Research On Load Balancing Method Based On Load And Load Change Rate

With the surge in the number of network users in China,the pressure on the network is huge,and the popularity of network applications such as e-commerce and live broadcasting has pushed the pressure on the network server to the limit.Network application suppliers and domestic researchers also continue to put forward solutions.At first,the problem of high concurrency will be solved by improving the performance of server hardware,but the high maintenance cost and maintenance cost make this solution insufficient to meet the actual application requirements,so the load balancing technology of server cluster is proposed.This thesis mainly focuses on the scheduling algorithm of load balancing to solve the problem of concurrent access to large amounts of data.This paper presents a method based on LLCR（load and load change rate）.Firstly,aiming at the number of server connections,a prediction model of server connections based on two-way long-term and short-term memory network is proposed,taking the number of connections at the next time and the current time as the index to evaluate the server performance.The data set used for the training of the model is the cluster recently released by Alibaba cloud in cluster-Trace-v2018,compared with the prediction algorithm of cyclic neural network and long-term and short-term memory network,the mean square error is reduced by about 12.5%,the mean absolute error is reduced by about 18%,and the determination coefficient is increased by about 20%.Then,for the selection of server performance indicators,add two indicators: CPU utilization change rate and memory utilization change rate,monitor the performance through the average load of the server,and test the optimal server weight update cycle.The experiment shows that it can better target the concurrency of large amount of data.By setting up a server cluster test environment,the number of the concurrency is set to 100-7500,Finally,the performances of polling scheduling algorithm,weighted minimum connection number algorithm and destination address hash scheduling algorithm built in LVS（Linux Virtual Server）are compared in three aspects: overall response time,throughput and failure times.The experiments show that the above three aspects are improved by about 10.3%,9.8% and 7.5%respectively.