Adaptive Address Family Selection for Latency-Sensitive Applications on Dual-stack Hosts

TL;DR

This paper tackles latency-sensitive transport in dual-stack hosts by showing that IPv4/IPv6 latency differences are destination-dependent and time-varying. It formulates an online learning approach based on EXP3 with $\gamma=0.1$ to adaptively select the lowest-latency address family per destination, enabling dynamic exploration and exploitation. A DNS resolver-based prototype (updns) implements the learning and demonstrates convergence toward the best address family in simulations and real-world browser tests, yielding meaningful latency gains. The work motivates DNS-empowered hints for address-family selection and points to extensions for IPv6 multihoming and broader latency metrics as practical next steps.

Abstract

Latency is becoming a key factor of performance for Internet applications and has triggered a number of changes in its protocols. Our work revisits the impact on latency of address family selection in dual-stack hosts. Through RIPE Atlas measurements, we analyse the address families latency difference and establish two requirements based on our findings for a latency-focused selection mechanism. First, the address family should be chosen per destination. Second, the choice should be able to evolve over time dynamically. We propose and implement a solution formulated as an online learning problem balancing exploration and exploitation. We validate our solution in simulations based on RIPE Atlas measurements, implement and evaluate our prototype in four access networks using Chrome and popular web services. We demonstrate the ability of our solution to converge towards the lowest-latency address family and improve the latency of transport connections used by applications.

Figures (6)

• Figure 1: Cumulative Distribution Functions (CDF) of HTTP request completion times (RCT) are similar across address families but RCT ratios have significant disparities when paired temporally.
• Figure 2: An example of probe-to-anchor RCT series split into four segments using change-point detection to highlight dynamicity inside categories of \ref{['tab:ripe-pairs-categories']} and produce \ref{['tab:ripe-pairs-segments']}.
• Figure 3: A client in a dual-stack network using the DNS resolver selecting IPv6 as the best performing address family towards a given domain name based on an online learning algorithm fed by transport metrics.
• Figure 4: EXP3 is largely able to choose the address family with the lowest completion time in simulations using the RIPE HTTP tests data from probes to anchors.
• Figure 5: Our prototype is largely able to choose the best address family in experiments using four access networks when loading popular websites with Chrome.