Networking over Next-Generation Satellite Systems


Thanks to both the rapid deployment of the Internet and advances in satellite technology, the market for broadband satellite services is poised for substantial growth in the coming decade. Current communications satellite systems have generally been designed to provide either voice or data transaction (low data rate) services through small terminals, or trunking (high data rate, or broadband) services through large terminals. However, technological advances are enabling new systems that combine broadband data rates with small terminals, thereby providing more affordable ``last-mile'' network access to home and small business users worldwide. In particular, two types of broadband satellite systems are under development: high-power satellites deployed at traditional geostationary (GEO) orbits, and large constellations of satellites deployed at much lower (LEO) orbits.

In this thesis, we explore research problems that have arisen from this shift in satellite network architectures. When using GEO satellites to provide Internet access service, the performance of the Internet's Transmission Control Protocol (TCP) is degraded by the high latency and high degree of bandwidth asymmetry present in such systems. We therefore undertook a comprehensive study of TCP performance in the context of broadband satellite systems used for network access. We first studied whether TCP's congestion avoidance algorithm can be adjusted to provide better fairness when satellite connections are forced to share bottleneck links with other (shorter delay) connections. Our data suggests that adjustments of the policy used in that algorithm may yield substantial fairness benefits without compromizing utilization. We next demonstrated how minor variations in TCP implementations can have drastic performance implications when used over satellite links (such as a reduction in file transfer throughput by over half), and from our observations constructed a satellite-optimized TCP implementation using standardized options. We explored the performance of TCP for short data transfers such as Web traffic, and found that two experimental options relating to how TCP starts a connection, when used together, could reduce the user-perceived latency by a factor of two to three. However, because not all of these options are likely to be deployed on a wide scale, and because even the best satellite-optimized TCP implementation is vulnerable to the fairness problems identified above, we explored the performance benefits of splitting a TCP connection at a protocol gateway within the satellite network, and found that such an approach can allow the performance of the satellite connection to approach that of a non-satellite connection. Carrying this work one step further, we construct a satellite-optimized transport protocol that can be used in such a split-connection environment, and demonstrate how it outperforms TCP in a satellite environment characterized by large amounts of bandwidth asymmetry. In particular, our protocol, which we have dubbed the ``Satellite Transport Protocol,'' uses up to an order of magnitude less traffic on the bandwidth-constrained reverse channel than is needed by TCP.

In contrast to research on GEO systems, research on LEO systems is still in its infancy. While LEO systems are being designed specifically to avoid the high latencies found in GEO systems, their design is challenging from a packet routing perspective due to the highly time-varying nature of the LEO network topology. Moreover, even the most basic system properties of such constellations is not well documented in the literature. We constructed a detailed packet-level LEO network simulator and identified some fundamental delay performance results of commercially proposed constellations. We then explored whether or not geographic-based network addresses can be constructively used in the design of distributed or centralized packet routing systems. We found that the construction of a distributed packet routing algorithm based on geographic-based packet forwarding is fundamentally challenging due to subtle topological properties of LEO networks. However, we demonstrated that geographic-based addresses are useful in centralized routing systems, enabling significant reductions in routing traffic and on-board routing tables. Specifically, we constructed routing strategies based on geographic-based addresses that potentially reduce by an order of magnitude the amount of routing traffic exchanged between satellite nodes and a centralized routing center on the Earth's surface, and the number of forwarding table entries required for non-local destinations.

Broadband satellite networks are likely to become an important niche of the future Internet because of their unique ability to provide network access from almost any point on the globe, particularly those underserved by terrestrial infrastructure. However, because the design of Internet protocols is driven by the performance of the wired Internet, satellite network engineers must be vigilant in assisting in the design and deployment of satellite-friendly protocols and in considering how satellite networks interwork with the wired Internet. Broadband LEO networks are likely to be deployed later than their GEO counterparts, and the design of these networks is still in its infancy, particularly since such networks are significantly more ambitious technically than anything that has been previously attempted. It is our hope that the findings presented in this thesis may contribute to a better understanding of how to design protocols for these next-generation systems while stimulating further work in this area.

The individual chapters: