Research On Out-of-order Processing Method And Controller Design Based On RoCE Network

Remote Direct Memory Access(RDMA)allows user space applications to directly read or write to remote memory,enabling kernel bypass and memory zero-copy.Because of its high bandwidth and low latency,RDMA is widely used in data center networks.In this case,the processing of out-of-order data packets has become a problem that the data center NIC must face.However,the current mainstream data center RDMA network card is sensitive to out-of-order packets and can not cooperate with the traffic scheduling mechanism,which limits the transmission efficiency and application scope of the RDMA network card.Therefore,this thesis analyzes related technologies,and proposes an improved Selective Acknowledge RDMA over Converged Ethernet algorithm(SACK-RoCE)to implement the out-of-order detection and recovery method of RDMA.On this basis,this thesis also designs the hardware architecture of RDMA out-of-order recovery controller to enhance its reliable transmission performance.The main work of this thesis is as follows:(1)Aiming at the problems encountered in the application of RDMA protocol in the data center network,this thesis analyzes the related out-of-order recovery algorithms and designs SACK-RoCE algorithm.In this algorithm,2-bit bitmaps are used to accurately find out the outof-order packets.Meanwhile,a method for timing out the out-of-order holes is set to determine whether packets are lost.In addition,the improved SACK frame is used to help the sender to selectively retransmit lost packets.In terms of resource consumption,SACK-RoCE reduces more than 30% of the additional storage space compared with IRN and other algorithms,reduces the deployment and debugging costs of the algorithm,and expands the application scenarios of RDMA.(2)On the basis of SACK-RoCE,this thesis designs RDMA out-of-order recovery controller.The controller is divided into three parts: bitmap management,out-of-order detection and retransmission attribute management.In order to meet the requirements of data center NIC for fast processing of multiple connections,the controller sets up a 16 engines parallel mode.The combination of parallel processing mode and engine allocation module can reduce the occurrence of cache bumps and enable the controller to handle concurrent traffic.In addition,this thesis also sets the page and free-list management mechanism.In a word,the RDMA out-of-order recovery controller uses 6.8KB of storage resources to accelerate the access process to 256 K linked resources.(3)This thesis simulates the SACK-RoCE algorithm on the NS-3 simulation platform.The results show that the SACK-RoCE can still maintain 87.5% effective throughput in a lossy network environment with a packet loss rate of 1%.In real load simulation,the SACK-RoCE has12.05%～22.32% improvement in average FCT,tail FCT and average Slowdown compared with IRN.In addition,this thesis implements the hardware design of RDMA out-of-order recovery controller on Field Programmable Gate Array(FPGA),and verifies its functions.The comprehensive performance test of the hardware proves that the controller can achieve the bandwidth of 40 Gbps network card ports and support 512 concurrent links,meeting the requirements of RDMA enabled network interface cards(RNIC)deployed in the data center.