Resource management method, system and device and medium

Abstract

The invention discloses a resource management method, which comprises the following steps: in response to a received first instruction of creating a container by a user, judging whether resources in afirst resource pool meet resources required by a to-be-created container or not; in response to the fact that the resources in the first resource pool meet the required resources, creating a container by utilizing the resources in the first resource pool; in response to a received second instruction of destroying the container by the user, adding the resources released after destroying the container into a second resource pool; and preferentially returning the resources corresponding to the first type of nodes of which the resource utilization rate is greater than the threshold value in the second resource pool to the first resource pool according to a preset proportion every other preset time period. The invention further discloses a system, a computer device and a readable storage medium. According to the scheme provided by the invention, certain nodes can reserve a large amount of dominable resources, the waiting probability of containers with large resource requirements is reduced, and the overall resource utilization rate of the cluster is improved.

Description

technical field [0001] The present invention relates to the field of cloud computing, in particular to a resource management method, system, equipment and storage medium. Background technique [0002] In the past two years, the field of artificial intelligence has shown the following trends. First, the scale of deep learning, reinforcement learning and other models has become larger and larger, especially model training needs to consume huge CPU and GPU resources, so model training generally requires a server cluster support; The second is that model development and training are becoming more and more streamlined. Different R&D personnel participate in different links, and they also need their own opposing development and training environments. Usually, fewer resources are used in the development of the model, and more resources are used in the model training process. Therefore, it is necessary to provide independent development and training environments with different reso...

AI Technical Summary

Problems solved by technology

Although Kubernetes container orchestration brings scheduling automation and reduces manual intervention, there is still a pain point that has not been well resolved in the face of the coexistence of model development and model training.

The current Kubernetes scheduling strategy is relatively simple or mainly considers the relative load balance of each node in the cluster, resulting in a more balanced distribution of containers on each node, and the resource usage and remaining amount of each node are basically the same

When the R&D personnel need to train the model, if a single container requires a large amount of resources, the remaining resources of each node in the cluster will not meet the resource requirements for scheduling, resulting in scheduling failure, users cannot use platform services, and also cause a certain amount of waste

The limit case is that the resource utilization rate of each node in the cluster is not high, but none of the nodes can meet the resource requirements required by the training environment, and the user can only train the model when the resources are sufficient

Images