Research On Transparent Computing Based Architecture For Internet--of--Things

With the rapid development of the communication technol-ogy,mobile computing,and computer hardware,the development of the Internet-of-Things(Io T)architecture has gone through three stages,(1)the traditional Io T architecture based on wireless radio frequency and wireless sensor network,(2)the centralized Io T architecture based on cloud com-puting,(3)the distributed Io T architecture based on edge computing.Ac-cording to Statistica,the total number of connected devices will reach 75.44 billion by 2025,while the total amount of global data will grow from 33 ZB in 2018 to 175 ZB in 2025.Io T has shown its potential to revolutionize the whole society and attracted significant attention and investment from both industry and academia.Although edge computing provides new ideas for Io T applications,the explosive growth of heterogeneous devices and data also brings many challenges to Io T architectures:(1)Wide heterogeneity in devices and resources.An Io T system generally consists of various applica-tion scenarios.Devices include sensor nodes,Io T gateways,cloud comput-ing platforms,etc.Different devices require different hardware architec-tures and software resources,resulting in heterogenization of devices and software resources?(2)The management of heterogeneous devices and re-sources.Simple management operations such as updating software and set-tings which perform on different types of hardware and/or embedded operat-ing system versions are not an easy work,due to the tight coupling between hardware and software.(3)Dynamic service provisioning on lightweight Io T devices.Most Io T devices are limited by computing and storage capa-bilities,and can only write fixed software or functions according to scenario requirements,making it difficult to provide cross-platform and on-demand dynamic services for lightweight devices.(4)Computation allocation be-tween end devices and the edge.Lightweight devices will consume higher computing resources to process complex operations,it rises the problems of high computation latency and high energy consumption?(5)More and more data are collected by devices.Massive devices will bring explosive data growth,transmitting these data to the cloud computing center not only cannot guarantee data security,but may also bring enormous pressure to the network environment.The complexity,difficulty of use,and unmanageabil-ity of computing systems in the above challenges are mainly caused by the tight coupling of computer software and hardware.Transparent computing extends the computer’s system bus to the network,and separates comput-ing and storage,software and hardware to achieve decoupling.Transparent computing uses stream computing to execute software and services,so that end users can request applications on demand.Therefore,transparent com-puting has the characteristics of high flexibility,high scalability and high security.So,this paper studies the Io T architecture based on transparent computing and implements a system prototype to address the major chal-lenges.(1)Io T platform for heterogeneous devices.In order to solve the chal-lenges of wide heterogeneity in devices and resources and the management of heterogeneous devices and resources,we propose an Io T architecture based on transparent computing.The Io T architecture includes the end-user layer,the edge network layer,the core network layer,service & storage layer,and management layer.In transparent computing,the hardware and software of Io T devices are logically split,so that the operating system can also be regarded as a software resource.The end device sends a request to the server to obtain the required operating systems,applications,data and other resources,and executes it locally through streaming computing,so that the Io T architecture can support cross-platform and on-demand service pro-visioning for heterogeneous devices.The server enhances the manageability of the Io T system based on transparent computing through the unified man-agement of resources such as operating systems,applications,and data.We implement a prototype system of the architecture on a development board of the ARM instruction set,and tested the latency and power consumption of remote service loading on the server for a single development board.The experimental results demonstrate its effectiveness and superiority.(2)Io T platform for wearable devices.In order to solve the challenges of dynamic service provisioning on lightweight Io T devices and computa-tion allocation between end devices and the edge,this paper studies Io T platforms for wearable devices based on the Io T architecture of transparent computing and implements a prototype system,the lightweight device has the dynamic service capability by implementing the streaming computing of transparent computing on the lightweight device system.In order to im-prove the performance of dynamic services,this paper simple models the whole process of installing software on standard wearable devices,and mi-grates the steps that consume computing resources during the installation of software to the edge server.This paper implements a prototype system on a Cortex-M4F-based CPU,named TCWatch,and evaluates the performance of dynamic service provisioning on TCwatch in terms of the latency and energy consumption under different App sizes.The experimental results demonstrate that TCWatch can efficiently achieve dynamic service provi-sioning based on our proposed architecture.(3)Decentralized artificial intelligence accelerator.In order to solve the challenge of more and more data collected by devices,this paper studies the decentralized artificial intelligence accelerator under the Io T architec-ture based on transparent computing and implements the system through simulation.In order to make full use of unlabeled data from end and edge devices,this paper deploys an unsupervised learning model on the end and edge devices,which enable data to be processed locally without uploading to the cloud.In order to meet the needs of local training of lightweight devices,this paper uses resistive random access memory(Re RAM)based on processing-in-memory(PIM)technology as the devices’ s memory.We create a “Sharing While Training” method to reduce the delay caused by parameter sharing in the training process,and propose a “m-to-n” commu-nication pattern to reduce the network overhead when sharing parameters.We build a model of Re RAM using the simulator and test the performance of the accelerator on different artificial intelligence models.The experimen-tal results show that Re RAM can greatly improve the speed of training,and the in-memory computing method greatly reduces the power consumption caused by data migration between memory and CPU.There are 45 figures,15 tables,and 152 citations in this thesis.