Evaluation of migration costs in multi-core CPU scheduling
Stefan Wächtler
TU Dresden
Verteidigung der Belegarbeit
The advent of multi-core CPUs brings new challenges to operating-system
schedulers. Commonly used scheduling algorithms avoid the migration of
threads because of shared data structures among different CPUs.
Additionally, it is believed that migrations are very costly due to
private CPU caches.
The goal of my work was to quantify and compare the costs of migrations
and preemptions. Modern hardware was used to identify factors
influencing these costs. The results of my work are compared to the
observations of a paper written by Bastoni et al.