Research On Intelligent Congestion Control Scheme

In recent years,the increased demand for computer networks has resulted in networks carrying a large number of applications with different characteristics,a dramatic increase in network traffic,and frequent network congestion,which poses a great challenge to congestion control techniques.Traditional network congestion control schemes that rely on mathematical models cannot adapt to complex dynamic network environments,and algorithms applicable to specific scenarios become completely inapplicable when network conditions change,and may even negatively affect network performance.The congestion control scheme that uses packet loss or network delay as the congestion signal cannot respond in time because the congestion signal has a certain lag,and the feedback information of the congestion signal is not accurate enough,resulting in the control effect cannot be guaranteed.Congestion control algorithms continue to evolve,multiple congestion control algorithms are constantly proposed,and the coexistence of multiple versions of congestion control algorithms makes it extremely unfair for different algorithms to compete for network bandwidth.In this dissertation,we address the above problems and study the network congestion control scheme from the different perspectives of network devices,senders and receivers,exploring a model-free,adaptive,and intelligent congestion control scheme.Machine learning is a data-driven approach and does not rely on mathematical modeling,so this dissertation proposes intelligent network congestion control schemes using machine learning techniques.In addition,this dissertation uses new congestion signals instead of the conventional signals and compute more accurate congestion windows,which achieves the goal of reducing congestion,improving throughput,and reducing network latency.For flows using different congestion control algorithms,this dissertation adopts flow isolation methods and different scheduling strategies to enable different flows to compete fairly for network bandwidth and to mitigate the negative impact of congestion.According to the deployment locations of the congestion control schemes,the main contents and contributions of this dissertation can be divided into two aspects,terminal congestion control and network device congestion control,specifically summarized as follows.1.Queue management mechanism for open virtual switches in multi-tenant data center networks in terms of terminal congestion controlFor multi-tenant data center networks,network operators often cannot control the settings of tenants’ operating systems and congestion algorithms,resulting in uncontrollable interactions between tenants.In the virtual switch on the physical host,the network operator can monitor the tenant traffic and learn an optimized queue management policy based on the reinforcement learning that can automatically adapt to different network environments,regulate the behavior of the tenants,reduce the negative impacts between tenants,and provide efficient network services.Experiments show that the mechanism can protect the interests of individual tenants and improve fairness and overall network performance.2.INT-based TCP congestion window modulator in terms of terminal congestion controlDuring connection establishment,a suitable initial rate is calculated to optimize the short flow completion time.After the connection is established,the in-band telemetry technique is used to attain the queuing information of the network devices.The INT metadata is used as the new congestion signal,which is more accurate than the packet loss and latency signals.The INT data contains specific information about the degree of congestion and the location of the congestion.The receiver can calculate the accurate congestion window from the INT data and make a timelier response to the network congestion.This scheme strives to avoid congestion before it happens,instead of taking remedial measures after the congestion occurs.Experimental results show that this scheme can improve the throughput of the elephant stream and reduce the stream completion time of the mouse stream.3.Congestion algorithm-aware switch queue management and scheduling scheme in terms of network device congestion controlFor the scenarios where multiple congestion algorithms coexist,there are unfairness problems in network bandwidth allocation,which is caused by sharing internal network queues.Utilizing the programmable capability of programmable switches,network flows are classified by a decision tree algorithm,and then assigned to different queues for traffic isolation.For different types of queues,different scheduling strategies are used.This scheme improves the network throughput and the fairness between different flows.4.Active queue management scheme based on reinforcement learning in terms of network device congestion controlAiming at the problem that traditional congestion control algorithms cannot automatically adapt to dynamic environment because they rely on rules or mathematical models,the automatic learning capability of reinforcement learning is utilized to design an adaptive network congestion control scheme.The scheme is deployed on network device,because network device has the ability to sense the global traffic and have better control over the whole network than terminal devices.The network device continuously optimizes its queue management strategies to prompt the terminal to adjust the sending rate to adapt to the changes in network conditions,so as to achieve the purpose of improving the performance of the whole network.