Research On Load Balancing Mechanism Based On Intelligent Measurement Of Programmable Data Plane

As the scale of the data center network continues to expand,traditional network management solutions can hardly adapt to the current huge and complex network environment,and the network traffic management problems of data centers are increasingly prominent.The shortcomings of previous network monitoring solutions,such as low scalability,insufficient measurement accuracy,and the scheduling granularity of load balancing policy are too coarse,the decision delay is too large,and the load balancing effect is poor,are increasingly affecting the operational performance of data center network.The idea of separating the control plane and data plane in software-defined networking provides a new solution for network management solutions,and the emergence of P4 further releases the programming capability of the data plane,through which the tasks previously centralized in the controller can be partially decentralized to the local switches,making full use of the resources in the network while reducing the latency of decisions.To this end,the main work of the thesis is as follows.First,based on P4 and a programmable data plane,this thesis proposes a fine-grained,scalable,and high-coverage network measurement scheme Full Sight for local switch and controller collaboration,which provides a low-overhead,high-coverage measurement scheme for intra-switch data and inter-switch data,while Full Sight designs a packet-based behavior algorithm IMM that intelligently adjusts the measurement frequency based on the behavior of data packets.Based on the change in data behavior in the network,IMM automatically adjusts the network measurement frequency to reduce the overhead caused by network measurement.In addition,for the problem of limited memory in the programmable data plane,a rotating storage algorithm RM is designed to enable Full Sight to continuously record the data in the network with limited memory.Finally,an error weakening algorithm is designed for the clock asynchronization problem and the hash collision problem existing on the programmable data plane.Simulation experimental results show that Full Sight,Netsight,and Netseer schemes can all achieve full coverage measurement of packet loss data,while reducing the bandwidth overhead by two orders of magnitude compared to Netsight,with Full Sight bringing only 0.1%bandwidth overhead in the measurement process,while,using the RM storage algorithm Full Sight requires less than 0.001% of the switch memory in the measurement,demonstrating Full Sight's very good scalability.Secondly,this thesis proposes a distributed load balancing policy,Roll,based on Full Sight's intelligent measurement,to address the problems of existing load balancing policies,such as large decision latency and not applicable to asymmetric topology,etc.Roll is a distributed load balancing policy with packets and flowlet as the scheduling granularity,based on local queue information and probabilistic forwarding algorithm for traffic distribution,and can be applied to any topology.Based on Full Sight,Roll implements a module to collect and return local traffic information,Roll proposes a congestion value measurement algorithm based on local queue information according to the existing traffic characteristics of the data center,Roll uses a weighted probabilistic forwarding algorithm related to the congestion value in traffic forwarding to avoid fast congestion in the path.Roll overcomes the problem of difficult access to queue information on the switch and the single choice of routing ports.The simulation results show that compared with the coarse-grained flow scheduling policy ECMP,the flowlet-grained scheduling policy CONGA,and the packet-grained scheduling policy DRILL,Roll-pkt has different degrees of improvement,and the average FCT of the flow is only 50% of that of ECMP,68% of that of CONGA,and 89% of that of DRILL respectively.In terms of decision latency,Roll,a load balancing policy based on local decision making,has at least nearly double the decision latency compared to CONGA,while compared to HULA,Roll's decision latency is reduced by about 74.3%-92.8%;and in terms of switch queue fluctuation,Roll has an order of magnitude improvement compared to ECMP,and compared to the DRILL scheme,Roll-pkt has a 6.2% reduction in queue occupancy variance.Finally,this thesis proposes a method to fit the switch local queue information with a leastsquares estimation based on the existing traffic characteristics of the data center,which can be used to measure the path congestion corresponding to the port.And the fitted function is used as the congestion measurement module of the load balancing policy Roll,which is deployed to the programmable data plane.Experimental results show that Roll with this congestion measurement module can further reduce the FCT of Roll-pkt under different traffic loads,and the FCT of Roll-poly2 under Web-Search load is about 93.8% of that of Roll-pkt,while the FCT of Roll-poly3 under Data-Mining load is about 92.2% of Roll-pkt.