SGER--CSR/SMA: Thermal Aware Dynamic Resource Management for Datacenters

Arizona State University, Scottsdale AZ

Investigators

Abstract

Minimizing the energy cost and improving thermal performance of datacenters are among the key issues towards optimizing computing resources and maximally exploiting a datacenter's computation capability. In general, there is no automated feedback control loop between the physical environment of a data center, and the datacenter's management software/system software scheduler for making thermal and power-efficient resource management decisions. The energy cost and thermal distribution of a datacenter depends on the multi-scale properties of a large number of multi-modality system variables, such as cooling capacity, hardware thermal characteristics, and workload profiles. An integrated, system-oriented approach towards this problem could provide several significant benefits. Thiswork, will investigate approachers for a unique merger between physical infrastructure and the resource management functions of the cluster operating system to take a holistic view of data center management, and make global (at the level of a datacenter), power-aware, and thermal-aware job scheduling decisions. The project seeks to develop a comprehensive solution for on-line, dynamic and automated thermal management and control framework for the power-limited datacenters. The methods include an integrated two-pronged approach for incorporating fast thermal evaluation, using an abstract heat model, in the resource management functions of the datacenters; namely: new thermal aware scheduling techniques will be developed by making traditional scheduling techniques cognizant of thermal performance of their scheduling decisions by taking into account the spatio-temporal thermal implications of job placements. This entails making scheduling decisions by taking into account not only job characteristics but also physical placement of computational node, the thermal characteristics of heterogeneous systems, and thermodynamics. These thermal-aware scheduling techniques will be comprehensively evaluated using new performance metrics such as utilization per operating cost.