Grant: Collaborative Research and Development Grants

Due to the power wall in CPU chip design, modern general-purpose processors organize the ever-increasing transistor capacity of CPU dies into multiple independent processing units, or cores. Initially, CPU core configurations were homogeneous and static. However, there is a trend towards CPUs that support dynamic, heterogeneous core configurations. This means that CPU properties such as the number of active cores, the speeds of the active cores, and the cores' cache configurations, can be changed on the fly to fit the characteristics of the CPU's workload. Dynamic adjustment of the CPU configuration offers many potential benefits, including better performance, reduced energy consumption, and improved ability to target CPU resources to high priority workloads.This project addresses the problem of how to actively manage CPU configurations to achieve these benefits for realistic workloads. We focus on server-side workloads found in enterprise and cloud data centers. Because many important server-side applications rely on some sort of database, these applications are supported by a multi-tier software infrastructure that includes both operating systems and database management systems. Our approach is to extend this infrastructure to provide active management of CPU configurations on behalf of the applications that it supports.Both operating systems and database management systems have a role to play in CPU configuration management. The former has fine control over job scheduling and direct access to the CPU, while the latter has richer knowledge of the hosted applications' characteristics and objectives. To reflect this, we will explore a coordinated, multi-tier approach to CPU configuration management that will exploit the capabilities of both systems to achieve our performance, energy, and workload management objectives.