The Case for Redundant Arrays of Internet Links (RAIL)

TL;DR

This paper introduces RAIL, a network redundancy approach that enhances WAN performance and reliability by proactively replicating packets across multiple internet links, demonstrating significant improvements for VoIP and TCP applications.

Contribution

The paper proposes RAIL, a novel network architecture that uses multiple internet links for redundancy, along with a prototype and evaluation demonstrating performance gains.

Findings

Multiple links improve VoIP and TCP performance.

Delay padding is necessary for high disparity paths.

Two paths provide most benefits when managed properly.

Abstract

It is well-known that wide-area networks face today several performance and reliability problems. In this work, we propose to solve these problems by connecting two or more local-area networks together via a Redundant Array of Internet Links (or RAIL) and by proactively replicating each packet over these links. In that sense, RAIL is for networks what RAID (Redundant Array of Inexpensive Disks) was for disks. In this paper, we describe the RAIL approach, present our prototype (called the RAILedge), and evaluate its performance. First, we demonstrate that using multiple Internet links significantly improves the end-to-end performance in terms of network-level as well as application-level metrics for Voice-over-IP and TCP. Second, we show that a delay padding mechanism is needed to complement RAIL when there is significant delay disparity between the paths. Third, we show that two paths provide most of the benefit, if carefully managed. Finally, we discuss a RAIL-network architecture, where RAILedges make use of path redundancy, route control and application-specific mechanisms, to improve WAN performance.