Performance Optimization Of PFC Protocol In RDMA Lossless Networks

With the explosive growth of application data,data centers have increasingly high requirements for network transmission throughput and latency performance.More and more commercial data centers are gradually adopting Remote Direct Memory Access(RDMA)technology to achieve high throughput and low-latency data transmission goals.This technology relies on a lossless network that employs a priority-based flow control protocol(PFC protocol).The protocol sends information frames to the upstream port before the downstream input queue overflows,thereby suppressing its data transmission.However,while the PFC protocol suppresses upstream data transmission,it can also cause severe issues such as deadlock,head-of-line blocking,and pause frame flooding,posing significant risks to network data transmission.In order to mitigate these risks,this thesis conducts in-depth research on these problems and proposes corresponding solutions.The main contributions are as follows:1.In order to address the issues of deadlock and head-of-line blocking,this thesis analyzes the root causes of the problems and proposes improvements to the upstreamdownstream speed coordination mechanism.The thesis also introduces a multi-priority decoupling design.Specifically,the upstream-downstream speed coordination mechanism monitors the length of the downstream input queue using a feedback packet generator and periodically samples the queue length within a fixed time scale.Based on this,the speed difference between the upstream and downstream is calculated to adjust the sending speed of the upstream output queue.This mechanism ensures that the input queue does not overflow and avoids the complete stoppage of data transmission from the upstream output queue,thereby eliminating the wait condition and preventing deadlock.Additionally,this thesis proposes a multi-priority decoupling design.The congested flows’ destination addresses are calculated by the feedback packet generator in the downstream,and the upstream node utilizes the destination address information of the packets to differentiate between victim flows and congested flows.For victim flows,they are temporarily inserted into selected idle queues,achieving decoupling from congested flows and reducing the flow completion time of victim flows.Extensive experimental simulations demonstrate that the enhanced PFC mechanism can prevent the occurrence of deadlock and significantly mitigate the effects of head-of-line blocking,reducing the average flow completion time of victim flows by over 40%.2.To avoid pause frame flooding,this thesis analyzes the root cause,which is the inability to timely insert the head-of-queue data of each input queue into the congested output queue during congestion.This thesis proposes an incremental deployment scheme for memory on the switching nodes,allowing temporary storage of congested head-of-queue packets and their subsequent release after congestion relief.This approach effectively mitigates the head-of-line blocking issue,thus avoiding the propagation of congestion-induced pause frame storms.To minimize the usage of storage devices while covering all data transmission paths,this thesis establishes a mathematical model based on the network’s topology structure and the path preference of data flows.It proposes a greedy-based approximate deployment scheme.Experimental simulations demonstrate that by incrementally deploying memory,pause frames can be effectively contained within smaller network areas,reducing the flow completion time of victim flows and alleviating the occurrence of deadlocks.