GPU resource-oriented task scheduling method, device and system

Abstract

The embodiment of the invention discloses a GPU resource-oriented task scheduling method. . The method comprises the following steps of when a task (such as a first target task) needs to be executed by using a video memory resource in a server cluster, calculating the size of a video memory resource required by the first target task according to the attribute information of the first target task included in the first task request and a task resource use configuration table, and then, according to the available resource state table of the server cluster, determining the GPU meeting the video memory resource size required by the first target task, so that the task execution module executes the first target task by utilizing the target GPU. According to the scheme, when the task request is received, video memory resource estimation of the task is carried out, the available GPU is selected with the video memory resource as the minimum unit to complete the task, and the granularity of resource management and task scheduling is refined. If the video memory resources in one GPU meet the task requirements, the purpose that one GPU executes multiple tasks in parallel can be achieved, and the utilization rate of the video memory resources in the GPU is increased.

Description

technical field [0001] The present application relates to the field of computers, in particular to a GPU resource-oriented task scheduling method, device and system. Background technique [0002] In recent years, artificial intelligence technology has developed rapidly and matured, and many industries have introduced related technologies to improve production efficiency and competitiveness. For example, the intelligent retail industry introduces image intelligent recognition technology to realize unmanned sales, intelligent statistical prediction of sales data, etc. Related deep learning tasks such as image processing, speech recognition, and machine translation usually use a Graphic Processing Unit (GPU for short) on a server cluster as the main computing resource. GPU resource management and task scheduling of server clusters are important links in business systems. [0003] The traditional GPU resource-oriented task scheduling method is to assign a GPU to each task when...

AI Technical Summary

Problems solved by technology

[0004] However, a single working server in a server cluster is configured with multiple GPUs, and each GPU has a certain amount of video memory resources. Deep learning tasks need to occupy the video memory resources in the GPU.

Since a single task often does not need to use the memory resources of the entire GPU, the traditional method is likely to cause waste of memory resources in the GPU, and the utilization rate of memory resources in the GPU is low

Images