The simulation task in the communication simulation system is a CPU-intensive task.When there are many users that use the simulation system at the same time and the amount of simulation performed concurrently is large,the hardware requirements are high.In order to meet the requirements of simulation tasks for computing resources,there are generally two methods of vertical expansion and horizontal expansion.Vertical scaling increases the processing power of a single server.However,the increased computing power of this method is limited by the development level of computer hardware and has a low cost performance.Horizontal scaling is the use of multiple servers connected to form a cluster to provide computing services outward,which has a higher cost performance.Cluster-based next-generation simulation systems can be easily scaled out to meet the demand for computing power.This paper first gives the detailed design of the cluster system architecture design,function modules,and information interaction timing diagrams.It describes how to build a cluster system for communication simulation that can provide high computing power,and then the core content of the cluster system:load forecasting and task scheduling.Methods to start research.The horizontal expansion of the system can increase the computing power and reduce the hardware cost of the entire system.However,there has also been a problem of unbalanced use of node resources in the cluster.The load balancing problem of clusters directly affects the utilization of resources and the responsiveness of services.Therefore,using an effective task scheduling strategy to ensure cluster load balancing is of great significance.The implementation of load balancing can not be separated from the load monitoring.The load monitoring of the system has multiple monitoring indicators.In order to effectively apply the load conditions to the load balancing scheduling algorithm,multiple monitoring indicators need to be combined into a single value.At the same time,how to design an efficient and reasonable load monitoring cycle will also affect the monitoring of its own node’s resource consumption and the amount of data transmitted between nodes,and ultimately related to the external service capabilities of the entire cluster system.This article gives an effective method for the above monitoring problems.The general load balancing scheduling strategy is to perform the request scheduling allocation according to the real-time load condition of the nodes or migrate the tasks among the nodes.The communication simulation task in this paper has a long running time,and the load conditions generated by the system during different stages of operation are also different.This makes it hard to ensure the load balancing of nodes based on the balanced scheduling when the real-time load is requested;the communication simulation system only provides simulations.The features,their functional unity and non-disruptiveness make it impossible to use service transfer methods to ensure load balancing.For this reason,a scheduling method based on running time prediction and load forecasting is designed.Using the regression model in machine learning,the system configuration parameters are used as feature vectors to predict the running time and load.In the task scheduling,based on the prediction data,the scheduling strategy of the load window superposition is designed.Even if the load generated by the task fluctuates with the running time,the load balance of the system can be ensured.When the cluster is expanded,the addition of heterogeneous nodes makes the system need to generate a large number of seed sets for the new nodes to generate different prediction formulas,which increases the complexity of cluster expansion.In this paper,the method of normalization of heterogeneous nodes is proposed to make predictions.Formulas only need to be generated based on a specific node. |