A Utility Based Power-Aware Autonomic Approach for Running Scientific Applications

Abstract

Traditionally, computational power and performance of the high performance computing (HPC) systems are assumed to depend upon the number of processors and speed of supercomputers. Supercomputers run at their peak performance to efficiently execute scientific applications. Therefore, these supercomputers consume enormous amount of power that results in increased operational cost. The high power consumption translates into high temperature of the physical HPC systems, which in turn results in high failure rate and decreased reliability. Employing an aggressive cooling system does not improve the situation, because it involves an additional operational and infrastructure cost. Slowing down these HPC systems results in loss of performance that is also not recommended. Moreover, the execution of scientific applications is affected adversely through the variation in available computational resources. This variation is due to the computational resource utilization by other applications running at the computing nodes. Therefore, the execution of a scientific application should be monitored with respect to its performance objectives (deadline etc.). Failure in meeting the deadlines may result in high cost in terms of revenue loss to the service providers. These issues raise the motivation towards the designing of an autonomic approach for managing power consumption in HPC systems without adversely affecting to the performance of the system. In this paper, we present a utility based power-aware approach 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 the execution of the application.