| Abstract: |
Many people around the globe enjoy the amenities of a ubiquitously available communication infrastructure. This infrastructure is used in a variety of scenarios: It is used for personal
communication with friends or business partners, to obtain information or entertainment from a service provider, or for buying things without being physically present in a shop. Additionally, the
same infrastructure is used by electronic devices to share information. For example, sensor nodes measure temperature, humidity, air pollution etc., and share this information with a central
server which gathers aggregated knowledge from the individual measurement points. Or different local sites of a company might exchange a high data volume to synchronize and / or back up their
data base. Each of these applications impose different communication re- quirements on the infrastructure, ranging from different levels of security and privacy, to different allowable
transmission delays, to different requirements on the transmission reliability. However, all these applications run over the same network architecture, the Internet architecture. Currently, many
add-ons to the original architecture make this possible. However, it is questionable whether this add-on approach is scalable an can accommodate future applications. Therefore, many researchers
develop new network architectures from scratch, so called clean slate architectures, that are designed to be more appropriate for future challenges. In this dissertation we have developed such a
clean slate network architecture called Dynamic Protocol Stack (DPS) architecture. The DPS architecture tackles the challenge of providing a future-oriented architecture by introducing
flexibility at two levels. First, newly iv Abstract developed communication protocols can be included seamlessly in the architecture and second, the protocol stack used for the communication can
be adapted to the current environment and communication needs while transferring data. However, such a flexible network architecture poses a new challenge: How can it be implemented without
suffering from performance degradation introduced by the flexibility? We address this challenge by providing two different implementations of the DPS architecture. With the first implementation
we show how the DPS architecture can be implemented on a general-purpose CPU, and we show that it has the same performance as an implementation of the Internet architecture on the same system.
With the second implementation we show how the DPS architecture can make use of hardware acceleration for the individual network protocols. Therefore, we implemented the DPS architecture on an
FPGA-based system-on-chip platform. This implementation, called EmbedNet, allows for the dynamic run-time mapping of network functionality to either software or hardware. Thus it offers hardware
acceleration for arbitrary network functionality in the DPS architecture as dedicated hardware for a specific network protocol would in the Internet architecture. Additionally, we show in two use
cases the benefits of having a dynamic network architecture. First, we show how we can minimize protocol overhead by adapting the protocol stack to the current network conditions. Second, we show
how we can improve the system performance when adapting the hardware / software mapping to the current network traffic. Finally, we discuss how more advanced optimization algorithms could be
developed to optimize the provided network functionality. To summarize, with this dissertation we have shown that it is feasible to provide dynamic network architectures and our results suggest
that it is worthwhile to go further in that direction. |
