Emulating High-performance Networks with CNNet

Metaverse advocates creating digital twins in the virtual world to faithfully emulate the behaviors of the physical counterparts. Following this idea, this paper explores how to create high-performance digital twins for network devices through network emulation. Existing network emulators are primarily used for network configuration verification. They can not emulate high-performance networks, and lose the ability to carry out performance-oriented tasks, like stress-testing datacenter networks. We present CNNet, a novel network emulator that is designed to emulate high-performance networks and execute performance-oriented tasks. CNNet adopts a cloud-native design and achieves the performance goal with the user-space dataplane, which provides high-performance virtual network links for different emulation containers. To further improve the emulation performance, a novel graph partition algorithm is proposed to embed the emulated network into the underlying server cluster. The embedding achieves two goals simultaneously, including balancing the number of emulation containers running on each worker server, and minimizing the bandwidth consumption of the cross-server virtual links. When emulating a datacenter network with over 100 nodes, CNNet can achieve the bi-section throughput of 96Gbps, which is an order of magnitude higher than existing network emulators.