Untersuchungen von Kommunikationsmechanismen in heterogenen Mehrprozessorsystemen

Modern mobile communication standards are characterized by continuously increasing signal processing complexity and strict energy-efficiency, latency and performance requirements. To fulfill these conflicting requirements, heterogeneous multiprocessor system on chips are used, which embed specialized and programmable cores for time critical and complicated computation tasks. An integration of these heterogeneous processing elements in combination with intelligent DMA-Controllers can further increase the system throughput as well as energy efficiency of the multiprocessor. In this work, a multiprocessor system on chip containing several heterogeneous processing elements controlled by a dedicated scheduler for task dispatching is investigated. In these systems, the central scheduler is a bottleneck that limits the performance due to its managing tasks. Hence, this work analyzes concepts to increase the system performance in heterogeneous multiprocessors. As a suitable solution to reduce the managing overhead at the central scheduler, a controller is integrated into the heterogeneous processing elements in order to allow local task management and configuration. Subsequently, the managing overhead at the central scheduler is reduced, resulting in increased system performance. Furthermore, mechanisms for data preprocessing in upcoming cloud computing workload hosted in basestations, e.g. database processing, are studied. Therefore, functionality for data selection, data modification on bit-level and data distribution to processing elements are integrated in an intelligent DMA-Controller. Thus, recurring workloads at the processing elements are reduced and the system performance is enhanced. To further increase the system performance an extension of the intelligent DMA-Controller for processing of dynamic datastructures, which are often used in database systems, is investigated. This allows preprocessing of datasets directly at the memory by the intelligent DMA-Controller. Additionally, another mechanism is analyzed in order to also accelerate applications based on the frequently used client-server model. The integration of these mechanisms into the DMA-Controller again leads to significant performance increase of the multiprocessor system.