Deployment and Debugging of Real-Time Applications on Multicore Architectures

Abstract

It is essential to enable information extraction from software. Program tracing techniques are an example of information extraction. Program tracing extracts information from the program during execution. Tracing helps with the testing and validation of software to ensure that the software under test is correct. Information extraction is done by instrumenting the program. Logged information can be stored in dedicated logging memories or can be buffered and streamed off-chip to an external monitor. The designer inspects the trace after execution to identify potentially erroneous state information. In addition, the trace can provide the state information that serves as input to generate the erroneous output for reproducibility. Information extraction can be difficult and expensive due to the increase in size and complexity of modern software systems. For the sub-class of software systems known as real-time systems, these issues are further aggravated. This is because real-time systems demand timing guarantees in addition to functional correctness. Consequently, any instrumentation to the original program code for the purpose of information extraction may affect the temporal behaviors of the program. This perturbation of temporal behaviors can lead to the violation of timing constraints, which may bias the program execution and/or cause the program to miss its deadline. As a result, there is considerable interest in devising techniques to allow for information extraction without missing a program’s deadline that is known as time-aware instrumentation. This thesis investigates time-aware instrumentation mechanisms to instrument programs while respecting their timing constraints and functional behavior. Knowledge of the underlying hardware on which the software runs, enables the extraction of more information via the instrumentation process. Chip-multiprocessors offer a solution to the performance bottleneck on uni-processors. Providing timing guarantees for hard real-time systems, however, on chip-multiprocessors is difficult. This is because conventional communication interconnects are designed to optimize the average-case performance. Therefore, researchers propose interconnects such as the priority-aware networks to satisfy the requirements of hard real-time systems. The priority-aware interconnects, however, lack the proper analysis techniques to facilitate the deployment of real-time systems. This thesis also investigates latency and buffer space analysis techniques for pipelined communication resource models, as well as algorithms for the proper deployment of real-time applications to these platforms. The analysis techniques proposed in this thesis provide guarantees on the schedulability of real-time systems on chip-multiprocessors. These guarantees are based on reducing contention in the interconnect while simultaneously accurately computing the worst-case communication latencies. While these worst-case latencies provide bounds for computing the overall worst-case execution time of applications on chip-multiprocessors, they also provide means to assigning instrumentation budgets required by time-aware instrumentation. Leveraging these platform-specific analysis techniques for the assignment of instrumentation budgets, allows for extracting more information from the instrumentation process.