Coordinating processor and main memory for efficientserver power control

  • Authors:
  • Ming Chen;Xiaorui Wang;Xue Li

  • Affiliations:
  • University of Tennessee, Knoxville, TN, USA;University of Tennessee, Knoxville, TN, USA;University of Tennessee, Knoxville, TN, USA

  • Venue:
  • Proceedings of the international conference on Supercomputing
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

With the number of high-density servers in data centers rapidly increasing, power control with performance optimization has become a key challenge to gain a high return on investment, by safely accommodating the maximized number of servers allowed by the limited power supply and cooling facilities in a data center. Various power control solutions have been recently proposed for high-density servers and different components in a server to avoid system failures due to power overload or overheating. Existing solutions, unfortunately, either rely only on the processor for server power control, with the assumption that it is the only major power consumer, or limit power only for a single component, such as main memory. As a result, the synergy between the processor and main memory is impaired by uncoordinated power adaptations, resulting in degraded overall system performance. In this paper, we propose a novel power control solution that can precisely limit the peak power consumption of a server below a desired budget. Our solution adapts the power states of both the processor and memory in a coordinated manner, based on their power demands, to achieve optimized system performance. Our solution also features a control algorithm that is designed rigorously based on advanced feedback control theory for guaranteed control accuracy and system stability. Compared with two state-of-the-art server power control solutions, experimental results show that our solution, on average, achieves up to 23% better performance than one baseline for CPU-intensive benchmarks and doubles the performance of the other baseline when the power budget is tight.