Computing power scheduling method and device based on Kubernetes

Abstract

The invention discloses a computing power scheduling method and device based on Kubernetes, and relates to the field of artificial intelligence. The method comprises the following steps: analyzing computing resources required by a machine learning current training task to decompose the type and quantity of a target accelerator card, screening out a corresponding machine node set according to a preset strategy in combination with computing power global data of all machine nodes in a Kubernetes cluster, and locking the computing resources of the machine nodes used by the current training task according to the corresponding machine node set, and establishing an inter-node mutual trust interconnection network on the machine nodes used by the training task, running the machine nodes used by thetraining task, and learning a training program to complete training. The device comprises an initial module, an analysis module, a screening module, an establishment module and a training module. According to the invention, intelligent distribution of resources is achieved, manual adjustment is not needed, and the efficiency is greatly improved.

Description

technical field [0001] The present application relates to the field of artificial intelligence, in particular to a Kubernetes-based computing power scheduling method and device. Background technique [0002] In recent years, with the great success of deep neural networks in the field of artificial intelligence, many open source machine learning platforms have emerged in the industry. How to implement distributed task scheduling and improve resource utilization has become the main research content of distributed cloud deployment, edge device deployment, and end device deployment of deep neural networks. [0003] Deep learning has higher and higher requirements for computing power, including a sharp increase in parameter scale, and longer iteration time, from the previous hour level to the day level, or even the month level. For example, video services have more parameters and more complex models. In the face of terabytes of training data, it often takes a long time to train ...

AI Technical Summary

Problems solved by technology

Although the scheduling strategy of Kubernetes can meet the above requirements, the resource screening mechanism of its tags and selectors is only a simple matching of resources. It needs to manually specify the properties of machine nodes and accelerator cards, and cannot intelligently allocate resources.

Images