GPU resource elastic scheduling method based on heterogeneous application platform

Abstract

The invention discloses a GPU resource elastic scheduling method based on a heterogeneous application platform. The method comprises the following steps: a) obtaining GPU resource utilization rate information; b) setting a trigger threshold value and the number of times; c) screening and sequencing the capacity-reducing platform queues; d) screening and sequencing the expansion platform queues; with the steps of 1) selecting a to-be-reduced platform; 2) establishing a GPU node list; 3) processing a locking state node; 4) offline the to-be-migrated node; 5) adding to a resource queue; and 6) judging whether the volume reduction is finished or not. According to the GPU resource elastic scheduling method, flexible adjustment can be conducted according to the GPU load condition of the whole platform, so that maximized utilization of platform GPU resources is achieved, scheduling of the platform is mainly achieved through an existing scheduling assembly of an adaptive platform on the lower layer, dynamic resource monitoring, information collection and execution operation issuing are achieved through interface calling. And rapid and flexible deployment and implementation of cloud computing, big data, artificial intelligence and high-performance computing scene platforms can be satisfied.

Description

technical field [0001] The present invention relates to a method for flexible scheduling of GPU resources, and more specifically, to a method for flexible scheduling of GPU resources based on a heterogeneous application platform. Background technique [0002] Graphics processing unit (GPU) resources have been increasingly used in the fields of cloud computing, artificial intelligence, and high-performance computing in recent years due to their excellent parallel computing capabilities, higher bandwidth, and main frequency. . At the same time, because the price of GPU resources is generally higher than that of CPU, GPU resources are scarce resources in different computing application scenarios. In order to improve the utilization rate of GPU resources, it is generally realized mainly through resource scheduling. [0003] GPU resource scheduling can generally be divided into task-level scheduling, hardware-level scheduling, and node-level scheduling. The task-level schedulin...

AI Technical Summary

Problems solved by technology

However, this method must be unique to the scheduling system used on the same platform, otherwise it will lead to inconsistent scheduling information for GPU resources, resulting in resource occupation conflicts; the commonly used scheduling platform for physical node-level scheduling is the same as the hardware-level scheduling platform, and all GPU computing nodes belong to It is based on a certain resource management platform and does not involve sharing and multiplexing. Its elastic expansion is limited to the inside of the platform. It does not support heterogeneous resources and heterogeneous application expansion, which is not conducive to improving the overall GPU resource utilization of the platform.

Images