Robust control-theoretic thermal balancing for server clusters

Thermal management is critical for clusters because of the increasing power consumption of modern processors, compact server architectures and growing server density in data centers. Thermal balancing mitigates hot spots in a cluster through dynamic load distribution among servers. This paper presents two Control-theoretical Thermal Balancing (CTB) algorithms that dynamically balance the temperatures of different servers based on online measurements. CTB features controllers rigorously designed based on optimal control theory and a difference equation model that approximates the thermal dynamics of clusters. Control analysis and simulation results demonstrate that CTB achieves robust thermal balancing under a wide range of uncertainties: (1) when different tasks incur different power consumptions on the CPUs, (2) when servers experience different ambient temperatures, and (3) when servers experience thermal faults.