Method, system and related equipment for scheduling deep learning jobs

Abstract

The invention discloses a method, a system and related equipment for scheduling deep learning jobs. The method includes acquiring job requests for the deep learning jobs; determining target job description file templates from a plurality of preliminarily stored job description file templates according to deep learning library types and job types; determining identification of target job base mirror images from identification of a plurality of preliminarily stored job base mirror images according to the deep learning library types and the job types; generating target job description files according to the target job description file templates and the identification of the target job base mirror images; transmitting the target job description files to a container scheduler; selecting targetjob base mirror images from the preliminarily stored job base mirror images by the container scheduler according to the target job description files and creating at least one container for executing the job requests. The deep learning library types and the job types are carried in the job requests. According to the scheme, the method, the system and the related equipment have the advantage that the deep learning job scheduling compatibility rate can be increased by the aid of the method, the system and the related equipment.

Description

technical field [0001] This application relates to the field of artificial intelligence, in particular to a deep learning job scheduling method, system and related equipment. Background technique [0002] In recent years, deep learning technology has been more and more widely used in various industries. Major public cloud service providers at home and abroad have launched deep learning cloud services. This type of cloud service has become an inevitable choice for enterprises to lower the threshold of technology use and reduce the cost of software and hardware deployment. When cloud service providers provide deep learning services, they often need to consider many indicators such as cost, performance, resource utilization, reliability, scalability, maintainability, etc., and the pros and cons of the scheduling system largely determine the above indicators. . This is because the "on-demand" and "elastic" usage characteristics of cloud services need to be realized through in...

AI Technical Summary

Problems solved by technology

These complex factors make it difficult for the batch job scheduler to schedule deep learning jobs in a simple way, and users have to write some adaptation scripts that are not highly reusable

[0009] (2) Although deep learning commissioning jobs and online reasoning jobs are similar to traditional services, as an application service submitted by users, their life cycle is often relatively short, and their scheduling requirements are also different from those of web servers and databases. Typical System Services

For the service scheduler designed for system service scenarios with relatively stable quantity and life cycle, and lacking batch processing abstraction and mechanism, these special scheduling requirements are either completely impossible to achieve, or need to be assisted by complex external mechanisms

[0010] Neither of the two types of traditional schedulers can fully meet the complex and diverse scheduling requirements of multiple deep learning libraries and multiple types of deep learning jobs, which is an important obstacle to providing deep learning services in the public cloud

Simply using the original batch job scheduler or service scheduler not only cannot realize the scheduling strategy proprietary to deep learning, thereby reducing user experience and increasing the complexity of operation and maintenance; but also potentially affects the utilization of hardware resources and improves Operating costs of public cloud

Images