Absolute Bandwidth Scheduling via Group-based Partitioned Proportional Share Scheduling and Dynamic Weight Management on Varying-Speed Multiprocessors

Abstract

For the past few years there has been an increase in the use of compute intensive applications running on highperformance embedded systems based on multi-core platforms. These applications demand an absolute share of CPU bandwidth to guarantee a certain level of QoS (Quality of Service) and fulfil their timing constraints. Unfortunately, traditional proportional share or priority scheduling algorithms employed in general purpose operating systems are not able to provide an absolute share of processor resources for such time sensitive tasks. In this paper, we present an absolute bandwidth scheduling scheme which aims at providing an absolute share of CPU bandwidth to groups of soft real-time tasks regardless of the work load conditions and varying speeds of CPU. The proposed scheme provides a mechanism of CPU bandwidth allocation for groups of softreal time tasks by dynamically changing the overall weight of the group while maintaining the proportion of share of each task in the group. Our proposed approach works on top of a traditional proportional share scheduler and does not require any modifications to the kernel layer. To demonstrate the effectiveness and the correctness of our scheme, we have implemented a prototype using Linux cgroups and the existing completely fair scheduler (CFS). A series of experiments are conducted to prove that each soft real-time task in the group maintains its required absolute bandwidth.