Dual model method and system for sparse resource prediction

Abstract

The application relates to a double-model method and system for sparse resource prediction, and belongs to the technical field of cloud computing and artificial intelligence. The method comprises the following steps: collecting resource usage data from a container orchestration platform and extracting a 23-dimensional feature vector; training a CPU classifier to determine whether a container is active; training a CPU regressor to predict CPU usage rate only on active samples; and training a memory regressor to directly predict memory usage rate. The CPU uses cascade prediction of classification and regression, and the memory uses direct regression prediction. The application explicitly processes the sparsity of resource data through a double-model architecture, and processes the heterogeneity through separate modeling of the CPU and the memory. The application has the following advantages: through a two-stage architecture of a classifier and a regressor, the sparsity problem in the resource usage data is explicitly processed. The classifier focuses on detecting whether a container is in an active state, and the regressor is trained only on active samples and focuses on predicting the resource usage rate in the active period. The application can be widely applied to scenarios such as Kubernetes elastic scaling, resource optimization and cost saving.

Images