The shift to manycore architectures has highlighted the need for runtime power and performance management schemes to improve the reliability, performance, and energy-efficiency of processors. However, the design of management algorithms is challenging since power and performance are strongly dependent on the workload, which cannot be determined apriori and exhibit wide and rapid runtime variations.
This dissertation seeks to show that sensitivity analysis (derivative estimation) provides runtime power and performance information that enables the design of adaptive and low-complexity management algorithms. The contributions of the dissertation include: 1) controllers that achieve rapid regulation of the power and throughput of processor cores, 2) a chip-level power control solution that maximizes the performance of manycore processors subject to the power constraints set by the cooling system, and 3) an iterative algorithm for optimizing the energy consumption of cache memories. The proposed algorithms use runtime derivative estimation to adapt to the rapid power and performance variations caused by workload, and their efficacy is demonstrated via formal analysis and simulation experiments.
| Identifer | oai:union.ndltd.org:GATECH/oai:smartech.gatech.edu:1853/51805 |
| Date | 22 May 2014 |
| Creators | Almoosa, Nawaf I. |
| Contributors | Yalamanchili, Sudhakar, Wardi, Yorai |
| Publisher | Georgia Institute of Technology |
| Source Sets | Georgia Tech Electronic Thesis and Dissertation Archive |
| Language | en_US |
| Detected Language | English |
| Type | Dissertation |
| Format | application/pdf |
Page generated in 0.0056 seconds