Hierarchical Multiprocessor CPU Reservations for the Linux Kernel

Abstract

This paper presents ongoing work in the development of a scheduling framework that will improve the service guarantees for soft real-time applications deployed on Linux. The scheduler has been designed around the current kernel infrastructure, trying to keep the changes minimal, and basing the scheduling policy on strong theoretical results. The main goal is to achieve hierarchical distribution of the available computing power on multiprocessor platforms, avoiding alterations to the existing user interfaces. The proposed framework exploits the hierarchical arrangement of tasks within groups and subgroups that is already possible within the Linux kernel. However, it adds the capability for each group to be assigned a precise fraction of the computing power available on all the processors, using existing uni-processor resource reservation techniques. Tasks are scheduled globally within each single group, and the partitions assigned to each group need not to be static, but can be dynamically balanced. Furthermore, the proposed mechanism can be used to support a variety of possible partitioning schemes using processor affinities.