduces about 15%-27% in terms of average flow completion time,and decreases about 36%-57% in terms of average queue length compared to DCQCN and PCN.Third,this paper presents a lightweight congestion control simulation verification architecture based on Data Plane Development Kit(DPDK-CC).DPDK-CC introduces a scheme to implement the core components of the switch-side and host-side for congestion control mechanisms in datacenter networks with DPDK.In addition,this paper implements the deployment and related experiments of DQCC based on DPDK-CC,and verifies the feasibility of DPDK-CC design and the correctness of DQCC.