Abstract: To meet the performance demands and reduce resource consumption, researchers have proposed many elastic retractable algorithms and schemes. However, most of them only considered current states of the servers or applications that limit the applicability of algorithm and the performance of resource adjustment. This paper presents an application-oriented elastic retractable algorithm based on long-short term memory network and back-propagation neural network. The algorithm includes a workload prediction model, a response time prediction model and a resource adjusting model. With these models, the algorithm can predict the workload and response time of cloud computing applications and provide appropriate resource scheduling strategies. In order to improve the accuracy of workload prediction, we combine convolution operation with long-short term network for better data feature extraction and prediction. To improve the convergence speed of the model and effectively avoid the problem over-fitting, we use batch normalization in BP neural network. In the validation experiment, the mean absolute percentage error of the algorithm’s workload prediction is reduced to 3.4×10-4, and models of response time prediction and resource adjusting also perform well. In the actual operating environment, the algorithm can also provide appropriate adjustment suggestion to Docker container cloud platform. The results of experiments show that our algorithm had better performance in workload prediction and server adjusting than other models.

Key words: elastic retractable, workload prediction, container cloud, long-short term memory network