Edge-targeted And FPGA-accelerated Distributed Stream Processing System

The rapid development of the Internet of Things technology has led to the explosive growth of stream data,and the real-time requirements of various applications are also getting higher and higher.Traditional cloud computing model has been unable to process those high-speed and massive stream data efficiently and timely,processing stream data in the Edge is becoming more imperative.However,compared to resource-rich clouds,Edge has poorer infrastructures and limited computing power,thus the performance of stream data processing at the edge is also limited,which can not meet the growing demand of high requirements of real-time.Field Programmable Gate Array(FPGA)which has massive parallel computing power,extremely low power consumption,and supports for flexible reconfiguration,provides a new way to improve the computing power of edge cluster and realize more efficient stream data processing.Therefore,it is proposed to use FPGA to accelerate the stream processing systems deployed in the edge cluster: By deploying the FPGA boards on edge servers and integrating them with distributed stream processing(DSP)system in the edge cluster,FPGAs can be used to accelerate stream computing tasks,thereby improving the performance of the stream data processing.Specially,an Edge-oriented and FPGAaccelerated distributed stream processing system(F-Storm)is designed and implemented: In order to reduce the overhead of system deployment and running,a light-weight manager is designed to integrate and manage FPGAs.For the scenarios of insufficient FPGA resources,an adaptive accelerator-prioritized scheduling strategy is designed to perform task scheduling,it makes full use of FPGA resources and can achieving reverse offloading of tasks from FPGA to CPU.Meanwhile,the JVM-FPGA data transmission mechanism is optimized,data batching and data pipelining techniques are adopted to reduce the time overhead of data transmission between JVM and FPGA,thus reducing data processing delay.In addition,to solve the problem that the use of FPGAs complicates the development of stream applications,F-Storm provides an easy-to-use programming interface that greatly reduces the burden of developing stream applications.Developed on Storm,F-Storm reduces the data processing latency by 36%-75% compared to Storm,gains throughput improvement of 1.4x-3.2x,and greatly reduces CPU utilization of some key threads,improving the performance of stream data applications.