Mining Task Precedence Graphs from Real-Time Embedded System Traces

Real-time embedded systems have evolved from simple, self-contained single-processor computers to distributed multiprocessor systems that are extremely hard to develop and maintain. Execution tracing has proved itself to be a useful technology to gain a detailed knowledge of runtime behavior of software systems. However, the size and complexity of execution traces generated by modern embedded systems make manual trace analysis impossible. Therefore, software developers need tools to extract high-level system models from raw trace data.  In this paper, we address the problem of mining task precedence graphs (TPG) from embedded system traces. A TPG can be helpful in performing several crucial software development and maintenance activities: understanding legacy systems, finding runtime bugs, and detect and diagnose anomalies in running systems. We rely on the recurrent nature of real-time systems to solve the TPG mining problem.
We propose algorithms to train a TPG on a set of system traces, as well as an algorithm to detect anomalies in trace streams using a TPG. We evaluate our algorithms on industrial execution traces generated on production cars