Coping with limited communication resources in distributed real-time systems is a major challenge nowadays. With regard to high-dependable and safety-critical systems (e.g., flight control systems and advanced driver assistance systems) the use of a time-triggered architecture is advantageous, since the periodic task executions and message exchanges according to a static schedule maximize the predictability compared to event-triggered systems. These systems efficiently realize fault tolerance by means of online and active fault diagnosis that enables redundancy-based fault-specific recovery or degradation strategies, e.g., system reconfigurations. In this way, they are able to overcome a failure or malfunction of some of their constituent components and continue to operate. As many systems are becoming more and more complex, there is an ever-growing amount of network traffic for monitoring and diagnostic purposes. In many cases, this causes the communication network to become the limiting resource, which can negatively impact the lengths of schedules, i.e., lead to longer overall service times, and reduce the maximum level of integration of different services in one system.
The use of data compression can help to reduce network traffic. However, due to the specific requirements and constraints of both time-triggered systems and diagnostic applications, some of which are contradictory, classical data compression algorithms cannot be straightforwardly applied. The difficulty lies in reconciling the needs for guaranteed data quality as well as temporal guarantees regarding information delivery. Lossless data compression does not support a certain worst-case compression ratio on short input sequences, but produces variable-length outputs. Therefore, it is difficult to guarantee the amount of information that will be encoded into a time-triggered message or, in turn, to guarantee the number of messages needed to transmit a certain amount of information. In lossy data compression, fixed-length outputs can be produced, but the quality of the reconstructed data might degrade. With respect to fault diagnosis, both incomplete or delayed data transmissions and reduced data quality lead to inaccurate fault identifications, which subsequently affects the fault recovery capabilities.
This thesis addresses the challenge of handling an increasing amount of network traffic within distributed time-triggered systems, with a particular focus on the data needed for diagnostic analyses. The thesis first presents a time-triggered architecture definition, including a compression model, that enables a systemwide coordination of data compression. Algorithms are then presented that go beyond the state of the art and realize online data compression in time-triggered systems. Specifically, the design of the novel algorithms allows the redundancy between multiple data streams to be exploited for compression by encoding the data simultaneously as tuples in a dynamic multidimensional product space. Correlated data streams are found as characteristic patterns in a low-dimensional subspace, which the compression algorithm covers by dynamically maintained dictionaries. This design also enables compressed data streams to be directly merged and split at arbitrary nodes within distributed systems, thus providing superior compression performance when the sources (or destinations) of multiple data streams are different. As key features, the presented data compression algorithms provide real-time capability through short worst-case compression and decompression times and, in addition, a worst-case compression ratio on short input sequences and a mechanism that guarantees a desired data quality with respect to the application.
The experimental evaluations show the advantages of compressed messages in resource-constrained real-time systems. Considering the trade-off between the reduction of communication times and the overhead in terms of computation times for compression, the analyses highlight the reduction potential in the response times of a time-triggered system. Compression allows to guarantee shorter deadlines and a higher integration of different services in one system. Providing a worst-case compression ratio on short input sequences, the online compression algorithms are not completely lossless, so the impact of degraded data quality is evaluated using a fault diagnosis use case and multiple test datasets that allow generalized conclusions. Especially with respect to the ability of the algorithms to compress multiple correlated data streams simultaneously, the evaluations show the scalability of the compression benefits for large systems. Savings from 40% up to 56% of the bits are observed in the analyzed scenarios. A comparison with other compression techniques highlights the advantages of the developed online data compression algorithms.