面向数据共享的模型训练服务系统  

A model training service system for data sharing

作　　者：魏宏原 华蓓[1,2] 林飞 Wei Hongyuan;Hua Bei;Lin Fei(School of Computer Science and Technology,University of Science and Technology of China,Hefei 230027,China;CAS Key Laboratory of Wireless-Optical Communications,Hefei 230027,China)

机构地区：[1]中国科学技术大学计算机科学与技术学院,安徽合肥230027 [2]中国科学院无线光电通信重点实验室,安徽合肥230027

出　　处：《网络安全与数据治理》2022年第8期20-29,共10页CYBER SECURITY AND DATA GOVERNANCE

基　　金：科技创新2030-“新一代人工智能”重大项目(2018AAA-0101200)。

摘　　要：数据驱动的人工智能应用需要大数据支持,然而现实中因隐私保护等原因,数据往往互不流通,而以孤岛形式存在。如何实现数据安全可用是当前亟待解决的问题。设计和实现了面向数据共享的模型训练服务系统,通过向用户提供数据功能服务接口而非数据本身,实现数据可用不可见。重点针对资源受限的数据共享平台,设计了高效的资源分配和作业调度方法,特别是通过自动资源缩放来应对多变的工作负载,达到优化用户体验和提高资源利用的目的。实验表明,相较于常规的作业调度方法,本系统在各种工作负载下都具有响应服务请求快、作业完成时间短的优点。Data-driven artificial intelligence applications require the support of big data.However,in reality,most of the data do not circulate with each other due to privacy leakage.How to achieve data security and availability is an urgent problem to be solved.This paper designs and implements a data sharing-oriented model training service system.By providing users with data function service interfaces instead of the data itself,data availability is invisible.This paper focuses on resource-constrained data sharing platforms,and designs efficient resource allocation and job scheduling methods,especially through automatic resource scaling to cope with changing workloads,to optimize user experience and improve resource utilization.Experiments show that,compared with the conventional job scheduling method,the system has the advantages of fast response to service requests and short job completion time under various workloads.

关 键 词：数据共享 模型训练服务 作业调度 资源分配 

分 类 号：TP14[自动化与计算机技术—控制理论与控制工程]