Run-time mapping: dynamic resource allocation in embedded systems

Many desired features of computing platforms, such as increased fault tolerance, variable quality of service, and improved energy efficiency, can be achieved by postponing resource management decisions from design-time to run-time. While multiprocessing has been widespread in embedded systems for quite some time, allocation of (shared) resources is typically done at design-time to meet the constraints of applications. The inherent flexibility of large-scale embedded systems is then reduced to a fixed, static resource allocation derived at design-time. At runtime, unanticipated situations in either the system itself or in its environment may render resources inaccessible that were assumed to be available at design-time. The increased flexibility obtained by run-time resource allocation can be exploited to increase the degree of fault tolerance, quality of service, energy efficiency and to support a higher variability in use-cases. The term run-time mapping is used to refer to resource allocation at run-time to meet the dynamic requirements of applications. A mathematical analysis of the run-time mapping problem shows that each of its subproblems, i.e. task assignment and communication routing, is computationally complex due to the constraints representing the limited resource capacities of the embedded platform. Even if one of these subproblems is solved to optimality, the second optimization problem is stillNP-hard. Therefore, two different heuristic techniques are presented to tackle the run-time mapping problem. The first approach discussed in this thesis is a deterministic technique. Both the resources requested by applications, and the resources provided by the platform are modeled as graphs. A divide-and-conquer algorithm exploits the graph structures in order to generate many small resource allocation problems. Each resource allocation problem is a knapsack problem, where the resource requests (the items) are assigned to a subset of the available resources (the bins). In case of an insufficient number of bins, the algorithm increases the set of bins by considering more platform resources, until it runs out of resources. A prototype run-time mapping system which uses this algorithm is evaluated on a many-core processing platform developed in the CRISP project. Multiple real-life applications (various beamforming applications, a GPS receiver, and a dependability monitor) have been tested successfully with the run-time mapper which determines the resource assignment