A Research On Datacenter Network Congestion Control Technology Based On In-network Telemetry(INT) Information

As the link speed of datacenter network reaches 100 Gbps,the conflict between throughput and latency becomes more pronounced.Applications such as highperformance computing,distributed deep learning,and data mining have put forward higher requirements on the performance of datacenter networks,which are mainly reflected in the expectation of higher throughput and lower transmission delay.The congestion control algorithm proposes a systematic solution to this contradiction.In recent years,with the development and popularization of switches that support the InNetwork Telemetry(INT)function,the INT-based congestion control algorithm attracts more and more attention because of its characteristics of accurate congestion control with accurate link state information.However,the current INT-based congestion control algorithm still has some defects.On the one hand,since the congestion control algorithm relies on ACK packets to transmit INT information,and its response time is at the RTT level.Therefore,the algorithm is insufficient to control the short flow.Especially for the short flow that can be transmitted within one RTT,the existing congestion control algorithm is completely out of control.On the other hand,the datacenter network has abundant multi-paths,but existing congestion control solutions only focus on single-path congestion,which causes the waste of network resources and limits the improvement of network transmission.This research proposes more complete solutions for above defects-UECC(UltraPrecise and Efficient Congestion Control)and MPICC(Multi-Path INT-based Congestion Control).UECC is an agile INT-based congestion control algorithm.It proposes a cooperative INT information feedback mechanism using switches and hosts,which manages network sending behavior within the first RTT of the flow and control the control information overhead.MPICC is a multi-path INT-based congestion control algorithm,which solves problems such as multi-path coordination and out-of-order data packets at the receiver caused by multi-path.This work conducts extensive and detailed experiments on UECC and MPICC.The experimental results show that compared with HPCC,on Cache Follower and Web Server workloads,UECC reduces the average queue length of the network by at least 42%;On the three workloads of Cache Follower,Web Search and Web Server,MPICC reduces the average flow completion time by 20.1%,14.4% and 39.5%,and reduces the 97 thpercentil flow completion time by 39.9%,18.9% and 57.9%,respectively.