TCPSpeaker

Abstract

The performance and fairness problems of TCP in 802.11 multi-hop wireless mesh networks (WMN) are well known [1, 2, 3]. Approaches to solve them fall roughly into two different categories: (i) those which maintain legacy TCP compatibility, and (ii) clean-slate protocols which fundamentally redefine WMN transport, breaking TCP compatibility. Clean-slate approaches such as block-switched and crosslayer transport protocols (e.g. HOP [4] and ATP [5]) break the end-to-end principle, utilizing intermediate WMN nodes for intelligent caching and pacing, demonstrating significant throughput and flow fairness improvements. Clean-slate approaches however, exclude TCP-speaking clients from the network and as such, their practical impact as well as realworld network evaluation potential is severely limited. We believe that the strengths of these two ostensibly opposed categories may in-fact be exploited by uniting them. We propose TCPSpeaker, a transport-protocol translator to “bridge” TCP flows entering and leaving the WMN at its edges. TCPSpeaker differs from other split-TCP approaches [3, 6], as it does not merely “split” TCP flows into multiple segments, but rather, it enables the removal and replacement of TCP entirely from within the WMN. To the mesh network users, TCPSpeaker presents a transparent, fully TCP-compliant interface. Inside the WMN, TCPSpeaker introduces the freedom to utilize a new transport protocol more suited to the broadcast nature of WMN forwarding (e.g. supporting opportunistic forwarding). TCPSpeaker may further prove useful beyond the scope of mesh networks, allowing any clean-slate transport to interact with legacy TCP end-hosts. We implement TCPSpeaker as a fully-functional Click Element [7] and evaluate its performance and behavior as an interface to TCP. We next discuss our design, implementation, and evaluation, identifying open questions and directions for further work.