Abstract
dc:description.abstractPower has become a critical constraint for the evolution of large scale High Performance Computing (HPC) systems and commercial data centers. This constraint spans almost every level of computing technologies, from IC chips all the way up to data centers due to physical, technical, and economic reasons. To cope with this reality, it is necessary to understand how available or permissible power impacts the design and performance of emergent computer systems. For this reason, we propose power bounded computing and corresponding technologies to optimize performance on HPC systems with limited power budgets. We have multiple research objectives in this dissertation. They center on the understanding of the interaction between performance, power bounds, and a hierarchical power management strategy. First, we develop heuristics and application aware power allocation methods to improve application performance on a single node. Second, we develop algorithms to coordinate power across nodes and components based on application characteristic and power budget on a cluster. Third, we investigate performance interference induced by hardware and power contentions, and propose a contention aware job scheduling to maximize system throughput under given power budgets for node sharing system. Fourth, we extend to GPU-accelerated systems and workloads and develop an online dynamic performance & power approach to meet both performance requirement and power efficiency. Power bounded computing improves performance scalability and power efficiency and decreases operation costs of HPC systems and data centers. This dissertation opens up several new ways for research in power bounded computing to address the power challenges in HPC systems. The proposed power and resource management techniques provide new directions and guidelines to green exscale computing and other computing systems.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (PhD)
- Level thesis:degree_level
- Dissertation
- Discipline thesis:degree_discipline
- School of Computing
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zou, Pengfei
- Contributors dc:contributor
-
- Rong Ge
- Valerie Taylor
- Amy Apon
- Feng Luo
- Xizhou Feng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://open.clemson.edu/all_dissertations/2618