Multi-Core Software

Process Scheduling Challenges in the Era of Multi-Core Processors

INTRODUCTION

In multi-core processor packages, each processor package contains two or more execution cores, with each core having its own resources (registers, execution units, some or all levels of caches, etc.). Even if the applications are not multi-threaded, multi-tasking environments will benefit from multi-core processors.

Design innovations of multi-core processor architectures mainly span the area of shared resources (caches, power management, etc.) between cores, core topologies (number of cores in a package, relationship between them, etc.), and platform topology (relation between cores in different packages, etc.). These innovations bring new opportunities and challenges to the system software. To exploit optimal performance, components such as the process scheduler need to be aware of the multi-core topologies and the task characteristics.

We start with a brief look at how the traditional process scheduler works and how the earlier challenges in the Symmetric Multi Processing (SMP), Non Uniform Memory Access (NUMA), and Simultaneous Multi-Threading (SMT) environments were addressed. We look at multi-core topologies with respect to core, cache, power management, and platform topologies. In the current generation of mainstream multi-core processors, the execution cores in a given processor package are symmetric and our focus in this paper is on such processors. Asymmetric multi-core processors are beyond the scope of this paper. We examine the need for a multi-core-aware process scheduler and look into the opportunities in this area. We examine different scheduling mechanisms for multi-core platforms under different load scenarios and the associated tradeoffs. With Linux* as an example, we examine how some of these scheduling mechanisms are currently implemented. Finally, we close this paper with a look at current and future research in this field.

Back to Top