Obtaining Accurate Bandwidth Estimations for the Internet of Things

Abstract

Bandwidth estimation is important for many Internet of Things (IoT) applications. Typical bandwidth estimation techniques involve saturating the link and causing temporary congestion; however, this may not be suitable for latencysensitive IoT applications. Packet pairs and trains are alternative bandwidth estimation methods that use packet inter-arrival time. However, these techniques are often noisy [14] and inaccurate. One source of this inaccuracy is caused by kernel interrupt handling and scheduling [22]. In this paper, we propose our PacketBurst bandwidth estimation technique. The main contribution of PacketBurst is to remove inaccuracies in packet train bandwidth estimate calculations that are caused by kernel interrupts. We observe that consecutive packets in packet trains may have exceedingly small differences in their receive timestamps. Using these timestamps in a packet train bandwidth estimate calculation can result in an inaccurate estimate. We identify four cases where kernel interrupts impact the packet train’s receive timestamps, and we show how to address the timestamp inaccuracies of each case using our PacketBurst technique. We show that PacketBurst estimates can reduce measurement errors of packet trains. When using six MTU-sized packets to estimate available bandwidth, the PacketBurst technique experiences an 2.75% error whereas packet trains experience a 21.6% error when the available bandwidth is 20Mbps.