Resource scheduling method, device, and storage medium

Patent metadata

Field	Value
Publication number	US-12399745-B2
Application number	US-202217744396-A
Country	US
Kind code	B2
Filing date	May 13, 2022
Priority date	Jul 7, 2021
Publication date	Aug 26, 2025
Grant date	Aug 26, 2025

Abstract

A resource scheduling method and apparatus, a device, and a storage medium are provided, and relates to the field of computer technology, and in particular to the field of deep learning technology. The method includes: acquiring a graphics processing unit (GPU) topology relationship of a cluster according to GPU connection information of each of computing nodes in the cluster; and in a case where a task request, for applying for a GPU resource, for a target task is received, determining a target computing node of the target task and a target GPU in the target computing node according to the task request and the GPU topology relationship, to complete GPU resource scheduling of the target task. The present disclosure can optimize the resource scheduling.

First claim

What is claimed is:

1. A resource scheduling method, comprising:
acquiring a graphics processing unit (GPU) topology relationship of a cluster according to GPU connection information of each of computing nodes in the cluster;
in a case where a task request, for applying for a GPU resource, for a target task is received, determining a target computing node of the target task and a target GPU in the target computing node according to task information of the task request and the GPU topology relationship, to complete GPU resource scheduling of the target task, wherein the task information comprises a task type of the target task and/or a number of GPUs applied for; and
executing the target task by using the target GPU,
in a case where the task type of the target task comprised in the task information is single-task and the number of GPUs applied for is 1, traversing each of the computing nodes according to the GPU topology relationship, and selecting a computing node with a lowest transmission performance as the target computing node;
in a case where the task type of the target task comprised in the task information is single-task and the number of GPUs applied for is larger than 1, traversing each of the computing nodes according to the GPU topology relationship, and selecting a computing node with a highest transmission performance as the target computing node;
in a case where the task type of the target task comprised in the task information is multi-task and the number of GPUs applied for is larger than 1, traversing each of the computing nodes according to the GPU topology relationship, and determining whether there are candidate computing nodes with GPU resources that are able to satisfy the target task in the cluster, and determining the target computing node according to whether the candidate computing nodes exist;
wherein after determining the target computing node of the target task, a GPU connection level selected for the target task is written into an annotation of the target task, and determining the target GPU in the target computing node according to the task information of the task request and the GPU topology relationship comprises:
reading the GPU connection level selected for the target task; and
selecting a corresponding GPU as the target GPU according to the GPU connection level and the GPU topology relationship,
wherein the GPU connection information comprises a GPU connection relationship and a GPU connection manner, and the acquiring the GPU topology relationship of the cluster according to the GPU connection information of each of the computing nodes in the cluster, comprises:
acquiring the GPU connection manner, the GPU connection relationship, and a node connection relationship of each of the computing nodes in the cluster;
determining a GPU connection level of each of the computing nodes according to the GPU connection manner of each of the computing nodes; and
obtaining the GPU topology relationship according to the GPU connection level, the GPU connection relationship, and the node connection relationship of each of the computing nodes,
wherein the GPU connection level is used to represent a transmission performance between GPUs.

2. The method of claim 1 , further comprising:
in a case where the task type of the target task comprised in the task information is multi-task and the number of GPUs applied for is larger than 1,
when a traversal result indicates that there are candidate computing nodes with GPU resources that are able to satisfy the target task, selecting a computing node with a highest transmission performance from the candidate computing nodes as the target computing node.

3. The method of claim 1 , further comprising:
in a case where the task type of the target task comprised in the task information is multi-task and the number of GPUs applied for is larger than 1,
when a traversal result indicates that there are no candidate computing nodes with GPU resources that are able to satisfy the target task, for each task in the target task, selecting a computing node, in which a number of GPUs is able to satisfy the task, from the computing nodes as the target computing node, according to a descending order of transmission performances of the computing nodes.

4. The method of claim 1 , wherein the determining the target GPU in the target computing node according to the task request and the GPU topology relationship, further comprises:
acquiring a GPU connection level and a GPU connection relationship of the target computing node according to the GPU topology relationship; and
selecting the target GPU from GPUs comprised in the target computing node according to the GPU connection level of the target task, and the GPU connection level and the GPU connection relationship of the target computing node.

5. An electronic device, comprising:
at least one processor; and
a memory connected communicatively to the at least one processor, wherein
the memory stores instructions executable by the at least one processor, and the instructions, when executed by the at least one processor, enable the at least one processor to perform operations of:
acquiring a graphics processing unit (GPU) topology relationship of a cluster according to GPU connection information of each of computing nodes in the cluster;
in a case where a task request, for applying for a GPU resource, for a target task is received, determining a target computing node of the target task and a target GPU in the target computing node according to task information of the task request and the GPU topology relationship, to complete GPU resource scheduling of the target task, wherein the task information comprises a task type of the target task and/or a number of GPUs applied for; and
executing the target task by using the target GPU,
in a case where the task type of the target task comprised in the task information is single-task and the number of GPUs applied for is 1, traversing each of the computing nodes according to the GPU topology relationship, and selecting a computing node with a lowest transmission performance as the target computing node;
in a case where the task type of the target task comprised in the task information is single-task and the number of GPUs applied for is larger than 1, traversing each of the computing nodes according to the GPU topology relationship, and selecting a computing node with a highest transmission performance as the target computing node;
in a case where the task type of the target task comprised in the task information is multi-task and the number of GPUs applied for is larger than 1, traversing each of the computing nodes according to the GPU topology relationship, and determining whether there are candidate computing nodes with GPU resources that are able to satisfy the target task in the cluster, and determining the target computing node according to whether the candidate computing nodes exist;
wherein after determining the target computing node of the target task, a GPU connection level selected for the target task is written into an annotation of the target task, and determining the target GPU in the target computing node according to the task information of the task request and the GPU topology relationship comprises:
reading the GPU connection level selected for the target task; and
selecting a corresponding GPU as the target GPU according to the GPU connection level and the GPU topology relationship,
wherein the GPU connection information comprises a GPU connection relationship and a GPU connection manner, and the acquiring the GPU topology relationship of the cluster according to the GPU connection information of each of the computing nodes in the cluster, comprises:
acquiring the GPU connection manner, the GPU connection relationship, and a node connecti

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Xu Binbin
• Tang Liang
• Zhao Ying
• Li Shupeng
• Shi En
• Qian Zhengyu
• Xie Yongkang

Classifications

• G06F9/5044 Primary

  considering hardware capabilities · CPC title

• G06T1/20

  Processor architectures; Processor configuration, e.g. pipelining · CPC title

• Y02D10/00

  Energy efficient computing, e.g. low power processors, power management or thermal management · CPC title

• G06N3/098

  Distributed learning, e.g. federated learning · CPC title

• G06F2209/509

  Offload · CPC title