Distributed machine learning task online scheduling method based on side cloud cooperation

Abstract

The invention discloses a distributed machine learning task online scheduling method based on edge cloud cooperation. The method comprises the following steps: step 1, modeling an ML task scheduling problem of minimizing total completion time by using integer linear programming; 2, establishing a lower bound of a target function based on the irrelevant parallel machine model, replacing xjsw and njof the original problem with a new variable njsw, and customizing the original problem again; 3, decoupling the re-customized problem into two optimization problems; 4, for each arrived ML task, sequentially distributing Dj data blocks of the ML task to a worker; and step 5, recording a newly arrived ML task set as [Ja], and after all tasks in the [Ja] are allocated with works, sequentially judging whether each task is using the worker or not. The method has the advantages that cloud and edge server providers can be helped to maximize resource utilization of the cloud and edge server providers so as to obtain economic benefits as high as possible, and meanwhile, dynamic scheduling is performed according to the task load and the demand difference of each user so as to enable the overall training time to be shortest.

Description

technical field [0001] The present invention relates to the technical field of edge computing, in particular to an online scheduling method for distributed machine learning tasks based on edge-cloud collaboration. Background technique [0002] With the breakthrough of 5G and the emergence of edge computing, it has become a trend to push the frontier of artificial intelligence (AI) to the edge, and a new field - edge intelligence (EI) has been generated. Many large companies have conducted extensive research on EI and produced corresponding products. For example, IBM has developed IEAM for synchronous training of machine learning (ML) models, and Google has designed special application integrated circuits for running AI interfaces on edge networks. Edge TPU. In the field of EI, how to train ML tasks is an important issue. ML tasks with the characteristics of computation-intensive and time-consuming usually use data parallel mode and parameter server (PS) framework for model...

AI Technical Summary

Problems solved by technology

[0003] However, training ML models in EI faces many challenges: (1) Compared with remote clouds, edge servers have the advantage of low latency, but their computing resources are relatively scarce. Therefore, how to balance the edge-cloud time Delay and computing power are very important for model training; (2) due to limited edge server resources, it is difficult for all workers and PSs of an ML task to be placed on the same edge server, so workers and PSs need to communicate frequently, which will further slow down However, the research on EI is still in its infancy. Most of the existing research on ML task scheduling focuses on cloud systems, which cannot capture the characteristics of edge networks. In addition, most of the work on EI is real-time reasoning and training support technology for ML tasks. , did not study the impact of the ML task scheduling problem of the edge-cloud network on the model completion time. Therefore, we propose an online scheduling method for distributed machine learning tasks based on edge-cloud collaboration

Images