A Learning-based Congestion Control Algorithm For The Internet

Learning-based congestion control algorithms gains more and more attraction,be-cause of its potential adaptability to different network environments.At the same time,the existing learning-based congestion control algorithms suffer from a series of prac-tical problems for deployment.First,we found that the performance of the existing learning-based algorithms in the evaluation metrics of congestion control,such as fair-ness,convergence and friendliness,is worse than that of the traditional congestion con-trol algorithms.Moreover,these algorithms have worse maintainability and higher overhead than the traditional congestion control algorithms.These problems make it impossible for the existing learning-based congestion control algorithms to be deployed in production environments.However,the traditional Internet congestion control algo-rithms in convergence,fairness,overhead,maintainability,and other evaluation met-rics are superior to the existing learning-based Internet congestion control algorithms,but because of the fixed control logic,these traditional congestion control algorithms cannot adapt to different network environments.Inspired by this,we designed Nunchucks,a new learning-based congestion con-trol algorithm.We decouple the design of the congestion control algorithm into the design of the parameterized congestion control class and the design of the learning al-gorithm by combining the expertise of human experts with the learning capabilities of machine learning algorithms.The parameterized congestion control class is designed by human experts and adopts the idea similar to the traditional congestion control al-gorithm.The difference is that we generalize the corresponding key parameters in the congestion control class for avoiding the disadvantages of the fixed control logic of the traditional congestion control algorithm.By elaborately designing the control logic of the congestion control class and carefully selecting the corresponding key parameters,we can analyze the convergence,fairness and other evaluation metrics of the conges-tion control class in theory.By considering the influence of the factors other than the congestion control algorithm on the performance objective in the Internet,we designed a learning algorithm based on Bayesian optimization algorithm.It makes Nunchucks have the ability to adapt to different network environments by automatically adjusting the key parameters of the congestion control class.Trace-driven emulations and real-production experiments show that compared with the existing learning based Internet congestion control algorithms,Nunchucks has lower overhead,better convergence and fairness,acceptable friendliness,and better adapt-ability to the unseen network environments.Meanwhile,Nunchucks significantly out-performs the traditional Internet congestion control algorithms in production environ-ments.