Software Concerns for Execution on Heterogeneous Platforms

Licentiate thesis, 2018

Context: Heterogeneous computing, i.e., computing performed on different types of execution units, such as CPUs, GPUs, FPGAs, has shown to be a feasible path towards higher performance and less energy consumption. Heterogeneous platforms are specialized on specific types of computation, e.g., parallel computing. However, this approach imposes a number of challenges on the software side. One of such challenges is related to software deployment, in which applications must be prepared to be executed in different target architectures. Further, the approach demands a robust inter-process communication solution, since these systems inherently distribute computation.

Objective: The objective of this thesis is twofold. First, to provide an overview of the state-of-the-art of software deployment on heterogeneous platforms, with emphasis to goals, concerns, challenges, and identification of topics of importance for further research. Second, to investigate the communication problem and propose a novel method that improves inter-process communication in distributed systems.
Method: Six papers were written as results of four studies: (i) a literature review in the form of a systematic mapping study on software deployment on heterogeneous platforms; (ii) a systematic evaluation of deployment methods in the context of a self-driving heavy vehicle; (iii) an investigation on data marshalling approaches and how they perform in the context of a cyber-physical system; and (iv) a novel message restructuring approach, also in the context of cyber-physical systems.

Results and Conclusions: The mapping study discussed the (i) concerns on the topic such as scheduling and software quality; the (ii) approaches available, such as frameworks; and the (iii) architecture solutions used, such as styles and principles. In the second study, we found that the performance decay is negligible when using sandboxed environments for deployment. In the third and fourth studies, we proposed and evaluated a data marshalling approach that decreases bandwidth consumption.
Future work: We intend to identify challenges that are currently faced in an industrial setting. In particular, a migration from a non-heterogenous platform to a heterogeneous platform can be studied in the context of a modern software development process, such as DevOps.