Online Resource Management in Thermal and Energy Constrained Heterogeneous High Performance Computing

Operators of high-performance computing (HPC) facilities face conflicting trade-offs between the operating temperature of the facility, reliability of compute nodes, energy costs, and computing performance. Intelligent management of the HPC facility typically involves taking a proactive approach by predicting the thermal implications of allocating tasks to different cores around the facility. This offers the benefit of operating the HPC facility at a hotter CRAC temperature while avoiding hotspots. However, such an approach can be a time-consuming process that requires complicated air flow models to be calculated for every mapping decision. We propose a framework in which offline analysis is used to assist an online resource manager by predicting the thermal implications of mapping a given workload. The goal is to maximize the reward earned from completing tasks by their individual deadlines throughout the day, while adhering to a daily energy budget and temperature threshold constraints. We show that our proposed techniques can earn significantly greater reward than traditional load balancing and thermal management schemes.