Research On RTT Based Congestion Control Algorithms For Stable Low Latency Datacenter Network

With the emergence and popularization of high-speed network,data center can provide lower latency and higher bandwidth for upper-layer applications.The resource disaggregation architecture of data center also requires the underlying network to provide a stable low-latency data path.Traditional TCP congestion control algorithms can no longer meet the needs of data center networks.At present,congestion control algorithms in data center can be mainly classified into two categories.One is based on ECN marking and the other based on RTT.The ECN marking-based method has poor scalability and cannot be deployed on a large scale network.The RTT-based congestion control algorithm is simple and effective,and does not depend on the underlying network devices.It is suitable for various network environments.However,the existing RTT-based congestion control algorithm has two problems.First,it is unable to maintain a stable queue length and cannot tolerate large-scale traffic incasts.Second,because of the lag in control loop suffering from the increasing queue length,the queue is unstable and oscillates when the RTT-based control algorithms converge.The longer the queue length,the larger the oscillation range.We propose SQCC algorithm to solve the problems of the existing RTT-based congestion control algorithm.Compared with the previous algorithms,SQCC makes improvements in two aspects.One is to design a new error function to maintain the queue length only in a limited range growing with the increase of the number of incast flows,and will not exceed the theoretical upper limit.The second is to design two selfadjustable parameters,the rate increment and the RTT threshold.It effectively controls the range of queue oscillation when the algorithm is stable and ensures a no-empty link.With the experience on SQCC,we design a new RTT-based congestion control algorithm,ULQCC,which reduces the queue length further more by innovations in two aspects.One is to set the allowable oscillation range of the algorithm.When the queue length is outside the oscillation range allowed by the algorithm,ULQCC adopts threshold-based rate adjustment to reduce the queue length quickly.When the queue length is within the oscillation range allowed by the algorithm,ULQCC uses the PID controller method to control the queue length in a fine-grained manner as close to the threshold as possible.This approach ensures the rapid convergence of the algorithm and maintains a stable low-latency queue.The second is to set the feedback timeout.When the ACK is still not received within the feedback timeout,the sender actively reduces the rate,which effectively alleviates the queue length oscillation caused by the feedback delay.We test and verify the effectiveness of the algorithms in NS3 network simulation.Compared with previous RTT-based congestion control algorithms,there is an 80%decrease in queue length by SQCC,with a lower oscillation range by 27% to 57% in large-scale incast scenario.ULQCC further reduces the queue length.In the same scenario,ULQCC reduces the queue length by up to 87% compared with SQCC.