Research On Congestion Control Schemes For Mix-flows In Data Center

In order to meet people's increasing demand for high-performance computing,the scale of data centers is growing rapidly.A single data center can accommodate thousands of servers that perform efficient data processing and high-speed data transmission.With the increase of data center services,the traffic volume injected into the data center network is increasing drastically.As a result,congestion occurred due to the inability to process excessive data on the bottleneck link,causing a slump in throughput.In particular,online services impose restrictions on the completion time of data transmission in data centers.An excessively long response time will greatly influence the user experience,thereby reducing the revenue of data center operators.Modern data center applications require high throughput(40Gbps)and ultra-low latency(less than 10 us per hop)for the network,and low CPU overhead.However,the standard TCP/IP stack cannot meet the high bandwidth and low latency requirements of the data center.In this thesis we study the data center congestion monitoring technology,and analyze the queue-based and delay-based congestion measurement standards and their models in detail.Then we study and summarize recent data center congestion control researches According to different goals,we categorize congestion control schemes into two main categories-schemes to optimize flow completion time and schemes to solve the incast problem.We introduce and analyze their advantages and disadvantages respectively.The queuing delay is an important part of the flow completion time.But the queue-based congestion control schemes are not sensitive to delay.They cannot effectively monitor or limit the packet queuing delay,resulting in unexpected flow completion time.To solve this problem,this thesis proposes a packet-queuing-delay-based congestion control strategy-QDCN(Queue Delay based Congestion Notification),which directly monitors and limits the packet queuing delay and its changes,and responses quickly to the dynamic network states.In addition,combining the calculation of packet queuing delay with the dequeue marking scheme,the network congestion signal is accelerated to the senders.Simulation results show that the congestion control strategy can effectively improve the flow completion time by 15% under certain conditions.In credit-scheduled schemes,a receiver keeps generating and sending credit packets until it receives a stop signal from the sender,which causes a large amount of credit packets wastes and decreases the network link utilization.Aiming at this problem,this thesis proposes a congestion control strategy-ECS(End-to-end Credit Scheduled Congestion Control),which is based on active credit packets termination.By monitoring the transmission status of the data traffic at the senders,this thesis establishes a prediction model to forecast the data flow deadline and calculate the number of required credit packets for receivers.The end-host will generate and send abundant credit packets and actively stop sending credit flows without specific stop signals.In this way,the waste of credit packets is reduced as much as possible while ensuring the network requirements.Simulation results show that this congestion control strategy can effectively reduce the waste of credit packets under high-load conditions and improve the completion time of small flows by about 25%.