Optimization based resource and cooling management for a high performance computing data center.

Area of Science:

• Computer Science
• Energy Efficiency
• High-Performance Computing

Background:

• High-performance computing (HPC) data centers face significant energy consumption challenges, particularly with numerous small-sized jobs.
• Existing methods often prioritize energy saving over processing efficiency, leading to suboptimal performance.
• Effective management of computing resources and cooling systems is crucial for energy reduction.

Purpose of the Study:

• To develop an energy-efficient job scheduling and resource provisioning strategy for HPC data centers.
• To minimize cooling system energy consumption without compromising job processing efficiency.
• To propose a novel approach prioritizing processing efficiency in energy reduction.

Main Methods:

• A hysteresis computing resource-provisioning algorithm was developed to adjust resources reactively.
• A dynamical thermal model was created to link computational load with cooling system performance.
• Constrained optimal control problems were formulated to minimize cooling energy usage.
• A two-step solution involving a thermal-aware resource allocation optimizer and an economic model predictive controller was implemented.

Main Results:

• Simulations using real job traces demonstrated significant energy savings in cooling systems.
• The proposed method effectively maintained high job processing efficiency.
• Guaranteed service quality was achieved alongside substantial energy reduction.
• The dynamical thermal model accurately reflected system interdependencies.

Conclusions:

• The proposed hysteresis algorithm and thermal-aware control effectively reduce energy consumption in HPC data centers.
• Prioritizing job scheduling and processing efficiency leads to substantial energy savings.
• The integration of thermal modeling and predictive control offers a viable solution for sustainable HPC operations.