Dynamic Thermal Management for Distributed Systems

In modern data centers, the impact on the thermal properties by increased scale and power densities is enormous and poses new challenges on the designers of both computing as well as cooling systems. Controltheoretic techniques have proven to manage the heat dissipation and the temperature to avoid thermal emergencies, but are not aware of the task currently executing or its specific service requirements. In this work we investigate an approach to dynamic thermal management with respect to the demands of individual applications, users or services. We show that the energy consumption and the temperature can be determined on a fine grained level and without the need for measurement, using information from event monitors embedded in modern processors. We extend the well-known abstraction of resource containers to an infrastructure for transparent energy and temperature management in distributed systems. The processing of a request can be throttled to meet the thermal requirements of the system, even if machine boundaries are crossed, e.g. by remote procedure calls in a client/server relationship. With this facility, energy consumption can be accounted and the resulting heat generation be controlled precisely without the need for expensive hardware. Experiments on a Pentium 4 architecture show that energy and temperature are accurately determined and thermal limits for the individual CPU and the whole distributed system will not be exceeded. Use cases and important implications of our approach are discussed.