Prof. Dr. Frank Bellosa
Mathias Gottschlag

Abstract:

Asymmetric multicore processors (AMP) integrate multiple core types with different power and performance characteristics in a single package. Using optimized scheduling, these processors can deliver higher performance per watt than a symmetric multicore processor. An application executes most efficiently on a certain core depending on how it uses resources like CPU and memory. Previous approaches analyze applications at coarse granularities, classifying each process or thread. In systems such as servers that have homogeneous processes with similar behavior in all threads, these approaches cannot distribute applications to core types effectively. However, applications generally go through different execution phases over time. These phases often differ in their resource usage and exist at both large and small scales. Whereas some systems already incorporate changing application behavior over large time intervals, it should also be possible to utilize shorter phases to save energy by migrating between cores at high frequency. In this work, we design and implement such a system that characterizes the small-scale phase behavior of applications between developer-defined points by monitoring memory accesses with performance counters. At runtime, it migrates the application thread to the optimal core for each execution phase. We evaluate our system on an AMD Ryzen processor. These processors allow asymmetric core configurations using frequency scaling. We fail to see reductions in power consumption with our system on these processors. We show that, contrary to available documentation, Ryzen does not have per-core voltage domains and conclude that these processors are not suitable as asymmetric platform.

BibTex:

@masterthesis{werling18savingpower,
  author = {Lukas Werling},
  title = {Saving Power Without Sacrificing Performance on Asymmetric Multicore Processors},
  type = {Masterthesis},
  year = 2018,
  month = feb # "28",
  school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
  }