High-performance cluster resource fair allocation method based on multi-agent reinforcement learning

Abstract

The invention relates to a high-performance cluster resource fair allocation method and system based on multi-agent reinforcement learning. The method comprises the following steps: S1, establishing a Markov game model of high-performance cluster resource scheduling; s2, collecting real cluster data, performing job playback by using a simulation environment, and constructing a high-performance cluster simulation environment; s3, training the strategy and state value evaluation network in a high-performance cluster simulation environment; wherein the strategy and state value evaluation network comprises an action strategy neural network NNactor and a value evaluation neural network NNcritic, and corresponding loss functions are respectively constructed for parameter updating. According to the method provided by the invention, the resource use fairness among the users can be maintained under the condition that the cluster resource utilization rate is not influenced.

Description

technical field [0001] The invention relates to the field of resource scheduling of high-performance clusters, in particular to a method and system for fair resource allocation of high-performance clusters based on multi-agent reinforcement learning. Background technique [0002] In recent years, deep learning has made tremendous progress and development in a large number of different fields, such as computer vision, image recognition, natural language processing, recommendation algorithms. In order to improve the accuracy of training results, the scale of the models that people use continues to increase, and the amount of training data continues to expand. For example, for the BERT model proposed by Google in 2018, the pre-training process used 16 TPU v3 chips, and it took 3 days to complete the pre-training task. It took 29 hours to train a Resnet-50 model on 8 Tesla P100 GPUs. With the continuous expansion of training calculations, in order to ensure that the training t...

AI Technical Summary

Problems solved by technology

Every user hopes that the jobs submitted by them can be run efficiently, but the limited resources of the cluster are doomed to make it impossible for every user's jobs to be executed immediately

Therefore, the problem can be modeled as a multi-agent scheduling problem: each user hopes to optimize the waiting time of his own job, but the amount of resources is rated, so a resource competition relationship is formed among users

Images