Design And Implementation Of PCIe Protocol Simulation Platform

With the rapid development of network communication technology,how to break through the limitations of hardware communication protocols and expand communication in a wider range of fields has become a popular research direction.The CCIX(Cache Co-herent Interconnect for Accelerators)specification protocol provides high-performance,low-latency communication interconnection for the next generation of cloud computing,artificial intelligence,big data,databases and other data center infrastructure.The CCIX conformance protocol can be delivered over the PCIe link with little modification.PCIe,as a high-bandwidth,point-to-point serial interconnection bus protocol used in various com-puting and communication platforms,has a very promising application prospect in cloud computing,data centers,and IT product lines.This thesis proposes to build PCIe under-lying communication software simulation analysis platform to avoid the disadvantages of traditional hardware configuration rigidity,long development cycle,and large workload by means of software simulation,and on this basis,conduct an in-depth analysis of the communication delay performance.Based on the research of CRC and FEC coding technology to ensure data integrity in network communication,this thesis carries out PCIe point-to-point network protocol simulation.First,the software reproduces the functions of each layer of the protocol.Based on the realization of multi-process communication in the form of shared memory of the ring queue,high-speed data transfer is carried out between the various layers of the protocol and between the layers and the data bus,which implements the End-to-end full-duplex synchronous communication.Secondly,considering the inconsistency between the hardware operating clock and the simulation clock,this thesis proposes a simulation clock advancement algorithm for flow control and bus transmission to analyze and process the delay more reasonably and accurately.Finally,based on the realization of the simulation platform,a human-computer interaction entry is provided,and intuitive parameter input and result output are performed through technologies such as web pages,data blocks,and back-end servers.On the basis of completing the PCIe point-to-point protocol simulation design,this thesis makes the PCIe protocol multi-point to multi-point simulation.Distributed machine learning,as the frontier of artificial intelligence in recent years,can carry out flexible and efficient data training through methods such as data partitioning or model partitioning when the amount of calculation is too large,the training data is too large,and the model scale is too large.And in the ordinary neural network structure,as the depth increases,it faces gradient explosion/dissipation.Res Net can effectively limit the gradient degrada-tion through Batch Normalization and convolution operations.Based on the theoretical analysis of Res Net-50,this paper adopts the data parallel optimization strategy under the typical AI cluster to establish a delay calculation model.Based on the effects of param-eters such as transmission rate,bandwidth,and bit error rate on AI performance under the Full-Mesh and Switch topologies,simulation analysis and theoretical verification are carried out.