摘要
物联网智能设备产生的大量并发AI数据流给云端处理中心带来了巨大的挑战,为了应对这一挑战,边缘计算将海量的并发AI数据流以流水线加工的方式将这些并发AI数据流分配给边缘服务集群内的计算节点处理。如何利用计算节点内有限的计算资源、以较低的成本提高并发AI数据流的处理与通信效率是本文研究的目标。提出了一种能够在处理并发AI数据流的计算节点内使用的通信模型,该通信模型结合DPDK的核绑定机制为并发AI数据流的接收过程、计算过程、发送过程均衡地绑定CPU核,还加入了数据流的分类计算、网卡端口的调度策略、缓冲环和全局网卡端口的负载监控。实验分析表明,并发AI数据流处理节点内的通信模型能够有效制定CPU核的均衡绑定策略,提高流处理节点之间的并发AI数据流的处理效率,还实现了多网口的均衡调度策略,使网卡端口的负载达到均衡状态,不会对端口造成太大的负载,同时带宽利用率和通信速率也大大提高,并且降低了边缘集群中流处理节点的部署成本,合理利用了节点内的计算资源处理并发AI数据流。
The large number of concurrent AI data streams generated by IoT smart devices has brought great challenges to the cloud processing center. In order to meet this challenge, edge computing processes the massive concurrent AI data streams and distributes these concurrent AI data streams to compute node processing within the edge service cluster. How to use the limited computing resources in computing nodes to improve the processing and communication efficiency of concurrent AI data streams at a lower cost is the goal of this paper. A communication model used in computing nodes that process concurrent AI data streams is proposed. Combined with the core binding mechanism of DPDK, the communication model binds CPU cores in a balanced manner for the receiving process, computing process, and sending process of concurrent AI data streams, and also adds the classification calculation of data flow, the scheduling policy of NIC ports, the buffer ring and the load monitoring of global NIC ports. Experimental analysis shows that the communication model within the concurrent AI data stream processing node can effectively formulate a balanced binding strategy for CPU cores, improve the processing efficiency of concurrent AI data streams between stream processing nodes, and achieve a balanced scheduling strategy for multiple network ports, so that the load of the network card port reaches a balanced state, which will not cause too much load on the port. At the same time, the bandwidth utilization rate and communication rate are also greatly improved, and the deployment cost of the stream processing nodes in the edge cluster is reduced, and the computing in the node is rationally utilized to handle concurrent AI data streams.
作者
黄东生
陈庆奎
HUANG Dongsheng;CHEN Qingkui(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《智能计算机与应用》
2022年第11期26-33,40,共9页
Intelligent Computer and Applications
基金
国家自然科学基金(61572325)
上海重点科技攻关项目(16DZ1203603,19DZ1208903)
上海智能家居大规模物联共性技术工程中心项目(GCZX14014)
上海市一流学科建设项目(XTKX2012)
上海市重点项目(9DZ1208903)。
关键词
并发AI数据流
计算资源
均衡
DPDK
流处理节点
concurrent AI data streams
computing resources
balance
DPDK
stream processing nodes
