Pangolin: Incast Avoidance in Large Datacenter Networks
Sae Hyong Park, Namseok Ko · 2023
In the fast-evolving domain of network technology, the efficient management of In-Network Functions is pivotal, especially in the realm of data center networks that deploy distributed applications. This paper presents an exploration of the incast problem, a significant bottleneck in network performance. Incast occurs when multiple servers concurrently transmit data to a single destination, leading to congestion and network resource saturation. This issue is commonly observed in data merging within large-scale parallel distributed applications. The core contribution of this paper is a novel traffic engineering approach for datacenter networks. We propose the integration of Segment Routing with IPv6 (SRv6) and Programmable Data Plane technologies to effectively address the incast challenge. This innovative approach aims to optimize network performance and reduce the limitations and overhead associated with conventional methods, offering a comprehensive solution to enhance data center network efficiency.