Home // CLOUD COMPUTING 2026, The Seventeenth International Conference on Cloud Computing, GRIDs, and Virtualization // View article
Accelerated Flow Processing in Kubernetes Overlay Networks
Authors:
Srinath Vasudevan
Khaled Harfoush
Keywords: Kubernetes; Calico; Overlay Networks; Receive Flow Steering; Receive Packet Steering.
Abstract:
Kubernetes is a popular container orchestration tool which makes managing numerous containers simple through various abstractions. Pods are the abstraction for containers, and Kubernetes allows for pod inter-networking through the Container Network Interface (CNI) specification. Overlay networking is a technique which utilizes network tunneling across hosts to allow pods to communicate with each other using their private IP addresses. Previous work shows that overlay networking results in significant network performance degradation due to the packet encapsulation and decapsulation processing steps required for network tunneling, largely because all packet processing is usually performed on a single CPU core by default. Optimizations in the Linux kernel exist, such as Receive Packet Steering (RPS) and Receive Flow Steering (RFS), in order to parallelize packet processing per flow. The current literature does not thoroughly evaluate RFS, nor does it evaluate RPS when enabled at the container level in overlay networks. In this paper, we conduct a thorough measurement study and a performance analysis of these optimizations. Our results demonstrate the greatest improvements in average bitrate and CPU core load distribution when RFS is enabled at the host and container levels, although improvements are seen in other scenarios as well. This important work outlines actionable methods for overlay network performance gains that are applicable to public cloud systems.
Pages: 48 to 55
Copyright: Copyright (c) IARIA, 2026
Publication date: April 19, 2026
Published in: conference
ISSN: 2308-4294
ISBN: 978-1-68558-373-6
Location: Lisbon, Portugal
Dates: from April 19, 2026 to April 23, 2026