| In these years, with the development of information technology, increasing popularity ofthe network, explosion internet and the World Wide Web have obviously adds a lot of onlineinformation and possible service for user. Some primary zone business flow has been farbeyond the single Web server’s affordability, such as enterprise network, campus network andwide area network. The growing information and service depends on an internet infrastructurewhich can resist dramatic pressure, and one can provide a large number of Web requestingWeb server system. Therefore, considering the system scalability and cost efficiency, adistributed design Web server system (Web cluster) instead of a mainframe machine has beenappeared. The mainframe machine is the high performance Web server. The most popularWeb site using distributed/parallel structure in order to reduce the load of a single server.How to reasonable and effective scheduling distribution network business flow has beenworth consideration. Reasonable and effective scheduling distribution network load can helpus to solve some server equipment in busy and other server equipment in the idle question.For the Web server cluster, it is designed in order to solve the server performancebottleneck problem, to provide users with high quality Web service. The cluster system’s Webrequesting is evenly distributed to backend server by front-end distribution server, it realizesthe server performance maximize. For the front-end scheduling server, how to effectivelydistribute load is one of the main factors will be influence improving the Web server clustersystem performance.In the Web server cluster system, the load balance technology is the eternal coretechnology. Load balancing technology provides a good strategy to resolve system usability,it can very good shielding failure server, and realize the user service transparent switch. At thesame time, the use of load balance technology also can be reduce each server real-time load ina certain extent, improve web server cluster system, improve the performance of the servercluster throughput and reduce the user requests service average waiting time.This paper describes the related concepts and characteristics of the load balancetechnology, load balance’s main goal is through the balanced task to redistribute work loadand optimize resource utilization and the response time of the task to minimize interprocess communication overhead. Through the balance each processor nodes work load as the goal ofthe load balance to improve the parallel computing performance. For load balancing strategyclassification, according to the processor nodes binding request task time, load balancingstrategy is divided into static load balancing algorithm and dynamic load balancing algorithm.The static load balancing algorithm and dynamic load balancing algorithm was compared andanalyzed from the system performance, response time, the algorithm complexity anddevelopment cost, and so on. The Linux Virtual Server cluster (Linux Virtual Server, short forLVS) is analyzed. Because it is more popular on the current high load Web Server clustersapplied research, and research and analysis the LVS Web Server cluster load balancingalgorithm.As the subject of research, we first make a thorough understanding and analysis researchto load balance technology, and then introduce an open source project is introduced. It is moremature open source software projects LVS (Linux Virture Server), which builds in the Linuxsystem. The open source project configurate the weighted least connection schedulingalgorithm and make the weighted least connection scheduling algorithm into a study. In theweighted least connection scheduling algorithm, each server node corresponding weightsshows the server node processing ability, weights W C/R, which C denotes server nodeperformance, R denotes the current task connection number. The algorithm based on theserver connection request quantity and server node performance indicators as parameters fortask allocation. When increasing the number of concurrent connection access request, loadbalancer also increase the processing tasks, big task throughput will lead to load balanceroverload, load balancer overload further lead to load balancer appears system bottlenecks. Inthis paper, in view of the weighted least connection scheduling algorithm deficiency, we putforward the improvement measures, and configurate the improvement method to the Webcluster system. Through test othe cluster system’s Bytes Recv Rate (cluster system throughputrate) and TTFB Avg (cluster system’s average response time for client) value to analysis theinfluence of the improved algorithm to cluster system performance.Method one: this paper puts forward a new weight value distribution schedulingalgorithm based on the Linux virtual service (LVS) cluster system, due to the weighted leastconnection scheduling algorithm only consider the current connect to the server the number ofnodes and put it as a parameter, the current server connection number can not be a goodindication of the server load condition, because of different request task content on servermemory and CPU utilization is not the same. In order to better describe the server load condition, we use utilization rate of CPU and memory free rate two parameters to describe theserver nodes load condition. In the load distribution on the realization, the load balancerperiodically get the backend server’s CPU idle rate and memory free rate to calculate thebackend server’s weight. According to the right value modifying conditions, decide whetheror not to write weight into the IPVS scheduling strategy. Finally, according to the eachbackend server’s weight to distribute the client request tasks. This new weights distributionscheduling algorithm optimizes the load balancing problem of the weighted least connectionscheduling algorithm in the LVS cluster. In a certain extent, it improves the LVS clustersystem performance.Method two: we put forward the dynamic adaptive feedback load balancing strategy.Because the new weight value distribution scheduling algorithm is by the load balancerperiodically collect backend each server node’s load information to calculate the weights ofthe server, this method increases the cluster system of communication overhead, but alsoincrease the load balancer burden. Dynamic adaptive feedback load balancing strategy usesserver performance value and server nodes of the dynamic load value as evaluation servernodes load capacity index. Each backend server node adaptive collect their load information,then sents the load information to the load balancer to calculate the server’s weight value, inorder to avoid load balancer and single server nodes instantaneous overload, we alsointroduce the server node load redundancy value parameter. To avoid the load balanceroverload leads to system bottlenecks, and reduce the load balancing algorithm complexity,effectively solve the cluster system internal load balancing problem between each servers.Method three: we put forward an innovate dynamic load balancing algorithm based ontask classification, for a distributed system, many nodes and node between dynamic taskallocation and scheduling condition is the key element of affect the system performance.Therefore, considering the user request task type differences, and user demand for resourcestypes and different levels of real time demand, the user request task types are classify intoreal-time I/O consumption task and real-time CPU consumption task two types. Loadbalancer according to the user request task type to dynamic executing scheduling task, thisscheduling algorithm called the dynamic load balancing algorithm based on task classification.This algorithm is a very good solution to solve the problem of the system low throughput.Finally, in order to optimization the cluster system performance. We make the loadbalancing strategy for performance analysis and evaluation. We mainly tested the Bytes RecvRate (cluster system throughput rate) and TTFB Avg (cluster system average response timefor client) value in the weighted least connection scheduling algorithm, a new weight value distribution scheduling algorithm and dynamic adaptive feedback load balancing strategy toanalyze the performance of the cluster system.In this paper, analysis and research the weighted least connection scheduling algorithmin the LVS Web cluster system application environment of dense load. According to theweighted least connection scheduling algorithm’s insufficient, we put forward theimprovement of the Web server cluster load balancing strategy based on the weighted leastconnection scheduling algorithm, so as to realize the reasonable scheduling server resources,make the server cluster keep load balancing, and give full play to the cluster server ability toobtain the optimization response characteristics, and promote the whole Web server clustersystem performance. |
PDF Full Text Request