Abstract: Aiming at the problem of load skew between nodes in distributed messaging systems under cloud services,a dynamic load balancing strategy based on the role of replica is proposed and the algorithm is applied to Apache Kafka,the distributed streaming platform.Because the function of the messaging system is to read,write and store messages,the algorithm used CPU,disk and bytes in/out as the main load factors of nodes,and proposed the corresponding Leadership Movement strategy and Replica Movement strategy according to different load types.The feasibility of the algorithm is demonstrated from the perspectives of time cost,space cost and service availability,and the influence of parameters involved in the algorithm on the execution of the algorithm was discussed.Experiment results show that,the algorithm can achieve that the resource usage of each node in the cluster is not greater than the specified threshold.Compared with the default system,the standard deviation of cluster CPU occupancy rate decreases by 72.1%,the standard deviation of disk occupancy rate decreases by 86.1%,the standard deviation of bytes in rate decreases by 79.2%,and the standard deviation of bytes out rate decreases by 63.9%.The optimization effect is remarkable.

Key words: Apache Kafka, Cloud service, Distributed messaging system, Load balancing, Multi-replica mechanism