Model-based Deployment Optimization of Automotive Multi- and Many-Core Systems

The demands on computing platforms in automotive systems are constantly increasing due to the growing amount of software required by new vehicle functions. At the same time, vehicles typically consisting of about 100 mainly mono-functional electronic control units are driven towards centralized electrical/electronic architectures to handle the computational demands. Accordingly, the process of integrating software is becoming increasingly challenging, as a mixture of applications originating from heterogeneous functional domains needs to be deployed towards heterogeneous hardware architectures consisting of different types of processing units and accelerators. <br /><br /> The scope of this thesis lies on the challenges and techniques for integrating such applications from heterogeneous functional domains in the context of automotive systems onto centralized computing platforms, particularly in early design phases. For this purpose, it formulates one primary and three secondary research questions that guide through the remainder of this document. Specifically, they address (i) efficient deployment optimization techniques that support the aforementioned integration process, (ii) timing analysis techniques that allow analysing complex automotive systems, (iii) support for interference originating from concurrent memory access operations in the context of timing analysis, and (iv) techniques for increasing efficiency of the deployment optimization process.<br /><br /> As a result, this thesis not only insights into the specific execution characteristics of automotive systems, but also provides formal definitions based on the Amalthea Data Model that allow representing and analysing applications from heterogeneous functional domains, such as engine management systems or advanced driver assistance systems. Moreover, it investigates common automotive standards, and evaluates them with regard to the large variety of timing related metrics that are typically defined in terms of constraints or quality attributes of automotive systems. Based on the intermediate outcomes, it further investigates methods for quantifying these metrics using abstract system models. Furthermore, it proposes novel techniques for quantifying those metrics that, to the best of the author’s knowledge, were not or not sufficiently covered in academic literature, such as (i) event models for the trigger patterns defined in automotive standards, such as the Amalthea Data Model, (ii) notations for quantifying dropped activation events that originate from exceeding a process’s multiple-task activation limit, (iii) response-time analysis of processes in consideration of their multiple-task activation limit, (iv) a generalization of existing work to account for interference from concurrent memory access operations originating from different processing units using information available in the Amalthea Data Model, and (v) notations quantifying the durations a process remains in states according to the best trace format. A selection of the presented techniques is subsequentially validated towards their correctness and evaluated with regard to their (i) run-time efficiency for bounding individual metrics and (ii) the respective improvement in comparison to other techniques.<br /><br /> Afterwards, this work evaluates existing frameworks and optimization techniques with regard to their underlying optimization algorithm and supported degrees of freedom that are employed to find feasible deployments. Based on this outcome, it develops two (hybrid) genetic algorithms that (i) allocate processes to processing units while assigning priorities and (ii) cluster labels according to their accessing processing units and subsequentially map these clusters to memories. Our experimental results related to evaluating their performance in terms of effectiveness and run-time efficiency show that, compared to existing techniques, both approaches are significantly faster in constrained solution spaces, offer a higher scalability, and are capable of obtaining better solutions in terms of the respective optimization goals.<br /><br /> Finally, the aforementioned techniques are practically validated and evaluated towards their applicability on industrial systems. For this purpose, this work develops customized hybrid genetic algorithms tailored towards the specific deployment optimization problems, which consists of (i) models from two publicly available industrial systems, (ii) a customized industrial system that has been extended to account for network-on-chip hardware architectures, and (iii) a proprietary engine management system. The approaches were consistently capable of finding feasible deployments for all of these case studies in the context of single- and/or multi-objective optimization. In addition, they were capable of improving quality attributes, such as a system’s minimum relative earliness by up to 78%, the average relative earliness by up to 92%, and individual reaction latencies by up to 94%.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Bielefeld

Year

2022

Author dc:creator

• Krawczyk, Lukas

Repository record source_url

https://pub.uni-bielefeld.de/record/2978215

OAI identifier oai:identifier

oai:pub.uni-bielefeld.de:2978215