A power-aware autonomic approach for performance management of scientific applications in a data center environment

Abstract

In the recent years, computer servers and data center facilities that provide high performance computing (HPC) for scientific applications have largely increased in numbers and have become great consumers of electrical power. Supercomputers often run at their peak performance for an efficient execution of scientific applications, and therefore consume an enormous amount of power that results in increased operational cost. Furthermore, an increase in the power consumption results in an increase in the temperature of the physical HPC systems, which in turn translates into increased failure rates and decreased reliability. Slowing down these HPC systems by reducing the individual speed of the processors, results in a loss of execution performance of the scientific application, due to the variation in processing speed. Another cause of the degradation in the execution performance of scientific applications is the variation in the computational resource availability due to its utilization by other applications executing on the same computing node in a space shared manner. The variations in processor availability can lead to severe performance degradation in the execution environment due to load imbalance and a violation of the performance objectives, such as meeting a deadline, and therefore it may result in high penalty in terms of revenue loss to the service providers. In this chapter, a utility based power-aware approach has been presented that uses a model-based control theoretic framework for executing scientific applications. The approach and related simulations indicate that the performance and the power requirements of the system can dynamically be adjusted, while maintaining the predefined quality of service (QoS) goals in terms of deadline of execution and power consumption of the HPC system, even in the presence of computational resource related perturbations. This approach is autonomic, performance directed, dynamically controlled, and independent of (does not interfere with) the execution of the application.