弹性部署的分布式AI计算架构系统研究  

Research on distributed AI computing architecture system with flexible deployment

作　　者：李茜萌 陈春雨[1] 何恒翔 LI Ximeng;CHEN Chunyu;HE Hengxiang(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150000

出　　处：《应用科技》2020年第5期64-69,共6页Applied Science and Technology

基　　金：国家自然科学基金项目(61871142);基于人工智能架构的多传感器信息融合与决策系统的研究与实现(KY10800180032);中央高校基本科研业务费资助项目(No.3072020CFT0803).

摘　　要：为了解决嵌入式前端运行人工智能的现实需求与其自身性能不足的矛盾,本文提出了一种可弹性部署的分布式人工智能(AI)计算架构系统。本系统通过一套经过精心设计的网络通讯程序,将嵌入式设备需要计算的数据发送至云端,然后利用云端的GPU工作站集群进行高速计算。本系统能高效管理云端各个工作站的运行,可弹性部署的特性使其能依据计算负载的不同,方便快捷地增加或者减少算力,实现节能与算力的最优化。分布式的特性使本系统具有一定的灾备冗余特性,有效避免了因某几个计算节点崩溃导致整个系统瘫痪,并且运用了集群计算模式,大幅减少了系统对公网IP数量的要求。In order to resolve the contradiction between the practical needs of artificial intelligence(AI)algorithm running on the front end of embedded device and the insufficient performance,a distributed AI computing architecture system with flexible deployment is proposed in this paper.The system sends the data that embedded devices need to calculate to the cloud through a well-designed network communication program,then GPU workstation clusters in the cloud compute at high-speed.The system can efficiently manage each workstation in the cloud.Because the system can be deployed flexibly,it is capable to increase or decrease the computing power conveniently according to different computing loads.The characteristic of distributed architecture makes sure the system has certain redundancy attribute for disaster recovery.Thus,it effectively avoids that the collapse of some computing nodes causes the entire system crash.In addition,because the system works in cluster computing mode,it largely reduces the number of public network IP required by the system.

关 键 词：人工智能 分布式计算 深度学习 云计算 嵌入式设备 网络通讯 工作站集群 神经网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]