TCP challenges in hybrid military satellite networks; measurements and comparison

Abstract

Military tactical networks require reach back connectivity, and in most cases the only viable solution is a combination of satellite and radio links. This hybrid network is characterized by long delay, transmission errors and relative low capacity. Geostationary satellites add a substantial time delay. Terrestrial wireless packet networks might introduce comparable delays and jitter. Mobility and radio interference inherently introduce packet losses. In such environments, the widely deployed Internet Transport Layer protocols (TCP) are known to suffer severe performance degradation. This paper assesses the performance of widely deployed TCP variants. The default TCP algorithm in computers running Windows 7 is TCP NewReno, while Compound TCP is available. For Linux, a wide variety of variants are available, including the satellite-tailored variant TCP Hybla — the default flavor is CUBIC. This study shows that these TCP implementations exhibit significant different performances over wireless links. The evaluation and analysis is based on real TCP/IP traffic over an emulated link with characteristics of a combined terrestrial radio and satellite link. It was found that Windows 7 TCP variants performed poorly in lossy networks compared to CUBIC and Hybla, where the latter two outperformed the other two. Differences in performance of up to 40 per cent in favor of Hybla and CUBIC were measured. Furthermore, an unfair division of capacity, at the cost of Windows TCP implementations, was observed when CUBIC and Hybla competed for bandwidth against a Windows 7 transmitter. Depending on several factors, including the buffer size implemented in the network, this unfairness was measured to starve Windows-originated traffic with as low as 5 per cent of the total available bandwidth. Possible solutions improving fairness include tuning of the TCP parameters and avoid utilization of different TCP implementations simultaneously.