Design And Implementation Of A Simulation Tool For Intelligent Network Computing Application

At present,machine learning is widely used in the field of network research,but the network environment used by researchers is inconsistent,which makes it difficult to reproduce the algorithm results,which hinders the development of machine learning network research.At present,the mainstream tool ns3-gym uses ns-3 as a unified network simulation platform,and uses ZMQ sockets to achieve data interaction between ns-3 and OpenAI Gym.However,the ZMQ socket method makes its data transmission rate slower,and it can not meet the requirements well in the scenario of large data volume and high transmission rate.Also,the OpenAI Gym used by ns3-gym only provides an interface for reinforcement learning and is not compatible with deep learning.Therefore,there is an urgent need to design and implement a universal and high-speed machine learning network research algorithm tool.The general tool ns3-ai designed and implemented in this thesis is divided into two aspects:C++side and Python side.The data interaction between ns-3 and the general AI framework is realized through shared memory.The shared memory pool includes the main control block,control block and memory block.Among them,the memory block realizes the synchronous control of the shared memory through the version information stored in the memory block,so as to ensure the normal reading and writing of data in the memory block.In addition,ns3-ai also provides a high-level interface for deep learning and reinforcement learning.The template class specifies a standardized data format,which is convenient to use and guarantees flexibility.In this thesis,the actual application of ns3-ai is also put forward.Three examples of cognitive radio,large data volume interaction and channel quality prediction scenarios are proposed,and ns3-ai is tested from three aspects of data transmission rate,correctness and versatility.The performance is compared with the current mainstream tool ns3-gym.Specifically,the Cognitive Radio example uses the scenario and reinforcement learning algorithm provided in the benchmarking tool ns3-gym,and illustrates the high data transmission rate of ns3-ai through comparison with ns3-gym.The experimental results prove that the training results of ns3-ai and ns3-gym have the same trend,indicating that the training results of ns3-ai are correct and have the same functions as ns3-gym.In addition,the data transfer rate of ns3-ai from the C++side to the Python side is 6 times higher than that of ns3-gym,and the data transfer rate from the Python side to the C++side is 70 to 120 times higher than that of ns3-gym.The large data volume interaction scenario example proves that the running time of ns3-ai and ns3gym has a linear relationship with the set simulation time.When the simulation time is set to 200 minutes,the required time for ns3-ai is 2 minutes and 48 seconds,and the required time for ns3-gym is 19 minutes and 45 seconds.This shows that when the transmission data packet is large and the simulation time is very long,ns3-ai It always has 10 times higher performance than ns3-gym.The channel quality prediction example uses the LSTM algorithm and the FNN algorithm to perform channel quality prediction,so as to achieve the performance evaluation of the two algorithms,in order to illustrate the generality of ns3-ai.The experimental results show that the accuracy of LSTM is 50%higher than that of FNN,and when the user's moving speed is increased to 70m/s,FNN can no longer provide prediction,while LSTM still has a prediction accuracy of more than 50%.This experiment proves that ns3-ai has a good performance in the application of deep learning algorithm evaluation,and is a versatile tool that takes into account deep learning and reinforcement learning.In summary,ns3-ai is both correct,versatile,efficient and stable.At present,ns3-ai has passed the certification of ns-3 official organization,and it will be launched on the ns-3 APP Store in January 2020.