GPU resource-based data processing method and system, and electronic equipment

Abstract

The invention discloses a GPU resource-based data processing method and system, and electronic equipment. The GPU resource-based data processing method comprises the steps of creating N virtual GPU resources according to a resource configuration strategy and physical GPU resources; generating a resource mapping relationship between the physical GPU resources and the virtual GPU resources accordingto a resource configuration strategy; and enabling each resource application module to obtain a target virtual GPU resource in the N virtual GPU resources according to the resource mapping relationship.

Description

technical field [0001] The embodiments of this specification relate to the field of computers, and in particular to a GPU resource-based data processing method, electronic device, and system. Background technique [0002] AI (Artificial Intelligence, artificial intelligence), especially DL (Deep Learning, deep learning) has been widely used in payment (face), loss assessment (picture recognition), interaction and customer service (speech recognition, text content extraction and filtering), etc. This scenario has achieved remarkable results. Typical DL tasks require powerful computing support, so most of the current tasks run on acceleration devices such as GPU (Graphics Processing Unit, Graphics Processing Unit). GPU is a high-performance computing acceleration device that is currently widely used in AI, For deep learning training and online services, the hardware of the physical GPU is driven by the GPU driver to realize the training of AI tasks, online services, etc. [...

AI Technical Summary

Problems solved by technology

However, the isolation of AI learning applications cannot be achieved, and there are resource preemption and mutual interference between different AI learning applications

And cannot schedule computing tasks to different nodes, therefore, cannot support migration, resulting in inflexible scheduling

For example, the solution based on remote system calls: at the time of dynamic running, the GPU resource request is intercepted, and then forwarded to the target to obtain the bound GPU resource. Since the resource has been bound, it cannot support hot migration (that is, it cannot be dynamically and timely Migrate the application between different physical GPUs on the same machine or between physical GPUs on different machines), which leads to inflexible scheduling

Images