GPU-based cluster resource allocation method and device

Abstract

The invention discloses a GPU-based cluster resource allocation method. A GPU cluster comprises a plurality of GPU cards. The method comprises the steps of obtaining to-be-processed tasks, wherein theto-be-processed tasks comprise a large task and a small task; the big task refers to a to-be-processed task of which the required resource quantity is greater than or equal to one GPU card; the small task refers to a to-be-processed task of which the required resource quantity is smaller than that of one GPU card; according to the priority sequence of the to-be-processed tasks, allocating one ormore GPU cards to the big tasks for execution; only one big task is allocated to each GPU card; obtaining the resource surplus of each GPU card executing the task; and for each GPU card with the resource surplus, traversing the unexecuted small tasks according to a priority sequence, if the resource surplus of the GPU cards is found to meet the unexecuted small tasks, allocating the resource surplus of the GPU cards to the small tasks, and updating the resource surplus of the GPU cards. Through the scheme of the invention, the GPU cluster resource utilization rate is improved.

Description

technical field [0001] The invention relates to the field of cloud computing, in particular to a method and device for resource allocation based on GPU clusters. Background technique [0002] Currently, resource scheduling software commonly used in GPU clusters (systems that manage a large number of GPU cards) include pbs (PortableBatch System) and slurm (Simple Linux Utility for Resource Management). Using existing resource scheduling software makes the utilization rate of GPU clusters low A GPU (Graphic Processing Unit, Graphics Processing Unit) card that allocates processing tasks has a usage rate of about 30% to 60%, and about 30% of the resources are not used, resulting in waste. [0003] The unit of the scheduler software pbs and slurm to allocate GPU cards is the number of cards. When scheduling GPU resources, one or more GPU cards are allocated to tasks. When tasks are running, the utilization rate of GPUs is not high, resulting in Some GPU resources are wasted. At...

AI Technical Summary

Problems solved by technology

[0002] Currently, resource scheduling software commonly used in GPU clusters (systems that manage a large number of GPU cards) include pbs (PortableBatch System) and slurm (Simple Linux Utility for Resource Management). Using existing resource scheduling software makes the utilization rate of GPU clusters low , the GPU (Graphic Processing Unit, Graphics Processing Unit) card that allocates processing tasks has a utilization rate of about 30% to 60%, and there is about 30% of resources that are not used and cause waste

[0003] The unit of the scheduler software pbs and slurm to allocate GPU cards is the number of cards. When scheduling GPU resources, one or more GPU cards are allocated to tasks. When tasks are running, the GPU utilization rate is not high, resulting in Some GPU resources are wasted

At the same time, the resource scheduling of the existing scheduler software pbs and slurm has multiplexing of GPU cards, that is, more than two tasks can be run on one GPU card. Although this can make full use of GPU resources, but because a GPU card simultaneously Running multiple tasks slows down task execution, resulting in very low work efficiency and longer task completion cycles

Moreover, the current technology only implements resource scheduling, and there is no related technology for rational allocation of resources. In view of the problems existing in the existing GPU cluster resource scheduling, it is urgent to propose a method based on GPU cluster resource allocation. Method and device capable of improving GPU cluster utilization

Images