Workload-Guided Application Scheduling in Multi-Core Systems

Description


Heterogeneous multi-core processors have emerged more efficient as compared to homogeneous multi-core processors. This is due to the ability of heterogeneous multi-core processors to meet the different resource requirements of applications, and hence achieve power-efficient computing. However, one of the challenges of using a heterogeneous multi-core processor is to schedule different programs in a workload to the most suited core that can deliver the most efficient program execution.

Prior research mainly focuses on dynamic core selection based on sampling the behavior of neighboring or all cores. Although this dynamic method can identify program phase changes during runtime and make corresponding core switching, it does not help map applications statically nor lead to more intelligent dynamic core selections. UT researchers have come up with a method to map applications to the optimum core by analyzing the micro-architecture independent characteristics of that particular application.

The proposed technique is a workload-guided scheduling mechanism that employs fuzzy logic to calculate the suitability between program and cores by analyzing important inherent program characteristics. The obtained suitability is used to guide program scheduling in the heterogeneous multi-core system. This relationship between inherent program behavior and the corresponding resource requirements is important, in the sense that it can not only help map applications statically according to off-line profiling but also lead to more intelligent dynamic core selections. This technique thus helps in mapping applications statically to the proper cores based on the micro-architecture independent characteristics.

The experimental results show that the proposed suitability-guided program scheduling mechanism achieves up to 15% average reduction in energy-delay product compared with that of a random scheduling approach.


Benefits

  • Helps exploit the core diversity that exists in heterogeneous multi-core systems
  • Helps design a more intelligent dynamic program scheduling mechanism in heterogeneous multi-cores than the current trial and error approach
  • Has no additional power and performance cost as opposed to the existing trial-and-error approach
  • Scales well as the number of cores increases, while the existing trial-and-error approach does not


Market Potential/Applications

Heterogeneous multi-core systems


For further information please contact

University of Texas,
Austin, USA
Website : www.otc.utexas.edu