Prof. Dr. Frank Bellosa, Andreas Merkel

Abstract: 

Constant advances in processor microarchitecture promise new levels of performance and efficiency with each new hardware generation. However, depending on the type of software a computer is processing, running the CPU at the highest frequency setting may not be the most efficient mode of operation. Previous work in the area of power management demonstrated that it is possible to make an operating system energy-aware in order to run hardware at its most efficient performance state. With the emergence of multi-core processor architectures, new challenges arise for operating system software and CPU power management in particular. The main obstacles are caused by hardware dependencies between cores where all cores share a common interface to main memory and groups of cores are subject to a common clock frequency.

This work presents an approach to apply previous knowledge about the energetic characteristics of microprocessors to computer systems based on a multi-core design. By employing an energy-aware scheduling algorithm, which collects specific energy-related information regarding the processes in the system, the operating system can determine the most efficient performance state the respective processor should operate at. This concept is augmented to the multi-core case by coordinating the set of individual performance state requests according to processor topology and the policy objectives the user can specify. The side effects shared memory bandwidth has on our policy are overcome by employing a heuristic that models the memory-related behavior of our platform.

We implemented energy-efficient scheduling for the Linux operating system kernel. Our evaluations show that the proposed design allows to run a multi-core system at optimum efficiency under varying workload conditions. The significant gains in energy efficiency that are possible with some workloads come at no additional hardware requirements or costs.

BibTex:

@diplomathesis{lieder08energy-efficient-scheduling,
  author = {Johannes Lieder},
  title = {Energy-Efficient Scheduling for Multi-Core Processors},
  type = {Diploma Thesis},
  address = {System Architecture Group, University of Karlsruhe, Germany},
  month = nov # "~18",
  year = 2008,
  url = {http://i30www.ira.uka.de/teaching/theses/pasttheses/}
}