Network Services Management using Programmable Data Planes for Visual Cloud Computing
Network Services Management using Programmable Data Planes for Visual Cloud Computing
复制标题
DOI：
10.1109/icnc57223.2023.10074183
复制
发表时间：
2023-02
期刊：
影响因子：
--
通讯作者：
Alicia Esquivel Morel;P. Calyam;Chengyi Qu;Durbek Gafurov;Cong Wang;Komal Thareja;A. Mandal;Eric J. Lyons;M. Zink;G. Papadimitriou;E. Deelman
中科院分区：
文献类型：
--
作者：
Alicia Esquivel Morel;P. Calyam;Chengyi Qu;Durbek Gafurov;Cong Wang;Komal Thareja;A. Mandal;Eric J. Lyons;M. Zink;G. Papadimitriou;E. Deelman
视觉云计算（VCC）应用为边缘/云基础设施上的视频数据处理管道提供了高效的解决方案。这些应用及其基础设施需要端到端监控和细粒度应用流量控制，以满足用户对体验质量的要求。在本文中，我们利用独立于协议的数据包处理器（P4）带来的可编程数据平面的优势，为 VCC 应用提出了一种新颖的网络服务管理方法。具体来说，我们定义了一个自定义的固定长度应用报头，并通过使用多跳路由检测（MRI）、带内网络遥测（INT）的变体和交换机端口转发（隧道）功能来避免拥塞，从而提高视频流应用的性能。在评估实验中，我们在国家科学基金会支持的 FABRIC 基础架构上的 VCC 测试平台中使用 P4 每数据包遥测元数据，用于路由路径、入口/出口时间戳、给定节点的队列占用率和出口端口链路利用率。实验结果表明，我们的方法在数据包丢失和吞吐量指标方面都提高了性能。
Visual Cloud Computing (VCC) applications provide highly efficient solutions in video data processing pipelines on edge/cloud infrastructures. These applications and their infrastructures demand end-to-end monitoring and fine-grained application traffic control to meet user quality of experience requirements. In this paper, we propose a novel network services management methodology for VCC applications by leveraging the advantages of programmable data planes enabled by Protocol-independent Packet Processors (P4). Specifically, we define a custom fixed-length application header and use it to improve the performance of a video streaming application through congestion avoidance using Multi-Hop Route Inspection (MRI), a variant of In-band Network Telemetry (INT), and switch port forwarding (tunneling) capabilities. For evaluation experiments, we use P4 per-packet telemetry metadata for routing paths, ingress/egress timestamps, queue occupancy in a given node, and egress port link utilization in a VCC testbed on the NSF-supported FABRIC infrastructure. Our experiment results demonstrate performance improvement obtained with our methodology in terms of both packet loss and throughput metrics.