Linux Scheduler Performance for Beowulf Compute Nodes

In the last decade, the use of the Beowulf Cluster concept for High Performance Computing and Cloud Computing has exploded. In addition, the Linux operating system used by many of these computing platforms has also greatly advanced. Understanding the effects of the underlying kernel scheduler on the computational performance of compute nodes is one of the main concerns in customizing a Linux operating system. The overlying batch scheduling, message passing, and network communications are highly affected by subtle changes in kernel scheduler customizations. In this paper, we explore the computational performance of two types of Beowulf compute nodes, one using the O(1) scheduler and one using the new Completely Fair Scheduler, in various types of operating modes to determine an optimum scheduler and configuration.