الفهرس 
Chapter 1 : introductionOnly 14 pages are availabe for public view
 |  | from | 125 |  |  |
| | | from | 125 | | |
Abstract
Power management is a standout amongst the most essential issues identified regarding microprocessors design. Enhancing the performance of the processor often means increasing consumed energy. We have come as far as possible while increasing the performance of uniprocessors by adding more and more transistors. One reason for moving from uniprocessor architectures to multiprocessors is the energy efficiency issue. However, even homogeneous multiprocessor architectures were not good enough. Recently, the move to heterogeneous chip multiprocessors was apparently a good solution. One of the popular asymmetric architectures is ARM’s big.LITTLE architecture. All the cores in this architecture are CPU cores and they implement the same instruction set architecture but the cores have different microarchitectures. However, the power problem still exists as battery technology cannot scale to recent advance.The main objective of this thesis is to introduce a new clustering scheme to achieve better energy efficiency with minimum performance loss. This new scheme is mainly about the arrangement of different cores into clusters and the achievement of a balance between the number of big cores and LITTLE cores. An analytical model is introduced, that can be applied to processors with different number of cores. Then, this model is validated using the simulation of a case study of an octa-core processor. The proposed scheme also makes use of different frequency domains for different clusters to match the needs of each cluster and cores within the same cluster individually. The proposed scheme has been implemented based on the variation in performance and power efficiency for different types of cores implementing the same instruction set architecture. It also makes use of the idea of clustering and tying groups of cores to different frequency domains with a flexible scheduler mapping workloads to appropriate core/cluster. Experimental results showed an average of 73% power savings with an average performance loss of 16% while making use of the proposed scheme.