In-Sender Reordering Compensation for RACK in Container Overlay Networks

Abstract

Container orchestrators like Kubernetes utilize packet encapsulation to construct container overlay networks, allowing for transparent communications among containers. While providing flexible connectivity with a minimal requirement for host machines, the inadvertent interplay with encapsulationinduced extra in-host hops and TCP’s default loss detection algorithm, Recent Acknowledgment (RACK), causes irregular in-sender reordering and spurious retransmissions (SRs). For a deeper understanding of the problem, we examine the behavior of RACK loss detection algorithm in the context of the packet datapath of the container overlay networks, which is not presumed by the RACK standard. Then we quantify the SRs of TCP using RACK in the production-level container overlay networks. Based on the in-depth analysis of the root causes of the SRs in container overlay networking, we derive a SR model induced by in-sender reordering and present a compensation mechanism for in-sender reordering. Our prototype implementation, centered around eBPF running in the Linux kernel, validates that the proposed compensation mechanism reduces SRs by up to 98.6% while maintaining the latency and throughput overhead below 2.3%.

Keywords

• TCP loss recovery
• RACK
• container networking
• spurious retransmissions