Deep learning-oriented multi-type GPU cluster resource management scheduling method and system

Abstract

The invention discloses a deep learning-oriented multi-type GPU cluster resource management and scheduling method and system. The method comprises the following steps: dividing a GPU cluster into a plurality of GPU groups according to the model of a GPU, counting the idle operational capability of each GPU group, obtaining all users accessing the GPU cluster, and recording the minimum operationalcapability requirement of each user; and periodically accessing the job queue, obtaining the job to be processed with the highest priority in the job queue, and scheduling GPU cluster resources according to the job to be processed. According to the invention, GPUs of different brands and models are uniformly managed as one cluster for deep learning, the number of maintained GPU clusters is reduced, and the GPU cluster management complexity is simplified; the requirements of different users in deep learning can be met; reasonable user attributes are set according to user requirements, users donot need to be familiar with and care about GPU cluster environments, resource scheduling is carried out according to operational capability requirements and priorities of the users, resources meetingthe requirements can be automatically allocated through the scheduling method, and the resource utilization rate of different GPU type groups is increased.

Description

technical field [0001] The present application belongs to the field of high-performance computing, and specifically relates to a deep learning-oriented multi-type GPU cluster resource management and scheduling method and system. Background technique [0002] In many technical fields such as image, speech recognition, natural language processing, and reinforcement learning, deep learning has been proven to be very effective, and has reached or even surpassed human performance on some problems. However, deep learning has a great dependence on computing power, and the resource limitation of a single GPU often cannot meet the processing requirements for large-scale data and models. Multi-GPU parallel computing can effectively reduce the time for deep learning. Deep learning frameworks such as TensorFlow, Caffe, and PyTorch already support multi-GPU parallel computing, but when multi-GPU parallel computing, the best performance can only be achieved when the GPU models are the sa...

AI Technical Summary

Problems solved by technology

However, GPU cluster user management currently lacks such personalized requirements, and it is difficult to meet the needs of various users.

[0005] Moreover, the current GPU cluster management system has a relatively simple way of allocating job resources. It can only allocate the number of GPUs or specify a specific GPU. Users need to evaluate resource requirements and understand GPU parameters by themselves, which increases the difficulty of cluster users.

[0006] To sum up, the existing GPU cluster system scheduling function and user management cannot effectively meet the following requirements: different priorities of users, different computing power requirements; compatible with different GPU types, unified management in the same cluster, and easy expansion need

Images