Characterizing networking performance and interrupt overhead of container overlay networks

Abstract

Containers, an emerging service to manage and deploy applications into isolated boxes, are quickly increasing in popularity in the cloud and edge computing. In order to provide connectivity among multiple hosts, cloud providers adopt overlay networks, which not only impose significant overhead in throughput and latency in containerized applications, but also consume more CPU resources of the system. Through profiling and code analysis, this paper reveals that the overwhelming interrupts, as well as its load imbalance in the kernel processing contribute to the inefficiency of the container overlay networks. Specifically, every packet in container networks might raise multiple software interrupts compared to that in VM networks. Our results indicate that the container network throughput drops 2/3 and the tail latency increases more than 37 times if the interrupt overhead is not well optimized.