Can machine learning benefit bandwidth estimation at ultra-high speeds?

Abstract

Tools for estimating end-to-end available bandwidth (AB) send out a train of packets and observe how inter-packet gaps change over a given network path. In ultra-high speed networks, the fine inter-packet gaps are fairly susceptible to noise introduced by transient queuing and bursty cross-traffic. Past work uses smoothing heuristics to alleviate the impact of noise, but at the cost of requiring large packet trains. In this paper, we consider a machine-learning approach for learning the AB from noisy inter-packet gaps. We conduct extensive experimental evaluations on a 10Gbps testbed, and find that supervised learning can help realize ultra-high speed bandwidth estimation with more accuracy and smaller packet trains than the state of the art. Further, we find that when training is based on: (i) more bursty cross-traffic, (ii) extreme configurations of interrupt coalescence, a machine learning framework is fairly robust to the cross-traffic, NIC platform, and configuration of NIC parameters.