Processor allocator for chip multiprocessors

Chip MultiProcessor (CMP) architectures consisting of many cores connected through Network-on-Chip (NoC) are becoming main computing platforms for research and computer centers, and in the future for commercial solutions. In order to effectively use CMPs, operating system is an important factor and it should support a multiuser environment in which many parallel jobs are executed simultaneously. It is done by the processor management system of the operating system, which consists of two components: Job Scheduler (JS) and Processor Allocator (PA). The JS is responsible for job scheduling that deals with selection of the next job to be executed, while the task of the PA is processor allocation that selects a set of processors for the job selected by the JS. In this thesis, the PA architecture for the NoC-based CMP is explored. The idea of the PA hardware implementation and its integration on one die together with processing elements of CMP is presented. Such an approach requires the PA to be fast as well as area and energy efficient, because it is only a small component of the CMP. The architecture of hardware version of a PA is presented. The main factor of the structure is a type of processor allocation algorithm, employed inside. Thus, all important allocation techniques are intensively investigated and new schemes are proposed. All of them are compared using experimentation system. The PA driven by the described allocation techniques is synthesized on FPGA and crucial energy and area consumption together with performance parameters are extracted. The proposed CMP uses NoC as interconnection architecture. Therefore, all main NoC structures are studied and tested. Most important parameters such as topology, flow control and routing algorithms are presented and discussed. For the proposed NoC structures, an energy model is proposed and described. Finally, the synthesized PAs and NoCs are evaluated in a simulation system, where NoC-based CMP is created. The experimental environment took into consideration energy and traffic balance characteristics. As a result, the most efficient PA and NoC for CMP are presented