No abstract

No abstract

No abstract

Additionally, the main objective of the evaluation of the methodology with the dataflow benchmarks is to assess the advantages and disadvantages of assigning the parallelization capabilities to one framework or the other. The three applications selected prove to be a good representation of this problematic: one of them obtains better results when PREESM parallelizes -9× with PREESM, 4× with Apollo and 6× with the combination, always compared to Clang-; another only obtains speedups when Apollo is in charge -around 5×-; finally, the last one is much faster when the optimizations come from both frameworks -14× with PREESM, 75× with Apollo and almost 100× with the combination-. The latter is also used to functionally validate the changes added within SPiDER, proving the proposed methodology to successfully detect runtime changes and reconfigure the behavior of the multiversioning system accordingly, with similar speedups than those obtained statically -close to 100×.These two studies ensure the correctness of the proposed methodology, using to that end benchmark suites widely applied in the related domains; nonetheless, so as to guarantee that the proposed methodology fulfills the objectives for which it is envisioned and characterize its applicability, a real-life application is implemented using this methodology, considering it as a real use-case. This application has been extracted from a complex algorithm used to locate human brain tumor boundaries in intraoperative real-time during surgical procedures: specifically, the part of the algorithm selected to conform this use-case is the supervised classification stage of the hybrid classification algorithm specifically designed for this task, as it is the most time-consuming part of the algorithm. The results obtained show that, as expected, the largest speedups are reached when combining the optimization potential of both tools, with speedups close to 4× when compared to the sequential version. These results prove also to be competitive with those extracted from the state of the art, showing similar execution times using fewer Processing Elements (PEs) and applying automatic parallelization techniques instead of manually exploiting the parallelization, which in turn involves a non-negligible increase in the production times.In conclusion, the methodology proposed in this PhD has been proven to be successful in merging the parallelization potential of dataflow-based specifications with the optimization capabilities of automatic parallelization approaches, which is achieved by either exploiting separately the advantages of each domain when one fails to increase the throughput, or profiting from both potentials in those situations compliant with both worlds.