Dr. Andreas Weißel, Prof. Dr. Frank Bellosa

Abstract:

The development of processors in recent years with its astonishing increase in speed and complexity leads to problems in regard to power consumption and cooling. With shorter clock cycles and increased complexity a modern CPU has to consume far more energy than its simpler predecessors. The processors are reaching the limit of the current standard CPU-cooling method: fan and heatsink.

From a physical point of view the processor takes electric energy and transforms it to heat which has to be diverted from the processor. Here the development of processors becomes visible: from the early CPUs up to some types of the i486 which did not need any real heat sink to cool it up to the current Pentium 4 and graphic accelerators where an additional fan is necessary just to increase the heat sink's efficiency. The upcoming discussion on better cooling solutions like heat pipes, water cooling and so on also indicates a growing need for thermal processor design.

It is getting difficult in air conditioned server farms with racks containing several computers on very small space to increase the density of processors. The air conditioning required when using current CPUs is challenging, so limits on power consumption per rack are introduced to prevent an overload of the installed air conditioning units.

Modern processors and motherboards allow the measurement of the CPU temperature. This is a prerequisite of temperature control, either in hardware like the Pentium 4 or various software methods. All of these methods have drawbacks in either affecting all processes simultaneously or not being applicable to current processors.

This work provides a process-specific approach to temperature control utilizing performance counters to account the energy consumption of processes, which can be used to compute the temperature of the processor.

This thermal model allows the computation of a dynamical power consumption limit for the machine, the enforcement of which keeps the processor temperature lower than a set limit. Combined with the facility of resource containers this approach features flexible task specific throttling while maintaining a temperature limit.

BibTex:

@studythesis{kellner03tempcontrol,
  author = {Simon Kellner},
  title = {Event-Driven Temperatur Control in Operating Systems},
  type = {Study Thesis},
  address = {Operating System Group, University of Erlangen, Germany},
  month = apr # "~30",
  year = 2003,
  url = {http://www4.informatik.uni-erlangen.de/SA/pdf/SA-I4-2003-02-Kellner.pdf}
}