Model assisted automated HW/SW partitioning

The future of computing is heterogeneous. With technological trends to includeprocessing capability into a large variety of devices, today’s challenges indistributed computing are no longer solved on homogenous systems. Not onlyfederations of devices are used to tackle various difficulties, also single devicesnowadays provide a collection of diverse compute units.This thesis presents a generalized abstraction of algorithms, suited to analyseits potential for distribution. These algorithms might emerge from any domainof software tasks, such as number crunching, machine learning, classical signalprocessing, or control theory. Such an abstract view does combine the elementsof static and dynamic analysis of the algorithm, and deals with the concretepattern of its computational profile. As counterpart to the model of thecomputation, a representation of a dedicated compute arrangement is proposed,in order to deal with the diverse nature of a complex of compute-nodes. Thisprovides an approach to model heterogeneous systems. These systems can be acollection of embedded devices, like micro-controllers, but also a combination ofcloud servers with edge computing devices. With these two models in place anestimation procedure is given to predict the performance profile of the hardwaresetup upon serving the defined algorithm. The performance profile does notonly include the overall execution time, but also a detailed view on the dataflow and node allocation. Such an allocation profile for example allows drawingconclusions on how to improve the hardware arrangement, or the algorithm, tomore efficiently achieve its goal, based on concrete facts.In this work a set of algorithms is analysed by the procedure given, and itsdistribution pattern is dissected in detail. Exemplarily the impact of variousmodifications and extensions, like adding additional compute nodes, to thehardware setup is performed. For these adaptions the impact on the allocationprofile is provided, such that a concrete judgment on their impact can be made.Based on the presented results, future research directions in the domain ofheterogeneous and parallel distribution of algorithms are illustrated.