Evolving Switch Architecture toward Accommodating In-Network Intelligence

Abstract

Motivated by the breakthroughs of AI in both theory and applications, we are perceiving a great potential for network innovations from a new dimension of intelligence. However, as the engine of networks, switches are designed as dumb network elements with the sole purpose of forwarding network packets, and thus a barrier against the entrance of AI as an intrinsic part of the network is unconsciously erected. This article proposes an evolved switch architecture aimed at breaking this barrier and accommodating in-network intelligence. We enhance the current switch architecture by embedding an intelligence plane, which is externalized as an intelligent computation pluggable module of an evolved switch. The module employs an integrated solution of X86 CPU+GPU+DPDK, which provides a high-performance and high-throughput open platform for hosting in-network intelligence. We further conceive a flexible processing framework for an intelligent traffic measurement, recognition, and traffic regime. We also carry out extensive experiments to demonstrate the capability of the evolved switch by deploying a prototype in a campus network, with two promising application scenarios: in-network application identification and in-network anomaly detection.