On the Use of Kernel Bypass Mechanisms for High-Performance Inter-container Communications

Abstract

In this paper, we perform a comparison among a number of different virtual bridging and switching technologies, each widely available and commonly used on Linux, to provide network connectivity to co-located LXC containers for high-performance application scenarios.

Comparative Evaluation of Kernel Bypass Mechanisms for High-performance Inter-container Communications

“…Finally, OVS and Snabb follow. Comparing these performance with other evaluations present in literature (Ara et al, 2019), we were able to conclude that the major bottleneck for Snabb is represented by its internal L2 switching component. In addition, notice that while the maximum throughput in terms of Mpps is achieved with the smallest of the packet sizes (64 bytes), the maximum throughput in terms of Gbps is actually achieved for the biggest packet size (1500 bytes).…”

“…Despite the various techniques available to reduce the impact of the kernel on network performance, their use is often not sufficient to match the demanding requirements of typical NFV applications (Ara et al, 2019). In typical scenarios, the only solution to these problems is to resort to techniques that can bypass completely the kernel when exchanging packets both locally and between multiple hosts.…”

“…Finally, a preliminary work (Ara et al, 2019) was presented by the authors of this paper, performing a basic comparison of various virtual switching techniques for inter-container communications. That evalu-7 https://www.ntop.org/products/packet-capture/pf_ring/ ation however was limited to only a single transmission flow between a sender and a receiver application deployed on the same machine, for which SR-IOV was the most suitable among the tested solutions.…”

A Framework for Comparative Evaluation of High-Performance Virtualized Networking Mechanisms

“…Finally, OVS and Snabb, which lack similar optimizations, obtain inferior performance with respect to the other solutions. Comparing its performance with other evaluations present in literature that used Snabb only to connect directly two ports without a switching component in-between [3], we were able to conclude that its internal L2 switching component represents the major bottleneck for Snabb.…”

“…The same authors of this paper presented a preliminary work [3] that compared various virtual switching techniques based on kernel bypass for intercontainer communications. That evaluation was limited to techniques that bypass the Linux kernel and only a single unidirectional packet flow on a single host.…”